Suhas Pai
Chief Technology Officer, TMLS 2022 Chair
Click Speaker to Read Abstract
Click Speaker to Read Abstract
Click Speaker to Read Abstract
Click Speaker to Read Abstract
Click Speaker to Read Abstract
Click Speaker to Read Abstract
Loading...
Presenter:
Yoshua Bengio, Scientific Director, Mila / Full Professor, University of Montreal
About the Speaker:
Recognized worldwide as one of the leading experts in artificial intelligence, Yoshua Bengio is most known for his pioneering work in deep learning, earning him the 2018 A.M. Turing Award, “the Nobel Prize of Computing,” with Geoffrey Hinton and Yann LeCun.
He is a Full Professor at Université de Montréal, and the Founder and Scientific Director of Mila – Quebec AI Institute. He co-directs the CIFAR Learning in Machines & Brains program as Senior Fellow and acts as Scientific Director of IVADO.
In 2019, he was awarded the prestigious Killam Prize and in 2022, became the computer scientist with the highest h-index in the world. He is a Fellow of both the Royal Society of London and Canada, Knight of the Legion of Honor of France and Officer of the Order of Canada.
Concerned about the social impact of AI and the objective that AI benefits all, he actively contributed to the Montreal Declaration for the Responsible Development of Artificial Intelligence.
Talk Track: Technical / Research
Talk Technical Level: 7/7
Abstract:
Current large language models and other large-scale neural nets directly fit data, thus learning to imitate its distribution. Is it possible to do better? Consider the possibility, which we claim is actually the typical case, where a world model that captures causal structure in the world would require substantially fewer bits than the code that performs the kinds of inferences that we may desire from that knowledge. For example, the rules of the game of Go or the rules of logical reasoning are fairly compact, whereas inference (e.g., playing Go at Champion level, or being able to discover or prove theorems) may require a lot more computation (i.e., huge neural nets).
In fact, exact inference for many problems (in science, computing, or in coming up with Bayesian posteriors) is often intractable and so we resort to approximate solutions. With amortized probabilistic inference, we expend computation upfront to train a neural net that learns to answer questions (for example sample latent variables or parameters) in a way that is as consistent as possible with the world model, making run-time inference quick.
This also makes it possible to exploit the generalization ability of the approximate inference generative learner to guess where good answers (e.g., modes of a highly multimodal posterior distribution) might be. It also allows to decouple the capacity needed for inference from the capacity needed to describe the world model. This is unlike current state-of-the-art in deep learning, where inference and world model are confounded, yielding overfitting of the world knowledge and underfitting of the inference machine.
We have recently introduced a novel framework for achieving this kind of model-based ML, with generative flow networks (or GFlowNets), which have relations to reinforcement learning, variational inference and generative models. We’ll highlight some of the advances achieved with GFlowNets and close with our research programme to exploit such probabilistic inference machinery to incorporate in ML inductive biases inspired by high-level human cognition and build AI systems that focus on understanding the world in a Bayesian and causal way and generating probabilistically truthful statements.
Presenter:
David Rosenberg, Head of ML Strategy, Office of the CTO Bloomberg
About the Speaker:
David Rosenberg leads the Machine Learning Strategy team in the Office of the CTO at Bloomberg. He is also an adjunct associate professor at the Center for Data Science at New York University, where he has repeatedly received NYU’s Center for Data Science “Professor of the Year” award. Previously, he was Chief Scientist at Sense Networks, a location data analytics and mobile advertising company, and served as scientific adviser to Discovereads, a book recommendation company first acquired by Goodreads and later Amazon. He received his Ph.D. in statistics from UC Berkeley, where he worked on statistical learning theory and natural language processing. David received a Master of Science in applied mathematics, with a focus on computer science, from Harvard University, and a Bachelor of Science in mathematics from Yale University. He is currently based in Toronto.
Talk Track: Technical / Research
Talk Technical Level: 5/7
Talk Abstract:
We will present BloombergGPT, a 50 billion parameter language model, purpose-built for finance and trained on a uniquely balanced mix of standard general-purpose datasets and a diverse array of financial documents from the Bloomberg archives. Building a large language model (LLM) is a costly and time-intensive endeavor. To reduce risk, we adhered closely to model designs and training strategies from recent successful models, such as OPT and BLOOM. Nevertheless, we faced numerous challenges during the training process, including loss spikes, unexpected parameter drifts, and performance plateaus.
In this talk, we will discuss these hurdles and our responses, which included a complete training restart after weeks of effort. Our persistence paid off: BloombergGPT ultimately outperformed existing models on financial tasks by significant margins, while maintaining competitive performance on general LLM benchmarks. We will also provide several examples illustrating how BloombergGPT stands apart from general-purpose models.
Our goal is to provide valuable insights into the specific challenges encountered when building LLMs and to offer guidance for those debating whether to embark on their own LLM journey, as well as for those who are already determined to do so.
What you’ll learn:
You’ll learn about the challenges that are often faced when designing and training LLMs, as well as some approaches to address these challenges. You’ll also see how domain-specific datasets can benefit LLMs.
Presenter:
Jesse Cresswell, Senior Machine Learning Scientist, Layer 6 AI
About the Speaker:
Spoke at TMLS 2022 about Privacy and Fairness. I lead a team of ML scientists at TD’s research lab. My team focuses on Trustworthy AI.
Talk Track: Advanced Research & Technical
Talk Technical Level: 6/7
Talk Abstract:
Neural networks are infamous for making wrong predictions with high confidence. Ideally, when a model encounters difficult inputs, or inputs unlike data it saw during training, it should signal to the user that it is unconfident about the prediction. Better yet, the model could offer alternative predictions when it is unsure about its best guess.
Conformal prediction is a general purpose method for quantifying the uncertainty in a model’s predictions, and generating alternative outputs. It is versatile, not requiring assumptions on the model and being applicable to classification and regression alike. It is statistically rigorous, providing a mathematical guarantee on model confidence.
And, it is simple, involving an easy three-step procedure that can be implemented in 3-5 lines of code. In this talk I will introduce conformal prediction and the intuition behind it, along with examples of how it can be applied in real-world usecases.
What you’ll learn:
Conformal prediction is a mathematical/statistical procedure which may be difficult to understand for less technical audience members. In my talk I will cut through the mathematics to provide the intuition of conformal prediction using visual aids and real-world examples. The audience will take away several settings where they ca immediately apply the technique.
Presenter:
Bharat Venkitesh, Senior Machine Learning Engineer, Cohere
About the Speaker:
Bharat Venkitesh is a senior machine learning (ML) engineer at Cohere, focused on model compression, model efficiency, and inference optimizations. Previously, Bharat was a ML research engineer at Huawei, Noah’s Ark Lab, Montreal, where he worked on scaling up model training and model compression for natural language processing and speech recognition applications on edge devices. He has a master’s degree from University of Waterloo and a bachelor of technology degree from Indian Insitute of Technology Hyderabad.
Talk Track: Technical / Research
Talk Technical Level: 5/7
Talk Abstract:
The rise of transformer-based language models has seen a boom in model sizes, since these models’ performance scales extremely well with size. With this comes the challenge to develop solutions to make inference on these models more efficient. We’ll show how these behemoth multi-billion-parameter models are optimized for production and how the inference tech stack is established. We’ll cover the key ingredients in making these models faster, smaller, and more cost-effective, including model compression, efficient attention, and optimal model parallelism.
Presenter:
Ezequiel Lanza, AI Open-Source Evangelist, Intel
About the Speaker:
Passionate about helping people discover the exciting world of artificial intelligence, Ezequiel is a frequent AI conference presenter and the creator of use cases, tutorials, and guides that help developers adopt open source AI open source tools.
Talk Track: Technical
Talk Technical Level: 3/7
Talk Abstract:
OpenFL is a Python 3 framework for Federated Learning. Designed to be flexible, extensible and easily learnable tool for data scientists is a community supported project that enables organizations to collaboratively train a model without sharing sensitive information, originally developed by Intel Labs and the Intel Internet of Things Group. The team would like to encourage any contributions, and aims to be community-driven. It employs narrow interfaces and allows running all the processes within Trusted Execution Environments (TEE) to provide confidentiality of data and models, integrity of computation, and enable attestation of compute resources.
To protect information while still leveraging ML models to automate scan analysis, Intel Labs and UPenn used data from over 71 medical institutions to apply and test the efficacy of federated learning for brain tumor edge detection. Click for PDF
With FL hardware and software, sensitive data can be secured at the source, while the AI model still benefits from a larger data set.
Learn how you can adopt, contribute and secure federated learning.
What you’ll Learn:
How Federated Learning can be implemented in a secure way.
Presenter:
Suhas Sreehari, Staff Research Scientist, Oak Ridge National Lab
About the Speaker:
Dr. Suhas Sreehari is a Staff Research Scientist at the Oak Ridge National Lab in the US. His research interests are in explainable AI models, signal processing, and Bayesian estimation and optimization. He is the recipient of several important research awards such as the SIAM Imaging Sciences Best Paper Prize, the IEEE Young Author Award, and the Presidential Scholar Award. Dr. Sreehari holds a PhD in electrical engineering from Purdue University, and is a Senior Member of the IEEE.
Talk Track: Technical
Talk Technical Level: 5/7
Talk Abstract:
With large language models such as GPT-4, LaMBDA, and LLaMA, research and commercial interest in natural language understanding (NLU), analysis, and generation has soared. However, these models typically require hundreds of billions of parameters and enormous volume of data to perform self-supervised learning. Furthermore, there is a lack of control and explainability in these models.
Although not to directly compete with GPT-level performance, in this work, we propose a ground-up approach to NLU that is based on graphs. The motivation for this research is two-fold:
The first point is useful to understand the sparsity of language data that are currently used in LLMs. Although we do not directly answer this, what we are discovering is whether the LLMs can learn effectively from much smaller data if the data are projected to lower-dimensional manifolds that are related to fundamental language concepts such as grammar and social constructs. The second point is of paramount importance in light of ethical and factual concerns currently associated with LLMs.
Here is a succinct outline. We formulate an algorithm that is able to read through text sequentially, updating the core graph. The graph nodes represent characters, objects, ideas, and events. The edges represent their relationships. Unlike traditional graphs, edges are time-stamped to account for dynamic behavior and relationships that could be time-varying and/or causal. We then learn collectively-exhaustive contexts that span the entire text. We also discuss loss functions for the context learning. We then use our proposed “context mixture models” (CMMs) and to perform robust Bayesian inference. We finally demonstrate the promise of such an approach through text summarization, question answering, and explainability analysis.
What You’ll Learn:
This work, to the best of our knowledge, is new. We are proposing “context mixture models”, which is similar to Gaussian mixture models, but is more flexible because of the looser parameterization of the probabilistic models defining context in language. Further, we approach a well-explored field from a different angle. Instead of having large language models train on huge datasets, we aim to find a lower dimensional subspace governed by grammar rules, context, and social conditioning — to make these models way lighter, but still effective.
Presenter:
Jekaterina Novikova, Director of Machine Learning Research, ARVA / Cambridge Cognition
About the Speaker:
Jekaterina is leading Machine Learning research efforts at AI Risks & Vulnerability Aliance and at Cambridge Cognition. Over the last decade Jekaterina worked in the intersection of Language Technology and Machine Learning, for interdisciplinary applications including Natural Language Generation, Machine Learning for Healthcare, Spoken Dialogue Systems, and Human-Robot Interaction.
Now, her research focus is on trustworthy machine learning: she works on ensuring qualities such as fairness, transparency, privacy, and robustness, and implementing them in real-world use cases. Jekaterina has been recognized as one of the “Top 25 Women in AI in Canada”, “30 Influential Women Advancing AI in Canada”, and received the “Industry Icon” award from the Applied Research in Action program at the University of Toronto. She has several best research paper nominations at international academic conferences. More information on her work can be found at jeknov.github.io
Talk Track: Technical
Talk Technical Level: 4/7
Talk Abstract:
Language models have become essential components in various NLP applications, such as machine translation, sentiment analysis, and question-answering systems. However, there is a growing concern that these models may perpetuate and amplify bias present in the data on which they were trained. Detecting bias in large language models is crucial because biased language models can amplify existing societal biases, such as racism and sexism, or perpetuate harmful stereotypes. When not properly controlled, the use of biased language models in decision-making processes, such as those used in loan applications, healthcare, hiring, and criminal justice, can result in unfair outcomes.
Bias in NLP models can arise from historical patterns, biases, and stereotypes encoded in the training data. For example, a study of classifiers trained to predict the occupation of a person based on their biography was more likely to predict “surgeon” for men’s biographies compared to women’s. In our work, we focus here on masked language models such as BERT that are fine tuned to perform specific tasks. We want to understand how pretraining and fine tuning each contribute to the observed biases of these models.
Such an understanding not only provides practitioners actionable guidance on how to responsibly use and finetune pretrained language models, it also highlights ‘infeasibility conditions’, i.e. situations where it is impossible to fully mitigate fairness-specific risks of downstream applications of a model.
What You’ll Learn:
This is and extremely hot topic right now (think of the “Pause Giant AI Experiments” open letter and similar high-scale concerns of AI advancements) and it was not properly presented and discussed at TMLS previosly
Presenter:
David Emerson, Applied Machine Learning Scientist, Vector Institute
About the Speaker:
David Emerson is an Applied Scientist at the Vector Institute working on applied research and development in natural language processing (NLP) and federated learning. He obtained his PhD in Applied Mathematics at Tufts University in 2015 working on numerical methods for partial differential equations. Since graduating, he has focused on machine-learning research and worked in various applied research and development positions at companies in Toronto. His areas of interest include NLP, generative modeling, federated learning, and machine learning theory, especially optimization.
Talk Track: Technical
Talk Technical Level: 6/7
Talk Abstract:
As the generalizability of large language models (LLMs) rapidly expands, prompting and prompt design have become an increasingly important field of study. For expressive LLMs, well constructed prompts can elicit remarkable performance on a wide variety of downstream tasks. However, there is significant variance in such performance depending on prompt structure, and manual optimization of prompts is often quite challenging. In this talk, we’ll discuss several state-of-the-art prompt optimization techniques, including both discrete and continuous approaches. Continuous prompt optimization approaches fall under the more general category of parameter efficient fine-tuning (PEFT) methods. We’ll briefly consider two such approaches in the form of Adapters and LoRA, the latter of which produces similar or better performance to full-model fine tuning on many tasks.
What You’ll Learn:
1. Motivations for why parameter-efficient fine tuning (PEFT) methods are important for large language models.
2. An in-depth understanding of current state-of-the-art approaches for PEFT.
Presenter:
Freddy Lecue, AI Research Director, JPMorgan Chase & Co.
About the Speaker:
Dr Freddy Lecue (PhD 2008, Habilitation 2015) is AI Research Director at J.P.Morgan in New York. He is also a research associate at Inria, in WIMMICS, Sophia Antipolis – France. Before joining J.P.Morgan he was the Chief Artificial Intelligence (AI) Scientist at CortAIx (Centre of Research & Technology in Artificial Intelligence eXpertise) @Thales in Montreal, Canada from 2019 till 2022. Before his leadership role at the new R&T lab of Thales dedicated to AI, he was AI R&D lead at Accenture Technology Labs, Dublin – Ireland. Prior joining Accenture he was a research scientist, lead investigator in large scale reasoning systems at IBM Research from 2011 to 2016 Before his AI R&D lead role, Freddy was a principal scientist and research manager in Artificial Intelligent systems, systems combining learning and reasoning capabilities, in Accenture Technology Labs, Dublin – Ireland. Before joining Accenture in January 2016, he was a research scientist and lead investigator in large scale reasoning systems at IBM Research – Ireland.
Talk Track: Technical / Research
Talk Technical Level: 5/7
Talk Abstract:
Financial institutions require a deep understanding of the rationale behind critical decisions, and particularly the predictions made by machine learning models. This is particularly true in high-stake decision-making such as loan underwriting and risk management. This has a led to a growing interest in developing interpretable and explainable Artificial Intelligence (XAI) techniques for machine learning models that can provide insights into the model’s decision-making progress. The development of these models will be crucial for the widespread adoption of machine learning (including deep learning) in finance and other regulated industries. This presentation will first give an overview on the recent progress of adopting and enriching XAI techniques in finance, and then review recent challenges where XAI techniques would need to be pushed further to embed the latest advancement of AI techniques including Large Language Models.
What You’ll Learn:
TBA
Presenter:
Nima Safaei, Senior Data Scientist, Scotiabank
About the Speaker:
Oral talk/presentation in various word-calss AI/ML conferences such as TMLS 2021/2022, CORS/Informs, ICML 2021, NVIDIA GTC 2020/2021
Talk Track: Advanced Technical
Talk Technical Level: 5/7
Talk Abstract:
Explainability is a necessity when applying AI/ML models for high-risk applications in the presence of many exogenous factors, latent variables, and disturbance processes. The good examples are financial, and econometric applications where, regardless of the model performance, the explainability of the outcomes is crucial for storytelling – extracting value from the AI/ML model in terms of some financial metrics and communicating them effectively across the company and with stakeholders.
In complex environments with many exogenous factors, multiple causes and effects are interacting in complex ways through mechanisms that are often only partially understood. When some variables are latent (unobserved), the causal structure imply a complicated set of constraints on the distribution of observed variables.
Explainability in ML is two folded, Casual Explainability (also known as interpretability) and Counterfactual Explainability. While the former addresses ‘why’, the latter addresses ‘how’ small and plausible perturbations of the input variables modify the output? For financial storytelling, the former is much more important than the latter because the reasoning through a causal chain brings more confidence on outcomes. A higher confidence can be translated as a lower (market/capital/operational) risk of the false alarms.
One major challenge in high-risk applications with complex environments is that the explanations provided by machine should be verified by a subject matter expert (SME), otherwise, the outcomes cannot be trusted for further decision making. The question is when we realize that the machine may beat the human by leveraging the information beyond the SME’s knowledge? Should we expect the machine to leverage a level of intelligence which we cannot handle it? In this talk, I try to discuss and answer the above questions through a causal inference lens.
What You’ll Learn:
Explainability in Financial Applications thru. Causal Inference
Presenter:
Alistair Johnson, Scientist, SickKids
About the Speaker:
Dr. Johnson is a Scientist at the Hospital for Sick Children. He received his Bachelor of Biomedical and Electrical Engineering at McMaster University and successfully read for a DPhil at the University of Oxford. Dr. Johnson strives to overcome barriers in data sharing, and his work on the MIMIC databases demonstrates the immense potential of publicly available healthcare data. His research focuses on the development of new data structures, algorithms for deidentification, and new machine learning methodologies for medical data.
Talk Track: Advanced Technical
Talk Technical Level: 5/7
Talk Abstract:
It remains unclear whether large language models (LLMs) trained primarily with general web text are the right tool in highly specialized, safety critical domains such as clinical text. Recent results have suggested that LLMs encode a surprising amount of medical knowledge. To investigate whether smaller domain-specific LLMs retain utility, we conduct an extensive empirical analysis of 12 language models, ranging from 220M to 175B parameters, measuring their performance on 3 different clinical tasks. We show that relatively small specialized clinical models substantially outperform all in-context learning approaches, even when fine-tuned on limited annotated data. Further, we find that pretraining on clinical tokens allows for smaller, more parameter-efficient models that either match or outperform much larger language models trained on general text. We release our trained models used under the PhysioNet Credentialed Health Data license and data use agreement.
What You’ll Learn:
In what cases fine-tuning an LLM to specialized domains (like clinical text) is worthwhile
Presenter:
Dr. Ehsan Amjadian, Head of Data Science, Royal Bank of Canada (RBC)
About the Speaker:
Dr. Ehsan Amjadian is the Head of Data Science at the Royal Bank of Canada (RBC), where he has led numerous advanced AI products and initiatives from ideation to production and has filed multiple patents in the areas of Data Protection, Finance & Climate, and Computer Vision & its applications to Satellite Images. He earned his Ph.D. in Deep Learning & Natural Language Processing from Carleton University, Canada and is presently an Adjunct Professor of Computer Science at University of Waterloo. He is published in a variety of additional Artificial Intelligence and Computer Science domains including Recommender Engines, Information Extraction, Computer Vision, and Cybersecurity.
Talk Track: Advanced Technical
Talk Technical Level: 5/7
Talk Abstract:
Large language models (LLMs) are powerful AI systems that can generate natural and artificial language content such as programming languages for various tasks and applications. However, LLMs also pose significant security and privacy risks, such as leaking sensitive information from their training data, producing unsafe or malicious code, and enabling adversarial attacks by malicious parties. Some examples of adversarial attacks on LLMs are PromptInject, which can hijack the model’s goal or leak its prompt, differentiable language model attack, which can fool text classifiers by fine-tuning a pre-trained language model, and gradient-based attack, which can generate perturbations that exploit the model’s vulnerability. In this talk, we review some of the main challenges and threats associated with LLMs for code and natural language, and survey some of the existing and proposed solutions to mitigate them. We will discuss some of the ethical and legal implications of using LLMs, and suggest some directions for future research and development.
What You’ll Learn:
1. Enumerating Privacy & Security Concerns with Large Language Models (LLMs)
2. Provide some understanding of LLms
3. Talk about Other Risks Associated with the use of LLMs
4. Go through some of the approaches that can mitigate some of these risks
Presenter:
Wendy Foster, Director of Engineering and Data, Shopify
About the Speaker:
Wendy Foster is a Director of Engineering and Data at Shopify, where she leads the teams building merchant analytics, helping entrepreneurs understand and make best, supported data informed decisions for their businesses.
Talk Track: Business & Strategy
Talk Technical Level: 2/7
Talk Abstract:
Data science is a balancing act—math and science have their role to play, but so do art and communication. Storytelling can be the binding force that unites them all. In this talk, I’ll explore how to tell an effective data story and illustrate with examples from our practice at Shopify.
What you’ll learn:
Compelling communication is a core requirement for building data science influence in businesses.
Presenter:
Alik Sokolov, Co-Founder and CEO, Responsibli
About the Speaker:
Alik’s professional background is in AI consulting as a machine learning and data science team leader, and venture capital as a research associate in one of Peter Thiel’s funds. Alik is also a seasoned AI project leader and educator in the machine learning field, having taught and developed the machine learning course at the University of Toronto Master’s of Mathematical Finance program, as well as many workshops and classes around the world. Alik is also a PhD candidate and Vanier Scholar at the University of Toronto, studying applications of machine learning in quantitative finance and he has several publications at the intersection of quantitative finance, AI, and responsible investing.
Talk Track: Executive Business/Strategy
Talk Technical Level: 2/7
Talk Abstract:
It’s no secret that deep learning requires a lot of compute power for both training and inference. Often businesses evaluate the value of machine learning at enterprise from cost perspective, i.e. people (salaries) and infrastructure (cloud spending). Increasingly a lot of organizations adding an environmental mandates to their business goals, such as reducing CO2 emissions from their compute workloads.
In this workshop we are responding to this need. We will be building on top of the previous workshop on optimizing machine learning workloads at TMLS 2022 and demonstrate the reduction of environmental cost of the optimized workloads versus status quo. We will do a survey of the landscape of tools available now, and show some of the free open-sourced tools we built at CentML for energy profiling.
What you’ll learn:
A unique, hands-on industry perspective on the rapidly evolving world of NLP / LLM’s, presented in an executive-accessible format
Presenters:
Miguel Mendez, Generative AI and CCAI Technical Lead, Deloitte | Anand Nimkar, Generative AI Capability Leader, Deloitte | Preeti Shivpuri, Executive Director, Deloitte
About the Speakers:
Miguel Mendez is a Generative AI and CCAI technical leader in Deloitte’s Omnia AI practice where he leads large deliveries across multiple industries
Talk Track: Business Strategy
Talk Technical Level: 5/7
Talk Abstract:
Additional Questions: How dangerous is Generative AI?
What You’ll learn:
Understand where Generative AI fits into business
Presenter:
Bala Gopalakrishnan, Chief Data Officer, Pelmorex Corp
About the Speaker:
Bala is Chief Data Officer at Pelmorex Corp, the parent company of The Weather Network, MeteoMedia, El-Tiempo, Clima, WetterPlus, Pelmorex Data Solutions brands.
Pelmorex is a leading player in Digital Advertising and Weather B2B/B2C solutions. The AI powered products of Pelmorex are used by B2B customers worldwide.
Bala leads the Product, Engineering and technology talent that boasts of Canada’s best GIS, AI/ML, Big Data Engineering and UX talents.
Bala has a Master’s of Science in Computer Engineering from The University of Alabama, and an MBA from Ivey Business School. Prior to joining Pelmorex, Bala lead engineering and software teams at Johnson Controls (Fortune 100) and Eutech Cybernetics which was rated by Gartner as one of the cool companies in smart city solutions. Bala was ranked in the top 50 global data analytics leaders in a survey by Corinium in 2022.
Talk Track: Case Study
Talk Technical Level: 5/7
Talk Abstract:
Weather affects every facet of life. It changes our behaviour patterns of what we buy, what we choose to eat. We have one of the best historical weather data sets going back decades for every point in the globe. Correlating it with sales data and able to scale hundreds of thousands of models at product / store level, we are able to predict weather driven demand for short and long term to help decision makers in supply chain, man power planning and marketing
What you’ll Learn:
Packaging ML in an end to end SAAS platform to solve a key industry problem
Presenters:
Richard Zuroff, Senior Vice President Growth, BlueDot | Alexei Nordell-Markovits Director AI Delivery, Moov AI
About the Speakers:
“Richard is the Senior Vice President of Growth at BlueDot, a global infectious disease intelligence company that uses human and artificial intelligence to detect and assess the risks of emerging and endemic disease threats. Richard began his career at McKinsey where he advised clients on advanced analytics, before founding the strategy and governance practice at Element AI, and continuing to work on democratizing the use of AI at DataRobot. Richard has spoken about the social and economic impact of AI/ML at the World Economic Forum, Partnership on AI, and Brookfield Institute. He is the author of a recent book chapter on Explainable AI for Financial Services, and has published in the Canadian Journal of Law and Technology. Richard holds an MBA, two Law degrees, and a BSc in Cognitive Science from McGill University.
Alexei has been building systems and teams within the context of modern AI for the better part of a decade. He has worked in diverse subfields, including process optimization, Mlops and Responsible AI. He is an AI director at Moov AI where he works with clients helping them discover and then implement AI solutions within business processes. In his spare time, he enjoys reading junky science-fiction.”
Talk Track: Case Study
Talk Technical Level: 3/7
Talk Abstract:
Rapid progress in large language models (LLMs) has ignited interest in a wide array of exciting generative use cases for both consumer- and enterprise-facing applications. As adoption accelerates, it is critical to ensure these systems remain safe and trustworthy. Researchers are exploring technical solutions (such as constitutional AI) but in this talk we discuss two design patterns that involve putting humans-in-the-loop (HITL) of production-ready generative AI systems. HITL approaches to responsible generative AI are well aligned with regulators’ push for human oversight of “high risk” systems and are more likely to help companies create a strategic moat (compared to just using foundational models).
We present two real-world case studies. BlueDot’s global disease surveillance team uses LLMs to identify concerning outbreaks of infectious diseases around the world. This is challenging because each outbreak is unique and has potentially conflicting media reports, but it is essential that an LLM-based summary be factual and specific – not hallucinated based on past events. This has implications for how internal experts (epidemiologists) use the system to create value and potential IP, and what can be exposed externally to users. Moov.AI needed to develop two question and answer services incorporating a knowledge base and decided to use a LLM to produce the content for these systems. The challenge was to avoid hand-crafting every aspect of the question and answer style while managing the reputational risk that inappropriate or off-style answers would create. The solution involved different HITL at different steps for fine-tuning and data enhancement.
We conclude by summarizing the lessons learned from these two use cases and the implications for other organizations that are considering generative AI use cases.
What you’ll learn:
Real-world case studies; insights from practitioners
Somayeh Sadat, Assistant Director, CARTE, University of Toronto | Alex Olsen, Senior Research Associate, CARTE, University of Toronto
About the Speakers:
At CARTE, Alex collaborates with faculty, industry partners, and students to bring the latest advances in AI to a broad audience. He has contributed to machine learning-based research in a wide variety of fields, and runs training for learners at every level.
Somayeh Sadat is the assistant director at the Centre for Analytics and Artificial Intelligence Engineering (CARTE), University of Toronto. In this role, she facilitates and oversees partnerships with industry and government to collaboratively conduct state-of-the-art research and translate and commercialize effective solutions in analytics and artificial intelligence, where applicable with leveraged funds. She also helps nurture the next generation of engineering talent through facilitating experiential learning opportunities with external partners. Somayeh holds a Ph.D. from the University of Toronto, and has over fifteen years of experience in research and education, business development, and consulting.
Talk Track: Case Study
Talk Technical Level: 1/7
Talk Abstract:
The rapid advancements in artificial intelligence (AI) have given rise to a new era of large language models (LLMs) such as ChatGPT and GPT-4. These sophisticated AI systems are capable of understanding and generating human-like text, making them invaluable tools for communication and problem-solving. However, the complexity of these models often leaves the general public feeling overwhelmed and unsure of their implications. This talk aims to demystify LLMs, explaining their inner workings in an accessible manner, while highlighting their potential impact on our society.
The presentation will begin by offering an introduction to the concept of LLMs, touching on their history and development. We will explore the key principles behind these AI systems, such as neural networks and natural language processing, in a way that is easy to grasp for a non-technical audience. The primary focus will be on ChatGPT and GPT-4, two of the most advanced and widely-discussed models currently available.
Next, we will delve into the practical applications of these LLMs, demonstrating their versatility in tasks ranging from content creation to customer support. By showcasing real-world examples, attendees will gain a better understanding of how these AI models can transform various industries and streamline everyday tasks.
Finally, we will address some of the ethical and societal implications brought about by the increasing adoption of LLMs. Topics such as data privacy, algorithmic bias, and the potential for AI-generated misinformation will be discussed, providing a balanced perspective on the benefits and challenges of these revolutionary technologies.
By the end of this talk, attendees will have a clearer understanding of large language models and their impact on our lives. Join us for an enlightening journey into the world of AI communication, where we will explore the potential of ChatGPT, GPT-4, and other LLMs to shape our future interactions with technology.
What You’ll Learn:
I have experience giving this talk to a completely non-technical audience, for which I’ve had great feedback. The aim of this talk is for people to walk away feeling like they actually understand the functionality of LLMs, without needing a technical background.
Michel Dubois, Principal Director, Experimental Development, Mila | Cameron Schuler Chief Commercialization Officer & VP, Industry Innovation, Vector Institute | Kirk Rockwel, Chief Operating Officer, Alberta Machine Intelligence Institute | Alex LaPlante, Interim Head, Borealis AI, Royal Bank of Canada (RBC)
About the Speakers:
As Director, AI Activation at Mila, the Quebec Artificial Intelligence Institute, Michel Dubois actively participates in the development of AI for the benefit of all. He holds a master’s degree in mathematics and is currently a PhD candidate in engineering (machine learning). He is also the author of a patent on the mathematical optimization of high-bandwidth signal switching. Over the past 28 years, Michel Dubois has consolidated his experience in several aspects of machine learning and artificial intelligence. Before joining Mila, Michel Dubois held the role of Associate Partner at IBM, and, in his previous role, he was Vice President of Artificial Intelligence at Newtrax, a Montreal start-up that quickly experienced international success, reaching an average annual revenue growth of 80% and opening offices in Santiago, Perth, Moscow and London.
Cameron Schuler is the Chief Commercialization Officer & Vice President, Industry Innovation at the Vector Institute. He is the former Executive Director of Amii, where, for 8 years, he led one of the top-ranked Machine Learning and AI groups in the world. Cameron’s multifaceted career has covered finance, business & product development, consumer products, IT and general management from start-ups to mature companies. His industry experience includes Alternative Energy, Banking, Consumer Products, Information Technology (Consumer and Enterprise), Investment Sales and Trading, Life Sciences, Manufacturing, Medical Devices, Oil & Gas, and Oil & Gas Services. Roles have included COO, CFO, President and CEO and he was COO & CFO of a food manufacturer whose products lead to sales of over 250 million units.
Kirk Rockwell is a project and operations manager with more than 20 years of experience managing technology and innovation initiatives, across a broad range of sectors in partnership with academia, all levels of government, and some of Canada’s largest companies.
Within an arm’s length government agency, he had responsibility for investment of public funds into research and innovation projects across multiple sectors including energy, clean tech, agriculture and health, utilizing technologies like AI and Machine Learning, Nanotechnology and Genomics.
He has experience with numerous public/private partnerships (P3s) including membership on advisory and governance boards where the stakeholders include large multinational corporations, the governments of Alberta and Canada, universities, and other small or medium-sized companies.
He has a Diploma in Environmental Technologies, a Bachelor of Science Degree in Environmental Sciences, and a Master’s of Business Administration Degree with a specialization in Innovation and Entrepreneurship.
Kirk is currently the VP, Public Strategy and Gants at the Alberta Machine Intelligence Institute (Amii) in Edmonton, Alberta, Canada. He resides in Edmonton with his wife and two daughters.
Alex LaPlante is currently the interim Head of Borealis AI, RBC’s R&D lab for artificial intelligence, where she and her team build and deploy leading-edge AI solutions to complex business problems found across the enterprise. Previously, Alex led Borealis’ business development and product management teams, which saw her design the company’s long-term product vision and build strategic partnerships with RBC stakeholders. Before joining Borealis, Alex held a number of leadership roles at the intersection of technology and finance. She holds a PhD in operations research from the University of Toronto.
Talk Track: Panel Discussion
Talk Technical Level: TBA
Talk Abstract:
TBA
What You’ll Learn:
TBA<
Presenter:
Vanessa Pizante, Data Scientist, Sobeys
About the Speaker:
I am a data scientist at Sobeys working on the personalization team to help deliver personalized offers to customers through our Scene+ rewards program. My academic background is more mathematics focused, as I have an undergraduate degree in Applied Mathematics from the University of Calgary and a Masters in Statistics from the University of Toronto. During my career, I’ve also learned a lot about software development, design and MLOps; and enjoy combining these newer skills with my math-focused education to build effective and long-lasting machine learning and measurement frameworks.
Talk Track: Case Study
Talk Technical Level: 3/7
Talk Abstract:
Personalization programs can have a significant impact on a business, but measuring their effectiveness can be challenging. Our team at Sobeys faced the task of building a system that could communicate the value of our personalized offers to external stakeholders and inform our internal team about what works and what does not. This required us to develop an experimentation and measurement platform that could measure the effects of a hierarchy of treatment variants, compute these effects across different customer segments, and aggregate the results consistently. In this talk, we will share our experience in creating a configurable, automated process that can handle repeated experimentation and measurement.
Simple frequentist A/B measurement approaches proved to be insufficient when measuring the effectiveness of our complex hierarchy of treatment variants, across an independent hierarchy of customer segments. The issue is that contradictory results can regularly arise when slices of hierarchical data are tested independently. For instance, two segments can show a positive lift individually, but the overall lift across both segments could be negative (this phenomenon is known as Simpson’s Paradox). Multilevel models can be leveraged in both the frequentist and Bayesian paradigm to help address this. However, in our use case the Bayesian paradigm offered several useful advantages including the capability to explicitly define the structure of the hierarchy via prior specification and better handling of smaller customer segments.
In general, attendees of this talk can expect to learn:
What you’ll learn:
Bayesian statistics at scale and taking an automated approach to complex Bayesian modelling is a topic I struggled to find much about. Also dealing with nested experiments in a robust manner, at scale, within this automated framework.
Presenter:
Ashley Varghese, Data Scientist, Canadian National Railways (CN)
About the Speaker:
Ashley is a data scientist at Canadian National Railway. She works in the automated inspection program overseeing the development and retraining aspects for the inspection of rail cars. She has over 12 years of research experience in computer vision and deep learning. Her research papers were published in multiple international conferences and journals. She has previously worked as an AI Scientist with Qii.AI and as a researcher with TCS Innovation Lab. She holds an MTech in Computer Science from the International Institute of Information Technology, Bangalore.
Talk Track: Case Study
Talk Technical Level: 4/7
Talk Abstract:
CN is one of the organizations which realized the importance of AI very early and adopted some of the latest technologies in railroad operations. We at CN have been evaluating and adopting some of the latest approaches to carry out automated inspections of railcars that improve overall operational efficiency. The automated inspection of the railcars is achieved by leveraging computer vision and machine learning techniques. For automated inspection of railcars, the train passes through the portals which have equipped with two generations of cameras and are positioned to capture all sides of the railcar including the undercarriage. Once these images are captured, they are sent to the inference engine to detect the defects. However, developing machine learning pipelines and training robust models comes with its own challenges.
In this talk, I would be covering how computer vision and machine learning techniques are used for developing some of the use cases, and various challenges associated with developing these machine learning pipelines. One of the key challenges is selecting the right data for training which must be representative of the actual data from the portal. Self-supervised learning-based techniques are adopted to identify the unique samples from the pool of unlabeled datasets.
Another challenge associated with the model performance is image quality. The quality of the image captured at the portals is affected by different weather and lighting conditions. In addition to these, there are several other challenges such as subjectivity in defect classification, lack of samples for the defective class and imbalanced datasets, etc. I will pick one of the use cases we developed and then go over discussing some challenges faced and how we tackled these challenges to improve the performance of the use cases. By choosing the right data and computer vision approaches, it is possible to develop effective solutions for the automated inspection of cars.
What you’ll Learn:
This talk provides insight into how computer vision and machine learning are applied in the railroad domain for the inspection of rail cars. It includes the challenges, strategies, and practical considerations that are associated with developing an ML-based automated inspection pipeline for a real-world, safety-critical application.
Presenters:
Yatong Li, Managing Director, Sixty Degree Capital | Ryan Shannon, Investor, Radical Ventures | Vik Pant, PhD, Founder, Synthetic Intelligence Forum | Michelle Yu, Investor, Georgian
About the Speakers:
With over 10 years of experience in technology, finance and investment banking, Yatong leads Sixty Degree Capital’s investment in the technology space. He joined Sixty Degree Capital as an Associate in 2017 and quickly rose through the ranks to become a Managing Director on the technology investment team where he focuses on enterprise software opportunities. During his six year tenure at the firm, he managed to build the firm’s portfolio in the tech space with over 12 portfolio companies — including GrubMarket, MioVision, Pragma, DataGrail, Arctic Wolf, MacroMeta, Tact.AI, Lime, and Schrödinger.
Ryan Shannon is a member of the investment team at Radical Ventures.
Prior to joining Radical, Ryan was a Private Equity Investor at TPG Capital in San Francisco, where he focused on Leveraged Buyouts, Corporate Carve-outs, and Take-privates of North American businesses. Previously, Ryan worked as an Investment Banker in the Financial Sponsors group at Barclays in Los Angeles.
Ryan received an HBA from the Ivey Business School at Western University, where he graduated as an Ivey Scholar, and an MBA from Harvard Business School.
Vik is the Founder of the Synthetic Intelligence Forum, a global community of practice focused on Data Science and applied AI. SIF is a professional networking, knowledge sharing, and collaborative learning hub for professional data scientists, AI researchers, and ML engineers.
He is an Adjunct Professor in the Faculty of Information (iSchool) at the University of Toronto and the Department of Geography, Environment, and Geomatics at the University of Ottawa.
Vik earned a doctorate from the Faculty of Information (iSchool) in the University of Toronto where his thesis was unanimously accepted by the examination committee As-Is and without any changes. His scholarship is focused on game-theoretic optimization of strategic coopetition in complex multi-agent systems. His research has been published in numerous peer-reviewed scholarly journals and he has also presented his academic research at refereed scholarly conferences and juried workshops.
Michelle is an Investor at Georgian, a growth-stage firm investing in high growth technology companies that harness the power of data in a trustworthy way. Michelle focuses on sourcing investment opportunities, investment due diligence, and advising companies post-investment on go-to-market, market analysis and product strategy. Prior to joining Georgian, Michelle worked on the investment team at Vectr Ventures in Hong Kong and for BDC Capital in Montreal. Michelle has a Bachelor of Commerce with a Double Major in Finance & International Management, with a Minor in Political Science, from McGill University.
Talk Track: Panel Discussion
Talk Technical Level: 3/7
Talk Abstract:
As a Toronto based Canadian VC focusing on global investments, Sixty Degree Capital invests in software that’s transforming industries, and the digital infrastructure that supports it. We have invested in lots of exciting portfolio companies including Arctic Wolf in the Cybersecurity space, DataGrail innovating on Data Privacy, Paperspace providing MLOps platform, MacroMeta solving Edge Computing bottlenecks, Pragma work in the Gaming Infrastructure space, Tact.ai and Radius Agent as Vertical SaaS products etc.). We believe AI/ML is a fundamental building block for next generation software, so Sixty Degree has spent lots of time to research, source, invest in the sub sectors. While we’re a direct VC fund, we’ve allocated capital to become LPs of notable venture funds such as Lightspeed, a16z, NEA, Tiger, NfX, Initialized, Foundation Capital and Upfront Ventures, which gives us a special information advantage on what are the top tier Silicon Valley VCs are investing.
What You’ll Learn:
Insider information on how investors are allocating capital in AI/ML now
Presenter:
Rokshana Stephny Geread, Data Scientist – Computer Vision, Biosymetrics
About the Speaker:
Steph joined BioSymetrics in January 2021, with a background in biomedical ECE engineering (MASc) and electrical engineering (BEng). She has diverse work experience, starting with terminal web development for Bell Canada before building, validating, and delivering a production-ready machine learning model for breast cancer imaging diagnosis. This last project was a collaboration with University Health Network and St Michael’s Hospital, Toronto, and resulted in three journal articles.
“I transitioned from the telecommunications industry to healthcare because I wanted to make a positive impact on patients’ quality of life. I wanted to contribute to bridging the gap between technology and healthcare, which has the potential to enhance the overall healthcare system.
Talk Track: Case Study
Talk Technical Level: 4/7
Talk Abstract:
Ebrafish are a popular model organism in developmental biology, molecular genetics, and toxicology studies due to their small size, low breeding costs, transparent embryos, morphological identification, and similarity of their genome to the human genome. They are particularly well suited for large scale phenotypic screening in drug discovery studies, where observable characteristics (phenotypes) can be monitored due to biological effects from a small molecule or gene modification (via CRISPR gene editing). Using imaging and computer vision analysis, some zebrafish phenotypes that can be monitored include morphology, heart rate, and ejection fraction (the amount of blood the heart pumps). To scale up these experiments, some logistical challenges need to be addressed. First, standardizing the positioning of live zebrafish embryos is necessary to produce consistent imaging data and maximizing the value of experiments. Second, up to thousands of images need to be labelled for specific organs to analyze specific phenotypes, a time-consuming task if performed manually. To address these challenges, we have developed a comprehensive zebrafish phenotyping assay that uses custom 3D-printed mounting platforms to affix embryos in a standardized orientation for consistent image and video acquisition. We then developed a computer vision analysis pipeline that couples a fine-tuned instance segmentation model with downstream phenotyping analysis, providing quantitative. The model, based on the Detectron2 segmentation model architecture, is trained to automatically detect individual zebrafish embryos and its organs from images generated by our screening assay. We applied our approach to study several genes associated with cardiovascular disease using CRISPR injections, identifying subtle cardiac abnormalities. Our results demonstrate highly accurate segmentation on validation data, achieving average precision scores of 60% for the ventricle, and over 80% for other major organs. Our phenotyping platform is designed to close the gap between gene and phenotypes, thereby empowering and accelerating drug discovery programs with a higher probability of clinical translation.
What you’ll Learn:
Using AI driven technology and image processing to analyze cardiac phenotypes in zebrafish images and videos.
Presenter:
Zahra Kharal, Data Scientist (Principal) Venterra Realty & ML Engineer
About the Speaker:
Published 12 journal papers and more than 20 conference papers. I was a women in engineering rising star: here. Lastly, I was responsible for building the AI/ML department at Venterra, from it’s conception to where it is today; in less than 3 years, we have implemented many models in production and developed a propriety in-house app using AI that allows us to invest considerably better.
Talk Track: Case Study
Talk Technical Level: 3/7
Talk Abstract:
The multifamily real estate industry has witnessed an increased adoption of machine learning (ML) in recent years; allowing it to recover from its initial relatively slow adoption of ML technology and catch up to other industries in the field. The use of ML started to move into the multifamily real estate industry, offering new and innovative ways to analyze and operate properties, providing greater efficiency and optimization to investors, owners, and property managers. This presentation will showcase a collection of innovative ML case studies and application projects in multi-family real estate at Venterra Realty.
One case study at Venterra where ML has shown its value was in predicting renter trends and tenure tendencies. At Venterra, we started 20 years ago, capturing and structuring data, that we thought would eventually be useful for this purpose. By utilizing the 20 years of data that Venterra had, including 70+ features, Venterra built ML algorithms and pipelines that can more accurately help in identifying tenant behavior patterns, such as rent-paying habits and the likelihood of renewing a lease. This accumulative data was then used to make occupancy forecasts months in advance. The results of these models were used to understand tenant behavior better, improve retention rates and help planning and budgeting. An interesting question arose from this model: If we give the model results to the property managers, how will it influence their behavior in regard to tenant dealings.
Another interesting application of ML at Venterra was identifying maintenance needs and developing models for preventative maintenance. By analyzing data from millions of maintenance requests and several other sources, over many years, Venterra used NLP and classification techniques to predict request categories and thus assign these requests to the correct department in almost real time. Additionally, using deep-learning time-series models Venterra was able to predict when repairs will be necessary and schedule them proactively, while also predicting the budget required for these repairs.
Through these case studies and application projects, attendees will gain an understanding of how ML is influencing operations at Venterra. They will see to succeed in ML endeavours, data strategies and proper data collection methodologies needs to be made decades in advance, and learn how pipelines need to be developed to successfully put the models in production. The benefits that Venterra has received since then will also be briefly discussed. Ultimately, this presentation will inspire traditionally non-tech industries to explore the potential of AI and ML in their own operations, and will prepare them for what’s needed in order to succeed.
A bit about Venterra: Venterra specializes in the identification, development, finance, acquisition, and management of multi-family residential communities in the United States. We manage a portfolio of multi-family real estate assets totaling over $4.7 billion in value and have completed more than $8.7 billion of real estate transactions. Venterra has differentiated itself from its competitors by deploying industry leading technology to the benefit of both residents and investors. This technology includes a comprehensive suite of entirely web-enabled applications covering rent management, pricing optimization, purchase order management, utility billing, work order and capital expenditure management; an entire suite of back-end financial applications and customized workflows; and reporting and more recently the adoption of ML.
What You’ll Learn:
Through the case studies and application projects at Venterra Realty, attendees will gain an understanding of how ML is transforming the operations at Venterra and the multi-family real estate industry at large. They will learn specifically what data problems are encountered in the real estate domain, how the models were developed and what pipelines needed to be developed to successfully put the models in production. Issues encountered, unique to this industry, will also be discussed.
Presenter:
Adhithya Ravichandran, Senior Engineer, ML Platform, Loblaw Digital
About the Speaker:
Adhithya (Adhi), is a Senior Engineer in the Machine Learning (ML) Platform team at Loblaw Digital (LD), where he helps build a self-service platform to provide the best development experience that empowers Data Scientists to productionalize ML services. Previously he has held a range of positions spanning analytics, Machine Learning, data strategy, and building a foundational data stack across a range of companies – early stage, Series B-C startups and Canada’s largest banks. Outside of work he spends time practising Muay Thai and yoga, as well as advising a private equity firm in Latin America.
Talk Track: Case Study
Talk Technical Level: 5/7
Talk Abstract:
At Loblaw Digital, ML models play a crucial role in every part of our business – from helping customers search for products they want, providing them with personalized shopping experiences, ensuring the orders they place get to them on time. Behind the scenes, our MLEs are constantly developing, and improving the models that power our business. Furthermore, more use-cases and teams of MLEs are being stood up. To facilitate a high volume of ML models throughout their lifecycle, we have invested in building an ML observability stack in our ML platform. In this talk, we discuss how our centralized observability stack – in-house inference logging, along with Snowplow analytics, which captures user behaviour data. This stack enables our MLEs to have visibility into how ML models influence user interaction, and eventually, revenue in our e-commerce sites. We further discuss how these tools help us enable model improvements and drive real world business outcomes. We also discuss ideas for future feature improvements and additions to our observability stack.
What You’ll Learn:
Investing in building and customizing model observability helps measure performance and ROI of ML, which in turn leads to continuous ML improvements and builds confidence and trust from stakeholders
Presenter:
Arsene Fansi Tchango, Senior Applied Research Scientist, Mila – Quebec Institute in Artificial Intelligence
About the Speaker:
Arsene Fansi Tchango is a senior applied AI scientist at Mila, the Quebec AI Institute. His topics of interest are reinforcement learning, natural languange processing, and the application of deep learning techniques on graphs. Before joining Mila, Arsene obtained his PhD at INRIA, the French National Institute for Research in Digital Science and Technology and spent more than 5 years working in the industry.
Talk Track: Workshop
Talk Technical Level: 4/7
Talk Abstract:
The automation of the medical evidence acquisition and diagnosis process has recently attracted increasing attention in order to reduce the workload of doctors and democratize access to medical care.
However, most works proposed in the machine learning literature focus solely on improving the prediction accuracy of a patient’s pathology. We argue that this objective is insufficient to ensure doctors’ acceptability of such systems.
In their initial interaction with patients, doctors do not only focus on identifying the pathology a patient is suffering from; they instead generate a differential diagnosis (in the form of a short list of plausible diseases) because the medical evidence collected from patients is often insufficient to establish a final diagnosis. Moreover, doctors explicitly explore severe pathologies before potentially ruling them out from the differential, especially in acute care settings.
Finally, for doctors to trust a system’s recommendations, they need to understand how the gathered evidences led to the predicted diseases. In particular, interactions between a system and a patient need to emulate the reasoning of doctors.
In this talk, we propose to model the evidence acquisition and automatic diagnosis tasks using a deep reinforcement learning based approach that considers three essential aspects of a doctor’s reasoning, namely generating a differential diagnosis using an exploration-confirmation approach while prioritizing severe pathologies. We discuss metrics for evaluating interaction quality based on these three aspects, and demonstrate that the proposed solution performs better than existing models while maintaining competitive pathology prediction accuracy.
What you’ll learn:
Integrating doctor’s reasoning in a Machine Learning system
Presenter:
Shagun Sodhani, Research Engineer, Meta AI
About the Speaker:
Research Engineer at Meta AI, previously at Mila and Adobe Research.
Talk Track: Workshop
Talk Technical Level: 7/7
Talk Abstract:
PyTorch is one of the most popular machine learning frameworks and its latest iteration, PyTorch 2.0, was just released. This workshop will focus on a deep dive into the internals of PyTorch 2.0 to understand where the performance benefits come from. It will be useful for both practitioners and researchers as it will improve their knowledge of PyTorch and enable them to extend PyTorch for their (complex) usecases.
Note that I submitted a talk proposal by the title “”PyTorch 2.0 – Why Should You Care”” that focuses on the high-level API and does not go into implementation details. This workshop is much more technical than the talk.
What you’ll learn:
Understand the behind-the-scenes, nitty-gritty details of what makes PyTorch 2.0 so performant.
Presenters:
Phillip Schmid, Technical Lead, Hugging Face / Andrew Jardine, GTM Executive, Hugging Face
About the Speakers:
Philipp Schmid is a Technical Lead at Hugging Face with the mission to democratize good machine learning through open source and open science. Philipp is passionate about productionizing cutting-edge & generative AI machine learning models.
Talk Track: Workshop
Talk Technical Level: 5/7
Talk Abstract:
Learn how to build AI-powered experiences leveraging the latest technology in Machine Learning and Generative AI with Hugging Face and Amazon SageMaker. Hugging Face has become the central hub for Machine Learning, with more than 100,000 free and accessible machine learning models to process and generate text, audio, speech, images and videos.
We will discuss how over 15,000 companies are using Hugging Face to build AI into their applications, and participants will learn how they can do the same easily, leveraging Hugging Face open source models with the enterprise compliant environment of Amazon SageMaker.
What you’ll learn:
Learn how to build AI-powered experiences leveraging the latest technology in Machine Learning and Generative AI with Hugging Face and Amazon SageMaker
Presenter:
Denys Linkov, ML Lead, Voiceflow
About the Speaker:
Denys is the ML lead at Voiceflow focused on building the ML platform and data science offerings. His focus is on realtime NLP systems that help Voiceflow’s 60+ enterprise customers build better conversational assistants. His role alternates between product management, ML research and ML platform building. Previously he worked at large global bank as a senior cloud architect.
Talk Track: Workshop
Talk Technical Level: 6/7
Talk Abstract:
You see a demo of a Large Language Model (LLM) and want to integrate it into your product, but what’s next? In this workshop we’ll go through both the product planning and technical implementation of integrating a LLM into your product. This will include performance monitoring, integrations, testing, fine tuning, and the various strategies we experimented with at Voiceflow. We’ll share some of the challenges and solutions we faced, and each module will feature a notebook to follow along with.
What you’ll learn:
Real world case study, clear actionable steps, code that is similar to that running in production.
Presenter:
Moez Ali, Data Scientist, PyCaret
About the Speaker:
Moez Ali is an Innovator, Technologist, and a Data Scientist turned Product Manager with proven track record of building and scaling data products, platforms, and communities. Strongly opinionated tech visionary and a thought partner to C-level. Moez is the creator of PyCaret. PyCaret is an open-source, low-code, machine learning library in Python that democratizes the use of data science and machine learning. It is ranked in top 1% of Python software globally, It has 8 million+ downloads, 7K+ GitHub stars, 100+ contributors, and 1000+ citations. He is globally recognized for his open-source work on PyCaret. Moez ios am Keynote speaker and top ten most-read writer in the field of artificial intelligence. He also teaches AI, ML, and NLP courses in MBA program at Cornell University and MMA program at Queens University.
Talk Track: Workshop
Talk Technical Level: 3/7
Talk Abstract:
Accelerate Machine Learning with Pycaret
What you’ll learn:
In this talk, we will explore how PyCaret 3, an open-source machine learning library in Python, can significantly accelerate machine learning workflows. PyCaret 3 offers a low-code approach to building, training, and deploying machine learning models, making it an ideal tool for data scientists and developers who want to focus on the business problem rather than the technical details.
Presenters:
Akbar Nurlybayev, Co-Founder, CentML / Michale Shin, Senior Software Development Engineer, CentML / Yubo Gao, Research Software Development Engineer, CentML & Ph.D Student, University of Toronto / John Calderon, Software Engineer, CentML
About the Speakers:
Akbar, Co-founder of CentML, where we accelerate ML workloads by optimizing models to run efficiently on GPUs without sacrificing model accuracy. Prior to CentML, I lead a data organization at KAR Global, through digital transformation and modernization of their data platform.
Michael is helping CentML to build our open-source machine learning tools, that help ML practitioners to performance profile their models, figure out appropriate deployment targets and measure CO2 impact.
Yubo is responsible for the research aspects of CentML’s open source tools for ML practitioners that let users profile their models, figure out appropriate deployment targets and measure CO2 impact.
John is Software Engineer at CentML working on machine learning profiling tools
Talk Track: Workshop
Talk Technical Level: 4/7
Talk Abstract:
It’s no secret that deep learning requires a lot of compute power for both training and inference. Often businesses evaluate the value of machine learning at enterprise from cost perspective, i.e. people (salaries) and infrastructure (cloud spending). Increasingly a lot of organizations adding an environmental mandates to their business goals, such as reducing CO2 emissions from their compute workloads.
In this workshop we are responding to this need. We will be building on top of the previous workshop on optimizing machine learning workloads at TMLS 2022 and demonstrate the reduction of environmental cost of the optimized workloads versus status quo. We will do a survey of the landscape of tools available now, and show some of the free open-sourced tools we built at CentML for energy profiling.
What you’ll learn:
As described in the abstract, given that no one is thinking about profiling for energy and CO2 impact in ML world, the available tools are not widely used and as a result not very useful.
Presenters:
Ethan C. Jackson, Co-Founder & Research Lead, ChainML | Ron Bodkin Co-Founder & CEO, ChainML | Daniel Kur Machine Learning Scientist, ChainML
About the Speakers:
Ethan is co-founder and research lead at ChainML, a startup focused on building scalable, composable AI systems and applications. He is also a founding member of the Social AI Research Group at the University of Toronto. Ethan was previously an Applied ML Scientist with Vector Institute’s AI Engineering Team, where he served as technical lead for several applied AI projects in collaboration with many industry and public sector partners. Ethan holds a PhD in Computer Science from Western University and trained as a postdoc under the supervision of Graham Taylor at the University of Guelph.
Ron is co-founder and CEO of ChainML, a startup that is delivering Generative AI supercharged by Web3. Ron was previously VP of AI Engineering and CIO at the Vector Institute and before that was responsible for Applied AI in the Google Cloud CTO office. Ron also was co-founder and CEO of enterprise AI startup Think Big Analytics that was acquired by Teradata and VP Engineering at AI pioneer Quantcast. Ron was also co-founder and CTO of C-Bridge Internet Solutions that IPO’d. Ron has a Master’s in Computer Science from MIT and an honors B.S. in Math and Computer Science from McGill University.
Daniel is a Machine Learning Scientist at ChainML, a startup that is delivering Generative AI supercharged by Web3. Prior to that, Daniel worked in the machine learning team at ServiceNow, joining as part of the acquisition of Element AI. Daniel also has experience in finance, starting his career in investment banking working on mergers and acquisitions at TD Securities. Daniel has a Master of Business Analytics and Bachelor of Business Administration from the Schulich School of Business at York University.
Talk Track: Workshop
Talk Technical Level: 5/7
Talk Abstract:
As business interest in generative AI continues to grow, it’s crucial for developers to understand the technical aspects of integrating cutting-edge language models like GPT-4 into business applications with strong requirements for customization, performance, and reliability. In this hands-on workshop, we will present our reusable process for successful integration of GPT-4 into business applications. We will give details about our strategies for prompt engineering, context building, and response evaluation – all with the goal of helping LLMs to give their best possible answers, according to a diverse set of evaluation criteria. As a point of reference, we will share details about our experience developing a highly customized chat experience to the Space and Time Web3 Data Warehouse that includes special features like AI-assisted database querying, data visualization, and time series forecasting.
We will also discuss our process for ongoing development and adaptation of chat applications, with a focus on how we are anticipating the rapid emergence of new tools and frameworks, particularly in the open-source domain. By the end of this workshop, participants will be able to simplify their approach to business chat application development by adopting or extending our process, and they will understand how development processes are likely to evolve with rapidly changing AI and LLM ecosystems.
What You’ll Learn:
You’ll learn how to apply an end-to-end experimental approach to building robust chat systems with GPT-4 at the core. We will go into details about the strenghts and limitations of both proprietary APIs and open-source frameworks, giving a pragmatic view on how to be productive with these tools today.
Presenter:
Sophia Yang, Senior Data Scientist, Anaconda
About the Speaker:
Sophia Yang is a Senior Data Scientist and a Developer Advocate at Anaconda. She is passionate about the data science community and the Python open-source community. She is the author of multiple Python open-source
Talk Track: Workshop
Talk Technical Level: 2/7
Talk Abstract:
Do you use data visualization to understand data and tell stories? Do you need to visualize big datasets? Are you interested in leveling up your visualization skills with HoloViz? HoloViz is a high-level Python visualization ecosystem that gives you the superpower to help you understand your data at every stage needed.
In this talk, you will learn how to build visualizations easily even for big and multidimensional data, how to turn nearly any notebook into a deployable dashboard, and how to build interactive drill-down exploratory tools for your data and models without having to run a web-technology software development project. You will also learn how to turn your dashboard into WebAssembly and run your dashboard entirely in the browser with the magic of Pyodide and PyScript.
What you’ll learn:
A high-level overview of the high-level Python visualization ecosystem HoloViz.
Presenter:
Jim Dowling, CEO, Hopsworks
About the Speaker:
Jim Dowling is CEO of Hopsworks and an Associate Professor at KTH Royal Institute of Technology. He is a co-inventor of the open-source Hopsworks Feature Store platform and develops a free online course on Serverless Machine Learning.
Talk Track: Workshop
Talk Technical Level: 5/7
Talk Abstract:
Only 4 years ago, if you wanted to build a production ML system, you would need tens of engineers to build and maintain a ML platform, such as Michelangelo at Uber. Now, with serverless ML platforms, you can build an operate a ML system in minimal time.
In this tutorial, we will build a serverless ML system from three different Python programs that, when plugged together make up a production ML system. The programs are:
For this tutorial, you will need experience with programming in Python, a laptop and Internet access.
The example shown will be drawn from our course on serverless ML.
Presenters:
Amber Roberts, Machine Learning Engineer, Arize AI | Kyle Gallatin, Senior Software Engeineer I, Machine Learning, Etsy
About the Speakers:
Amber Roberts is a community-oriented Machine Learning Engineer at Arize AI, an ML observability company. Amber’s role at Arize looks to help teams across all industries build ML Observability into their productionalized AI environments. Previously, Amber was a product manager of AI at Splunk and the Head of Artificial Intelligence at Insight Data Science. A Carnegie Fellow, Amber has an MS in Astrophysics from the Universidad de Chile.
Kyle Gallatin is Senior Software Engineer I, Machine Learning, at Etsy; formerly, he was a data scientist and MLE at Pfizer; he has an MA in Molecular and Cell Biology from Quinnipiac University || Amber Roberts is Machine Learning Engineer @ Arize AI. Previously: PM at Splunk, Head of AI at Insight Data Science. Carnegie Fellow (2016–2017); MS Astrophysics.
Talk Track: Workshop
Talk Technical Level: 3/7
Talk Abstract:
While 80% of data generated is unstructured images, text, or audio, ML teams working with this unstructured data often ship models blind. This lack of visibility can create costly and time-intensive problems — and some heartache, too.
Internal embedding representations can be extracted from almost all types of deep learning models, giving an internal glimpse at what the model is “seeing.”
Embeddings are key to workflows that allow teams to identify issues, resolve them, and continually improve models and data. Join us as Amber Roberts, ML Engineer at Arize AI, and Kyle Gallatin, Senior Software Engineer I, Machine Learning at Etsy, as they discuss Etsy’s journey with embeddings, the challenges they’ve encountered, and best practices when troubleshooting unstructured data models.
Presenter:
Dr. Sean Wise, Professor, Toronto Metropolitan University
About the Speaker:
Dr. Sean Wise BA LLB MBA PhD is an expert on startups & venture capital. He uses this expertise in his various roles as: tenured Toronto Metropolitan University, formerly Ryerson University, Professor, bestselling author, international business speaker, and partner at Ryerson Futures, a seed stage venture capital fund and technology accelerator.
Dr. Wise has been an entrepreneur since age 13. His latest venture is the edtech startup ProfBot.ai, which is in stealth mode. He’s already sold one of his startups for a staggering $30 million dollars, demonstrating his keen eye for successful business opportunities.
He has been in the Venture Capital industry for more than 15 years and has mentored hundreds of innovation focused startups; collectively, these companies have raised more than $2.1 Billion in capital. Over the last seven years, Dr. Wise has invested in over 26 startups, making him a seasoned veteran in the world of entrepreneurship.
Dr. Wise spent five seasons as an advisor for CBC on the mega hit venture reality show Dragons’ Den (aka Shark Tank, in America) before moving in front of the camera as the host of the Naked Entrepreneur, which airs on the Oprah Winfrey Network. Dr. Wise writes a national column for the Huffington Post (titled: “Dear Professor Investor”) in which he answers some of the toughest startup questions posted.
Dr. Wise studied economics and engineering at Carleton University, earned his MBA and Law degrees from the University of Ottawa, and received his PhD in Business from the Adam Smith Business School at the University of Glasgow. He has published five books and more than two dozen peer-reviewed research papers and case studies of high-growth startups. His research on social network analysis for collective intelligence and innovation has been featured at MIT’s COINs conference and the Academy of Management. Dr. Wise is also a talented moderator having spent five years as the host of the Naked Entrepreneur, which aired on the Oprah Winfrey Network.
As faculty at Toronto Metropolitan University, Dr. Wise has been exploring the ethical and social impacts of AI use cases for campus. As an investor with Ryerson Futures, he has supported dozens of startups, including one of Canada’s leading AI companies: Ada.
Talk Track: Workshop
Talk Technical Level: 3/7
Talk Abstract:
Join us for a cutting-edge workshop at the AI conference that will explore how to deploy OpenAI chatbots as personalized digital tutors for undergraduate students in higher education. This workshop will delve into the use of natural language processing and machine leDeploying an OpenAI Chatbot as a Personalized Digital Tutor for Undergraduate Students in Higher Education, A workshoparning algorithms to create chatbots that adapt to the pace, style, and knowledge level of individual students.
Participants will learn how to design and implement chatbots that provide a 24/7 accessible and tailored educational experience, improving student engagement and satisfaction. Don’t miss this opportunity to discover how AI can revolutionize education and enhance learning outcomes for the next generation of students.
Please bring a short exam you have given recently, along with the correct answers and the grading rubric for that test to the workshop and YOU WILL LEAVE WITH YOUR OWN WORKING CHATBOT.
What you’ll Learn:
At this workshop, everyone will be building hands-on. They’re very own digital tutor to take back to their office.
Presenter:
James Cameron, Senior AI/ML Architect, NVIDIA
About the Speaker:
James is a Senior Solutions Architect from Nvidia where he works with companies to design, develop, and deploy their AI systems on the edge or in the data center. Previously he was a Team Lead at Patriot One Technologies where he designed and deployed many production AI/ML systems.
Talk Track: Workshop
Talk Technical Level: 4/7
Talk Abstract:
With the rise of LLMs and other generative deep learning models, people are starting to build use cases at scale that require large amounts of compute. In this workshop we will highlight useful approaches that will reduce inference costs, and decrease latency.
What you’ll Learn:
Practical examples and code samples. Succient, actionable items
Presenter:
Jordan Shaw, Creative Technical Leadership, Half Helix
About the Speaker:
Jordan Shaw is an artist and creative technologist. As a MFA graduate from OCAD University’s Digital Futures program, his work spans mediums ranging from strictly digital to interactive physical environments. He explores themes relating to the influence of technology in popular culture and the predefined expectations society has about their relationship with computers, technology, and the future. The manifestation of his creative output transpires through projects that focus on communicating the invisible, yet very physical components of technology and how these unnoticed pieces of technology impact, influence and alter our behaviour and surroundings.
Talk Track: Workshop
Talk Technical Level: 4/7
Talk Abstract:
This workshop will look at a variety of ways AI, ML and data science has been utilized in creative ways to contemporary artworks and creative practices. We’ll look at the use of Tensorflow.js, ml5js, PoseNet, BlazePose, and Node.js + p5.js and maybe some others as ways to add smart and engaging interactivity to creative projects such as interactive artworks.
What you’ll Learn:
Hands-on experience and opportunity to explore basic data training, using it to control either visual or hardware in a creative and collaborative way.
Presenters:
Royal Sequeira, Machine Learning Engineer, Georgian | Azin Asgarian, AI Technical Lead, Georgian
About the Speakers:
Royal Sequiera is a Machine Learning Engineer at Georgian. He did his Masters in Computer Science from University of Waterloo. In the past, he has worked at Microsoft Research India, LG Toronto AI Lab, and Ada Support Inc. in Toronto. In 2018, he founded Sushiksha, a mentorship organization that helps hundreds of students across India. In his free time, he reads books, meditates, and likes to learn new languages.
Azin Asgarian is currently an AI Technical Lead on Georgian’s R&D team where she works with companies to help adopt applied research techniques to overcome business challenges. Prior to joining Georgian, Azin was a research assistant at the University of Toronto and part of the Computer Vision Group where she was working on the intersection of Machine Learning, Transfer Learning, and Computer Vision
Talk Track: Workshop
Talk Technical Level: 5/7
Talk Abstract:
With the advent of Large Language Models (LLMs), the only way to access them have been via prompt engineering. In this workshop, we will introduce the concept of prompt engineering and how to interface with LLMs and external applications. Using various tools, we will demonstrate prompt engineering—from basic to advanced prompting. We will cover various topics such as zero-shot prompting, chain of thoughts prompting etc.
What You’ll Learn:
This will be a workshop on making the best of LLMs via prompt engineering LLMs via prompt engineering
Presenter:
Nick Acosta, Developer Advocate, Tecton
About the Speaker:
Nick Acosta enjoys helping developers automate feature pipelines as a Developer Advocate at Tecton, a feature platform for real-time machine learning. Nick has previously led Developer Relations teams at Fivetran and IBM.
Talk Track: Workshop
Talk Technical Level: 5/7
Talk Abstract:
Are you struggling to keep up with the demands of real-time machine learning? Like most organizations building real-time ML, you’re probably looking for a better way to:
Manage the lifecycle of ML models and features
Implement batch, streaming, and real-time data pipelines
Generate accurate training datasets
Serve models and data online with strict SLAs, supporting millisecond latencies and high query volumes
Look no further! In this workshop, we’ll walk through the concepts and code that will help you build a modern technical architecture that simplifies the process of managing real-time ML models and features. Nick will explain how feature platforms serve as the interface between raw data and real-time models.
Through a series of code examples and use cases, this workshop will demonstrate how feature platforms can be used to:
– Automate batch, streaming, and real-time feature pipelines for ML
– Orchestrate the transformation of data
– Store features in offline and online stores
– Generate accurate training datasets
– Serve features online for real-time inference
– Monitor data pipelines and online service levels
– Manage the flywheel of ML data
What You’ll Learn:
What are feature platforms, and how can they be used to make real-time machine learning easier?
Presenter:
Anouk Dutrée, Product Owner, UbiOps
About the Speaker:
AAnouk is a Product Owner at UbiOps. She studied Nanobiology and Computer Science at the Delft University of Technology, which spiked her interest in Machine Learning. Next to her role at UbiOps, she also frequently writes for Towards Data Science about various MLOps topics, she co-hosts the biggest Dutch data podcast, de Dataloog, and she recently rounded up a Master’s in Game Development. Her efforts in tech have been awarded twice with the Tech 500 award (T500), in both 2020 and 2021.
Track: Virtual Workshop
Technical Level: 4/7
Abstract:
Generative AI models are all the hype nowadays, but how do you actually deploy them in a scalable way? In this talk we will discuss best practices when moving models to production, as well as show an interactive example of how to deploy one using UbiOps. UbiOps is a serverless and cloud agnostic platform for AI & ML models, built to help data science teams run and scale models in production. We will pay special attention to typical hurdles encountered in deploying (generative) AI models at scale. Python knowledge is all you need for following along!
What You’ll Learn:
Deployment at scale doesn’t have to be difficult. Participants will learn how to deploy a generative AI model to the cloud themselves, and how to make sure it runs with the right resources (CPU,GPU,IPU etc.).
Prerequisite Knowledge:
Python knowledge and a basic understanding of what computer vision is.
Presenter:
Niels Bantilan, Chief Machine Learning Engineer, Union.ai
About the Speaker:
Niels is the Chief Machine Learning Engineer at Union.ai, and core maintainer of Flyte, an open source workflow orchestration tool, author of UnionML, an MLOps framework for machine learning microservices, and creator of Pandera, a statistical typing and data testing tool for scientific data containers. His mission is to help data science and machine learning practitioners be more productive.
He has a Masters in Public Health with a specialization in sociomedical science and public health informatics, and prior to that a background in developmental biology and immunology. His research interests include reinforcement learning, AutoML, creative machine learning, and fairness, accountability, and transparency in automated systems.
Track: Virtual Workshop
Technical Level: 4/7
Abstract:
The use of Language Models (LMs) has become more widespread in recent years, thanks in part to the broader accessibility of datasets and the ML frameworks needed to facilitate the training of these models. Many of these models are large – hence the terminology of Large Language Models (LLMs) – and serve as so-called foundation models, which are trained by organizations with the compute resources to train them. These foundation models, in turn, can be fine-tuned by the broader machine learning community for specific use cases, perhaps on proprietary data. One of the barriers that make fine-tuning these models is infrastructure: even with cloud tools like Google Colab and the wider availability of consumer-grade GPUs, putting together a runtime environment to fine-tune these models is still a major challenge. This workshop will give attendees hands-on experience on how to use Flyte to declaratively specify infrastructure so that they can configure training jobs to run on the required compute resources to fine-tune LMs on their own data.
What You’ll Learn:
This workshop has two main learning goals. First, attendees will learn the main concepts behind Flyte, a workflow orchestrator for data and machine learning. Many of these concepts are orchestrator-agnostic, such as containerization for reproducibility, declarative infrastructure, and type-safety. Secondly, they will also learn how to leverage the latest deep learning frameworks that optimize memory and compute resources required to fine-tune language models in the most economical way.
Prerequisite Knowledge:
Intermediate Python, working knowledge of Docker, and intermediate knowledge of machine learning.
Presenter:
Amrit Krishnan, Senior Applied Machine Learning Specialist, Vector Institute
About the Speaker:
Amrit is a Senior Applied ML Specialist at the Vector Institute, currently working on building tools for enabling ML deployment for healthcare. Amrit is passionate about Healthcare ML, Robotics, Open-Source and Software Engineering for productionizing ML systems.
Track: Virtual Workshop
Technical Level: 5/7
Abstract:
The use of ML in healthcare applications is rising steadily. Deployment of these systems requires a responsible approach, and regulation is lagging behind. At the Vector Institute, in strong collaboration with our stakeholders, we are building an open-source software framework to address this gap. Specifically, we are focussing on rigorous evaluation and monitoring of the ML system across patient sub-populations. We will show how we can generate evaluation and monitoring reports for end-users, using two use cases.
Additionally, we will also discuss challenges in implementing monitoring sub-systems in healthcare settings.
What You’ll Learn:
You will learn how to use an open-source toolkit to develop ML models on health data, with a focus on deployment and workflow integration.
Prerequisite Knowledge:
Clinical Data, ML, Software Development, Open-source
Presenters:
Rahm Hafiz, CTO & Co-Founder, Armilla AI | Dan Adamson, CEO & Co-Founder, Armilla AI
About the Speakers:
Rahm Hafiz is the co-founder and CTO of Armilla AI, a company helping institutions testing and building more fair, safe, trustworthy, and useful AI. In the past, Rahm headed AI initiatives at Outside IQ and Exiger where he worked with global financial institutions, government and regulatory bodies to bring innovative AI into reality. Rahm’s PhD work offers an efficient and modular framework for syntactic and semantic analyses and understanding of natural language by addressing some long standing NLP problems including correct processing of ambiguity. Rahm’s research on NLP has been published in over 10 reputable journals and conferences.
Dan Adamson is the Co-Founder and CEO of Armilla.AI, a company helping institutions testing and building more fair, safe, trustworthy, and useful AI. Previously, he founded and served as OutsideIQ’s CEO, since its inception in 2010 until its acquisition in 2017 by Exiger, where he remained as their President overseeing product and cognitive computing research. Dan also previously served as Chief Architect at Medstory, a vertical search start-up acquired by Microsoft. He is an expert on vertical search, investigative use of big data and cognitive computing with more than 15 years in the industry. He holds several search algorithm and cognitive computing patents, has been named among the most influential “must-see” thought leaders in AI and FinTech, in addition to being a recipient of numerous academic awards and holding a Master of Science degree from U.C. Berkeley.
Track: Virtual Workshop
Technical Level: 4/7
Abstract:
We discuss a new alignment templating technology that can be used to enhance the fairness, robustness and safety of the generative AI by understanding the expected behaviour of the model, measuring where the model is underperforming with synthetic test data, and iteratively improving the model with minimal humans in the loop fine-tuning approach.
This alignment platform can be used for use case specific tasks including coercing a generative model to be more fair towards under represented groups, less toxic and less misogynistic, more leaning towards a desired political viewpoints, to use specific tone while generating answers for customer service applications, to preserve PII, to guard models from generating potentially harmful responses etc.
Our alignment technology can interact with users as needed to adjudicate critical decision points to guide its intention-understanding, data generation, testing and tuning capabilities to be more contextual.
What You’ll Learn:
By the end of the workshop, participants will understand common shortcomings of Gen AI, how to use a self-governing alignment technology to overcome those shortcomings and to make GenAI useful for user specific tasks.
Prerequisite Knowledge:
Concepts of large language models, text 2 image models.
Presenters:
Raheleh Givehchi, Lead Data Scientist, Pelmorex Corp. | Hicham Benzamane, Team Lead, Data Engineering Insights, Pelmorex Corp.
About the Speakers:
Raheleh Givehchi, I am a lead data scientist with several years of experience in data science and machine learning. I specialize in delivering valuable insights through data analytics and advanced data-driven methods. Currently, I am working on the Weather Insight Platform (WIP) to provide data-driven solutions to weather-related challenges.
Hicham Benzamane – Team Lead, Data Engineering Insights
– Multiple years working experience in software development and Data engineering.
– Now leading a team of software and data engineers to leverage data and ML to empower Pelmorex customers with cutting edge insights.
Track: Virtual Workshop
Technical Level: 5/7
Abstract:
How to train millions of models while keeping cost very low and predict on daily basis with the highest performance.
What You’ll Learn:
– Real world Data optimization techniques for training and prediction
– Scale techniques using Cloud infrastructure
Prerequisite Knowledge:
– Basics on Machine Learning
– Basics on Cloud Computing
Presenter:
Chinmay Sheth, Senior Machine Learning Engineer, Royal Bank of Canada (RBC) | Colleen Gilhuly, PhD, Data Scientist, Royal Bank of Canada (RBC) | Haleh Shahzad, Director, Data Science, Royal Bank of Canada (RBC)
About the Speaker:
Chinmay Sheth is a Senior Machine Learning Engineer at the Royal Bank of Canada and is completing his MSc in Computer Science at McMaster University. His responsibilities include providing support for MLOps tooling, promoting ML models to production, and developing production-grade data pipelines.
Colleen completed her PhD in Astronomy & Astrophysics at the University of Toronto. She then turned to data science in order to combine her favourite aspects of research with a non-academic career. She joined the Chief Data Office at RBC in July 2022, working primarily on projects in NLP.
Haleh is a Director, Data Science at RBC with 9+ years of experience in AI (Deep learning, machine learning), software development and advanced analytics.
She is currently leading AI initiatives for cutting-edge solutions to maximize the impact of data across organization at RBC. She is also an instructor in the school of continuing studies at York University.
Haleh has a Ph.D. degree in Electrical and Computer Engineering from McMaster university where she was also part of the sessional faculty in the Electrical and computer engineering department.
Track: Virtual Workshop
Technical Level: 5/7
Abstract:
Large language models (LLMs) such as ChatGPT are able to interpret text input and generate human-like responses. Many individuals and companies are excited to use this technology, but integration remains a question mark. Applications using LLMs are also limited by their tendency to invent information and give unpredictable answers.
Microsoft’s Semantic Kernel is an open source, lightweight SDK which enables fast integration of LLMs into a wide range of applications. It also enhances the power of LLM with a structured approach to responses and the ability to refer to external sources of truth.
In this workshop, we will give a crash-course on Microsoft Semantic Kernel and demonstrate how to create a simple web app that harnesses the power of LLMs from OpenAI. This workshop will be hands on so please come prepared with an OpenAI API key.
What You’ll Learn:
– Learn the basics of LLMs, and their applications/impacts in industry
– Hands-on experience with Semantic Kernel’s skills, planner, memories, and chains
– You will be able to build your own web app with an LLM integrated into the backend
Presenters:
Kevin Laven, Energy & Resource Leader, Deloitte | Alex Gobolos, Solutions Engineer, Dataiku
About the Speakers:
Kevin Laven leads the Energy, Resources, and Industrial sector team within Deloitte Canada’s Artificial Intelligence practice. Kevin’s AI experience started in 2003 with a Masters degree at the University of Toronto Machine Learning Lab, and includes dozens of AI models for ER&I applications.
Alex is a Solutions Engineer at Dataiku. He works with customers to get value from all things data and analytics, from data access and exploration, to machine learning and AI. Alex has worked in the analytics space for the last 15+ years including project delivery, consulting, and solutions engineering across industries such as banking and insurance, healthcare, and manufacturing.
Track: Virtual Workshop
Technical Level: 4/7
Abstract:
Opportunity for business users and executives to be exposed to the following topics: How to validate use cases, Approaches to building models, Business case for deployment
What You’ll Learn:
High-level steps on creating an ML model as a business user
Prerequisite Knowledge:
Basic knowledge of ML
Presenter:
Alon Gubkin, CTO & Co-Founder, Aporia
About the Speaker:
In 2019, Alon Gubkin cofounded Aporia, the ML observability platform. Aporia is trusted by Fortune 500 companies and data science teams in every industry to ensure responsible AI and monitor, improve, and scale ML models in production. Alon, an ex-R&D team lead in the elite Unit 81 intelligence unit of the Israel Defense Forces, has led Aporia in raising $30 million from investors like Tiger Global Management and Samsung Next. For two years in a row, 2022 and 2023, Alon was named to Forbes 30 Under 30.
Track: Virtual Workshop
Technical Level: 4/7
Abstract:
In this workshop, we will explore how to improve your recommender system in production by monitoring your model and generating insights from production data. We will discuss how to track the behavior of different versions of your model, understand the performance of your model in different data slices, and detect data drift. By leveraging monitoring and insights from production data, you will be able to improve the performance and accuracy of your recommender system and drive better business outcomes. Join us to learn how to take your recommender system to the next level with real-time monitoring and actionable insights.
What You’ll Learn:
Everything you need to know to get the most out of your recommender system in production and turn them into a revenue machine.
1. How to track the behavior of different versions of your model.
2. Understand the performance of your model in different data slices.
3. How to detect data drift.
Prerequisite Knowledge:
Basic understanding of ML use cases and ML models in production.
Dan Adamson, CEO & Co-Founder, Armilla AI
About the Speaker:
Dan Adamson is the Co-Founder and CEO of Armilla.AI, a company helping institutions testing and building more fair, safe, trustworthy, and useful AI. Previously, he founded and served as OutsideIQ’s CEO, since its inception in 2010 until its acquisition in 2017 by Exiger, where he remained as their President overseeing product and cognitive computing research.
Dan also previously served as Chief Architect at Medstory, a vertical search start-up acquired by Microsoft. He is an expert on vertical search, investigative use of big data and cognitive computing with more than 15 years in the industry. He holds several search algorithm and cognitive computing patents, has been named among the most influential “must-see” thought leaders in AI and FinTech, in addition to being a recipient of numerous academic awards and holding a Master of Science degree from U.C. Berkeley.
Talk Abstract:
We discuss the risks of using generative AI in enterprise settings, as well as new alignment techniques to test and iteratively improve generative AI models’ safety, fairness, robustness, and fit-for-purpose capability.
Eric Duffy, Senior Director Business Development, Tenstorrent
About the Speaker:
Eric Duffy, from Tenstorrent, is a Computer Scientist by training and a Business Development professional by trade. Eric has previously worked in consulting for Data Management, FinTech and Cloud systems, and has worked with companies across public and private sectors to deliver high-performance analytical and Machine Learning solutions.
Today, Eric is focused on the intersection between AI/ML algorithms and next-generation computers.
Talk Abstract:
The RISC-V ISA has been described as the ‘Linux of the chip world’. It has been used in many small scale applications, but in recent times it has seen a rise in popularity for high performance system designs for AI and HPC applications… all of this, in spite of a relatively nascent software ecosystem. Learn why companies like Tenstorrent are turning to RISC-V over traditional x86 and ARM architectures.
Mark Gibbas, CEO, Weather Source
About the Speaker:
Mark has grown Weather Source from a small start-up to a leading provider of weather and climate technology. In its early years, Weather Source primarily provided weather information and consulting. During this time, Mark developed a vision for a scalable weather operating system that could serve clients across any industry and provide each with information relevant to their specific needs and locations of interest.
Today, Weather Source’s OnPoint Platform and related products provide actionable weather insights to a growing list of top-ranked companies.
Prior to founding Weather Source, Mark applied his expertise in meteorology and computer science at Applied Insurance Research (AIR) and TASC. Among his notable accomplishments is the development of long-range forecasting systems that laid the foundation for the use of such technology at WSI and The Weather Company.
Mark has also enjoyed working on projects for the United Nations’ World Meteorological Organization, where he advanced the meteorological capabilities of several Latin American countries. Mark holds a bachelor’s degree in meteorology with minors in computer science and mathematics.
Talk Abstract:
Weather impacts a wide range of business functions, this talk explores examples of these impacts and how to solve them.
Niv Hertz, Solutions Architect, Aporia
About the Speaker:
Niv Hertz is a Solutions Architect at Aporia, where he customizes ML monitoring to support any use case or need. Before joining Aporia, Niv worked as a Software Engineer and Cyber Security expert at other startups and in the elite 81 intelligence unit of the Israel Defense Forces. When he is not in front of a computer, Niv also enjoys hiking and basketball.
Talk Abstract:
Dive into the world of monitoring Large Language Models (LLMs). We’ll cover key points like tracking drift and measuring model performance, and highlight useful tools and alert systems. This talk aims to give data scientists and ML engineers a basic but solid starting point on their monitoring journey to ensure high performance of their LLMs, contributing to the development of next-gen models.
Yves Fogel, CTO & Co-Founder, Flipando
About the Speaker:
Yves is a visionary leader at the intersection of art and technology. As the CTO and co-founder of Flipando.ai, he is revolutionizing several industries by applying his GenAI expertise to create solutions that empower everyone to bring their ideas to life.
Yves recently completed his Master’s at NYU, where he developed a learning by doing approach to democratize GenAI tools and knowledge. He also launched an open-source project called VAPAI, an Unreal Engine plugin that makes Virtual Production accessible to everyone. With his technical skills and experience as a Growth Manager for an AI dev shop, Yves is leading his team to build the next generation of GenAI solutions. Join us to witness the power of Yves and Flipando.ai in action.
Talk Abstract:
In this 3-minute lightning talk, we will demonstrate the democratization of GenAI tools and knowledge by building a GenAI app from scratch. We will show how a learning by doing approach can empower everyone to bring their ideas to life. With GenAI, the possibilities are endless, and each person in each diverse industry or department can build their own apps: marketing, legal, finance, or teams developing the next GenAI app. Join us to witness the power of GenAI and see how we can build the next generation of GenAI solutions together.
Akbar Nurlybayev, Co-Founder, CentML
About the Speaker:
Co-founder of CentML, where we make machine learning affordable. Prior to CentML, I was a director of Data Science and Data Platform at KAR Global, where I was responsible for modernizing KAR’s data assets and establishing data science and machine learning practices.
Talk Abstract:
The speed of innovation in Generative AI space is breathtaking. Models are getting large, requiring more powerful and increasingly scarce GPUs. As a result the cloud computing costs are increasing too. Converting your models to use FasterTransformer, TensorRT, DeepSpeed and etc. often requires significant engineering effort. If there were only an automatic way to performance tune your models! Enter Hidet – state of the art modern deep learning compiler.
Business Leaders: C-Level Executives, Project Managers, and Product Owners will get to explore best practices, methodologies, principles, and practices for achieving ROI.
Engineers, Researchers, Data Practitioners: Will get a better understanding of the challenges, solutions, and ideas being offered via breakouts & workshops on Natural Language Processing, Neural Nets, Reinforcement Learning, Generative Adversarial Networks (GANs), Evolution Strategies, AutoML, and more.
Job Seekers: Will have the opportunity to network virtually and meet over 60 Top Al Start-ups and companies during the EXPO & Career Fair.
Ignite what is an Ignite Talk?
Ignite is an innovative and fast-paced style used to deliver a concise presentation.
During an Ignite Talk, presenters discuss their research using 20 image-centric slides which automatically advance every 15 seconds.
The result is a fun and engaging five-minute presentation.
You can see all our speakers and full agenda here