Aida Nematzadeh
Recording : TBD
Recent generative models have demonstrated remarkable capabilities, from solving intricate reasoning problems to creating highly realistic images. However, as these models grow more complex, evaluating them presents increasing challenges—particularly since we often have access only to their outputs, not the underlying mechanisms. This predicament mirrors a challenge faced by cognitive scientists: understanding human cognition by observing behavior
without direct access to the “cognitive model” itself. In this talk, I will explore how principles from cognitive science can illuminate the evaluation of generative models. I will discuss how cognitive science approaches, such as experimental design in human data collection, probing for specific capabilities, and developing automated evaluation metrics, can offer valuable insights into understanding and assessing these advanced models.
David Rolnick
Recording : TBD
Machine learning is increasingly being called upon to help address climate change, from processing satellite imagery to modeling Earth systems. Such settings represent an important frontier for machine learning innovation, where traditional paradigms of large, general-purpose datasets and models often fall short. In this talk, we show how an application-driven paradigm for algorithm design can respond to problem-specific goals and incorporate relevant domain knowledge. We introduce novel techniques that leverage the structure of the problem (such as physical constraints and multi-modal self-supervision) to improve accuracy and usability across applications, including monitoring land use with remote sensing, designing chemical catalysts for the energy transition, and downscaling climate data.
Machine learning practitioners often face significant challenges in formally integrating their prior knowledge and beliefs into predictive models, limiting the potential for nuanced and context-aware analyses. Moreover, the expertise needed to integrate this prior knowledge into probabilistic modeling typically limits the application of these models to specialists. In this talk, we present a regression model that can process numerical data and make probabilistic predictions at arbitrary locations, guided by natural language text which describes a user's prior knowledge. Large Language Models (LLMs) provide a useful starting point for designing such a tool since they 1) provide an interface where users can incorporate expert insights in natural language and 2) provide an opportunity for leveraging latent problem-relevant knowledge encoded in LLMs that users may not have themselves. We start by exploring strategies for eliciting explicit, coherent numerical predictive distributions from LLMs. We examine these joint predictive distributions, which we call LLM Processes and we demonstrate the ability to usefully incorporate text into numerical predictions, improving predictive performance and giving quantitative structure that reflects qualitative descriptions. This lets us begin to explore the rich, grounded hypothesis space that LLMs implicitly encode.
Identifying latent representations or causal structures is important for good generalization and downstream task performance. However, both fields have been developed rather independently. We observe that several methods in both representation and causal structure learning rely on the same data-generating process (DGP), namely, exchangeable but not i.i.d. (independent and identically distributed) data. We provide a unified framework, termed Identifiable Exchangeable Mechanisms (IEM), for representation and structure learning under the lens of exchangeability. IEM provides new insights that let us relax the necessary conditions for causal structure identification in exchangeable non--i.i.d. data. We also demonstrate the existence of a duality condition in identifiable representation learning, leading to new identifiability results. We hope this work will pave the way for further research in causal representation learning.
In this talk I will cover our work inferring cellular dynamics during differentiation and disease with neural ODE networks that are regularized to follow the data geometry or manifold. First I will cover Mioflow for dynamic optimal transport-based derivation of single cell trajectories from static snapshot scRNA-seq data. Then I will discuss RITINI, our recent graph ODE network which allows us to learn gene regulation that underlies cellular dynamics. I will showcase applications of these in triple negative breast cancer and human embryonic stem cell differentiation. Finally, I will cover learnable geometric scattering (LEGS) networks, a multiscale wavelet-based neural network that allows us to classify and characterize populations of cells, and their trajectories. I will also show some applications of these techniques to neural activity dynamics.
A major debate has recently emerged concerning whether large-scale AI systems are capable of human-like reasoning, with some arguing that their apparent reasoning capabilities are based instead on mimicry of their vast training data, and others arguing that these systems constitute an early form of AGI. To shed light on these issues, I will discuss work that aims to systematically evaluate, understand, and improve the reasoning capabilities of these systems. First, drawing on insights from cognitive science, I will argue that it is important not to treat reasoning as a monolithic capacity, but instead to carefully evaluate distinct modes of reasoning in isolation. I will present results that highlight interesting and surprising dissociations between these distinct modes, including findings suggesting that LLMs possess strong analogical reasoning capabilities, while displaying major weaknesses in physical and mathematical reasoning. Second, I will discuss the mechanisms that enable these reasoning capacities, drawing on both cognitive theories and work on architectural inductive biases in neural networks. Finally, I will highlight potential paths toward addressing the major remaining challenges for machine reasoning, including especially the need for object-centric visual processing, and improved methods for goal-directed planning..