learning-fast-and-slow.github.io
Learning in the brain is a complex process occurring over multiple timescales, from rapid adaptation to prolonged practice-dependent changes. Rapid motor adaptation, statistical learning, even one-shot learning happens within minutes or hours, whereas slower processes, like acquiring new motor skills, perceptual priors, and complex cognitive abilities, may take days, weeks, or even longer. Often, these learning processes have been studied in isolation. However, animals and their brains do not operate with clear boundaries between different learning types, especially in dynamic and unpredictable environments. Instead, learning can occur on a continuum of timescales and depend on interrelated neural mechanisms, with differential recruitment of parallel or sequential processes across tasks. By integrating perspectives from both fast and slow learning paradigms, this workshop seeks to uncover the underlying principles that govern this continuum, the specific roles played by various brain regions and circuits, as well as how to experimentally disambiguate between learning processes. Two common frameworks will be used to guide discussion and unify insights from multiple perspectives: Learning rules: differentiated based on feedback structure (e.g., supervised, unsupervised, reinforcement learning), activity dependence, and brain regions. Evolution of neural population activity: with an emphasis on changes to latent dynamical structure and representational geometry that support new task-relevant computations. A discussion on distributed and multi-timescale learning mechanisms is particularly timely due to recent advances in three key areas: (1) the ability to chronically monitor and perturb large neural populations, including multiple brain areas simultaneously, (2) the explosion of statistical and machine learning tools to extract structure from high-dimensional behavioral and neural data, and (3) insights from training artificial neural networks with bio-inspired architectures on increasingly complex behaviors. The workshop is designed for experimentalists and theoreticians working across different neural systems, with the goal of fostering not only the exchange of empirical findings but also mathematical tools and methodologies for studying learning mechanisms. This includes a discussion of identifiability issues and other challenges of fitting models to behavioral data, neural recordings, and perturbation experiments. This broader community can inspire new theory-guided experimental designs for testing specific predictions and enable the integration of theory and data on multiple levels – behavior, population activity structure, and cellular mechanisms. We will also discuss more conceptual questions such as the distributed nature of learning, what governs the dominance (if any) of different learning mechanisms or circuits, and when such degeneracy may lead to competitive or synergistic interactions. We will try to work towards a more holistic understanding of biological learning by re-emphasizing behavioral modeling, formalizing dynamics of multi-site plasticity, and incorporating a diversity of dense and sparse teaching signals.
