October - December 2022


AI Helps Ukraine consists of a full-day conference at Mila on December 8th 2022 and a series of online talks between October and December 2022.

December 8th, Thursday (in person at Mila, Montreal)

The times are given in the local Montreal time (Eastern Standard Time, UTC-5).

The conference will take place at the Agora of Mila, which is located at 6650, Rue Saint-Urbain, Montreal, QC, H2S 3H1.

Time (EDT) Session Speaker Title
09:00 - 09:30 Registration
09:30 - 09:45 Opening remarks
09:45 - 10:45 Keynote talk I Olga Russakovsky
Trustworthy (and trusted) computer vision Computer vision is at an interesting cruxpoint: the systems are now both accurate enough that they're increasingly being deployed in high-stakes settings, but also have enough errors that folks are losing faith in them (with good reason). This talk will reflect on some of the recent work from the Princeton Visual AI lab on building more trustworthy (and more trusted) computer vision. I'll talk about improving the benchmarks, mitigating the social bias of the systems, and creating better model explanations that are more understandable for end users.
10:45 - 11:00 Coffee Break
11:00 - 12:00 Panel discussion Yoshua Bengio, Anna Goldenberg, Olga Russakovsky, Ava Amini, Oleksandr Romanko AI for social good
12:00 - 13:30 Lunch Break
13:30 - 14:30 Keynote talk II Anna Goldenberg
Time series ML for deployment in healthcare There is growing support and excitement around AI integration into clinical practice. And yet, enabling AI in healthcare broadly is replete with many obstacles including fundamentally unresolved machine learning issues. In this talk I will cover several of our contributions to time series modeling including explainability, representation learning and generative modeling that we have worked on to date. I will also discuss some of the broader computer science contributions we had to make in order to get us closer to deployment. Finally, if time permits, I will touch on an important question of feedback loop, i.e. what happens when an ML solution is deployed and is starting to affect the labels that we are using to retrain and improve our models.
14:30 - 14:45 Invited talks Mariia Rizhko, Bohdan Naida Lightening talks by Ukrainian students
14:45 - 15:00 Coffee Break
15:00 - 16:00 Keynote talk III Ava Amini
AI to optimize biology The potential of artificial intelligence (AI) in biology is immense, yet its success is contingent on interfacing effectively with wet-lab experimentation. In this talk, I will focus on two settings, biomolecular and experimental design, in which we’ve developed new AI algorithms to optimize biology and inform the experimental process. I will first share recent work in developing a diffusion-based generative model that designs protein structures by mirroring the native protein folding process. Moving from molecules to experiments, I will present a method — evidential deep learning — for uncertainty quantification in neural networks and demonstrate its potential to guide key steps in experimental lifecycles, opening the door for sustained feedback between computation and experimentation in the biological sciences.
16:00 - 16:15 Invited talk Oleksandr Romanko AI for Education
16:15 - 17:00 Invited talks Talks by founders of AI for Good startups
17:00 - 17:15 Closing remarks
17:15 - 19:00 Networking

Virtual series

The times are given in Universal Coordinated Time (UTC). Click on the times to check your local time.

Date Time (UTC) Speaker(s) Title Stream
October 26th (Wednesday) 16:30 Timnit Gebru
Community rooted, independent AI research The Distributed Artificial Intelligence Research Institute (DAIR) was launched in December 2021 by Timnit Gebru as a space for independent, community-rooted AI research, free from Big Tech’s pervasive influence. Gebru believes that the harms embedded in AI technology are preventable and that when its production and deployment include diverse perspectives and deliberate processes, it can be put to work for people, rather than against them. With DAIR, Gebru aims to create an environment that is independent from the structures and systems that incentivize profit over ethics and individual well-being. In this talk, Gebru will discuss why she founded DAIR and what she hopes this interdisciplinary, community-based, global network of AI researchers can accomplish. She will discuss the incentive structures that make it difficult to perform ethical AI research, and give examples of how DAIR is hoping to forge a different path.
Crowdcast | YouTube
November 3rd (Thursday) 16:00 Yoshua Bengio
AI to explore molecular space and fight pandemics and climate change Machine learning research is expanding its reach, beyond the traditional realm of the tech industry and into the activities of other scientists, opening the door to truly transformative advances in these disciplines. In this lecture I will focus on two aspects, modeling and experimental design, that are intertwined in the theory-experiment-analysis active learning loop that constitutes a core element of the scientific methodology. Computers will be necessary to go beyond the currently purely manual research loop and take advantage of high-throughput experimental setups and large-scale experimental datasets. I will discuss methods related to active learning, reinforcement learning, generative modeling, Bayesian ML, amortized variational learning and causal discovery. I will discuss the notion of epistemic uncertainty and how to estimate it. I will motivate generative policies that can sample a diverse set of candidate solutions to a problem, be it for proposing new experiments or causal hypotheses. Finally, I will describe current research to help us with these questions based on a new deep learning probabilistic framework called GFlowNets and how we plan to apply these in areas of great societal need like the unmet challenge of antimicrobial resistance or the discovery of new materials to help fight climate change.
Crowdcast | YouTube
November 7th (Monday) 16:00 Irina Rish
Computational psychology and psychological computation: How AI and Brain Sciences can help each other The rapidly developing field of computational psychology/psychiatry aims to bridge the gap between the “legacy” approaches still prevalent in mental health care that rely on primarily subjective self-reports and subjective evaluations by doctors - in a stark contrast to other scientific and medical fields where objective tests (e.g. blood work, X-ray, MRI) are typically used to provide an objective basis for diagnosis and prognosis. The promise of computational psychiatry is to bring psychological diagnosis to a more solid ground based on objective measurements, potentially including brain imaging, wearable sensor data, as well as statistical analysis of speech and text data - the are where modern AI techniques can prove highly useful, I will briefly mention some past and ongoing work by our group and collaborators in this directions, from developing more robust statistical models across brain imaging datasets, aiming at learning invariant features of a mental state or a person across varying setting, to transferring knowledge extracted across different data modalities, following recent advances in large-scale pretrained deep neural network models whose transfer capabilities appear to go far beyond what was previously considered possible. However, the road between AI and brain sciences is bidirectional: among many ways in which neuroscience, psychiatry and psychology can inspire novel AI approaches, the ability of an AI system to maintain “meaningful” dialog with a person, and reaching certain objectives of the dialog (e.g., application of AI in semi-automated therapy), can be considerably improved by importing certain practices and ideas from the therapeutic dialog into AI language models (I will briefly summarize our prior work on depression therapy dialogues). Finally, given the rapid advances in the capabilities of large scale models in the past 1-2 years, sometimes called a “scaling revolution” these days, also increase the importance of developing AI systems that are aligned with human values and can avoid harming humans while remaining useful to them - the blossoming field of AI Alignment (as a subfield of AI Safety, which, beyond alignment problem, includes robustness and interpretability of AI systems). The vast amount of practical knowledge from the fields of psychology and neuroscience, and from psychiatric therapy practice, can be clearly very advancing the AI Alignment field.
Crowdcast | YouTube
November 11th (Friday) 16:00 Hannah Kerner, Inbal Becker-Reshef
AI and Earth observations for global agricultural monitoring and food security Earth-observing satellites are collecting terabytes of observations of the entire planet every day with unprecedented clarity. These globally-available datasets offer immense opportunities for providing time-sensitive information needed by decision-makers about issues important to society, such as agricultural production and food security. Analyzing these petabyte-scale Earth observation datasets requires AI and machine learning methods designed to meet the unique challenges and opportunities of Earth observation data. Harvest is NASA’s applied sciences program on agriculture and food security, committed to advancing the use of satellite Earth observations to benefit food security and agriculture in the US and worldwide. In this talk, Harvest Director Inbal Becker-Reshef and Harvest ML/AI Lead Hannah Kerner will talk about how Harvest is using AI and Earth observations to rapidly respond to critical events impacting global agriculture and food security, including the Russian invasion of Ukraine.
Crowdcast | YouTube
November 14th (Monday) 16:00 Max Welling
Generating and steering molecules with ML and RL After speech, text, image and video, a whole new application area is opening up for deep learning technology: accelerating molecular simulation. In this talk I will highlight two such methods in Transition Path Sampling and Ligand-Protein docking.
Crowdcast | YouTube
November 17th (Thursday) 19:00 Alexei Efros
Visual self-supervision in the Post-Dataset Era Starting in the early 2000s, datasets have revolutionized computer vision, arguably turning it from an art form into a science. But now, 20 years later, is our infatuation with fixed visual datasets and their training/test splits starting to cause harm? In this talk, I will describe some of our efforts to go past the fixed dataset paradigm and toward a more natural continuous online learning regime, where the distinction between training and test data is more blurred. I will present some of our latest work on test-time training, as well as a comically simple approach to test-time visual prompting / visual analogies.
Crowdcast | YouTube
November 21th (Monday) 16:00 Regina Barzilay Expanding the reach of molecular models in the drug discovery space Crowdcast | YouTube
November 24th (Thursday) 15:00 David Rolnick
Machine learning in climate action Machine learning (ML) can be a useful tool in helping society reduce greenhouse gas emissions and adapt to a changing climate. In this talk, we will explore opportunities and challenges in ML for climate action, from optimizing electrical grids to monitoring crop yield, with an emphasis on how to incorporate domain-specific knowledge into machine learning algorithms. We will also consider ways that ML is used in ways that contribute to climate change, and how to better align the use of ML overall with climate goals.
Crowdcast | YouTube
December 5th (Monday) 16:00 Sara Beery
Mapping urban trees across North America with the Auto Arborist Dataset Generalization to novel domains is a fundamental challenge for computer vision. Near-perfect accuracy on benchmarks is common, but these models do not work as expected when deployed outside of the training distribution. To build computer vision systems that solve real-world problems at global scale, we need benchmarks that fully capture real-world complexity, including geographic domain shift, long-tailed distributions, and data noise. We propose urban forest monitoring as an ideal testbed for studying and improving upon these computer vision challenges, while working towards filling a crucial environmental and societal need. Urban forests provide significant benefits to urban societies. However, planning and maintaining these forests is expensive. One particularly costly aspect of urban forest management is monitoring the existing trees in a city: e.g., tracking tree locations, species, and health. We introduce a new large-scale dataset that joins public tree censuses from 23 cities with a large collection of street level and aerial imagery, containing over 2.5M trees and 300 genera. This benchmark enables the exploration of automated urban forest monitoring across modalities and with respect to geographic distribution shifts, vital for such a system to be deployed at-scale.
Crowdcast | YouTube
December 12th (Monday) 16:00 Michael Bronstein Geometric Deep Learning: from Euclid to Drug Design Crowdcast | YouTube
December 15th (Thursday) 17:30 Volodymyr Kuleshov
Uncertainty-Aware Machine Learning for a Sustainable Food Supply Chain In the United States, 30–50% of all the food produced is wasted, representing a massive environmental problem—food waste contributes to up to 25% of all greenhouse gas emissions. One of the main drivers of food waste is supply chain inefficiency. This talk describes (1) novel machine learning methods for reasoning under uncertainty and (2) their application to a real-world supply chain control system that significantly improves efficiency and reduces waste. Specifically, I will present improved methods for calibrated uncertainty estimation and planning algorithms that benefit from accurate model uncertainties. This research extends widely used uncertainty quantification methods and demonstrates that most existing planning algorithms can be improved via these methods. I will also describe how these algorithms are deployed in a real-world supply chain control system. This system is currently deployed across hundreds of supermarkets in the US (handling ~10% of US produce volume) and has led to waste reductions of up to 50%.
Crowdcast | YouTube