Part of the Image-Based Scientific Machine Learning for Theories of Biological Dynamics Across Scales program

​

A long-standing goal of biology is to arrive at a mechanistic understanding of biological dynamics. Modern imaging and recording technologies produce high-dimensional, dynamic measurements of living systems across molecular, cellular, and tissue scales. However, turning these data into mechanistic understanding requires addressing several outstanding challenges. Most fundamental are questions such as: What is actually learnable from a given dataset? How much of the past is relevant for predicting the future? Which low-dimensional variables capture the dynamics that matter? Can one interpret them in terms of physically-measurable quantities? Can one extract insights into the computations performed by biological dynamics? Information theory provides a model-agnostic language for some of these questions. It quantifies statistical dependencies without committing to a specific mechanism, and it sets bounds on what any analysis method could extract. Importantly, the central quantities of information theory measure how well a system predicts, estimates, and controls the variables that matter to it, and so they let us ask not just how a system behaves but what it is for. This ties biology functions to optimal estimation and control theory and offers a functional, even teleological, account of living dynamics that purely descriptive methods cannot. This workshop will explore these topics by bringing together mathematicians, computer scientists, physicists, and biologists to develop and apply information-theoretic methods for learning the dynamics of biological systems directly from experimental data, including the spatiotemporal imaging that anchors this program.

 

The workshop will chart a path from fundamental concepts to working analysis pipelines. Throughout the workshop, we will emphasize interpretability via dynamical systems and control theory and connections to contemporary AI. Indeed, a system that senses and acts can be treated as a controller that must extract relevant information from its inputs and internal dynamics to steer its behavior, connecting control-theoretic and information-theoretic quantities and objectives. Further, the same objectives underlie much of representation learning in contemporary AI, including self-supervised and contrastive methods, and the variational bounds that make these quantities computable also make them usable at scale. The representations learned by deep models can in turn be probed and constrained with information-theoretic tools. These concepts will be explored through technical topics that include predictive information as a measure of predictability of biological dynamics; the information bottleneck and latent-variable models as ways to compress high-dimensional observations into compact, predictively relevant representations; and the practical problem of estimation, namely how to quantify entropy and mutual information from finite, undersampled data using modern neural and variational estimators as well as classical bias-correction methods. Thus we will focus on methods that produce mechanistic and dynamical insight, building toward shared estimators, benchmarks, and open tools that experimentalists can use.  These outcomes connect to the data-driven model discovery happening elsewhere in the program.