CLAUDIA CLOPATH: Local learning in recurrent neural networks 

LUKE CODDINGTON: Midbrain dopamine signals for optimizing new behaviors

CELINE DRIEU: Rapid emergence of latent knowledge in the sensory cortex drives learning

KATE NUSSENBAUM: Learning adaptive valence biases across development

JONATHAN PILLOW: Inferring learning rules from animal training data

CRISTINA SAVIN: Learning by directly modulating neural dynamics

ANDREW SAXE: An analytically tractable model of optimal replay for schema learning and few-shot generalization

ALISON COMRIE: Selective generation of neural dynamics for learning in changing environments

LAURA GRIMA: Algorithms for learning to forage from scratch

Animals in the real world often face new circumstances, requiring the brain to learn and flexibly adapt under ambiguity. In particular, survival in novel, complex, and changing environments critically depends on an animal’s ability to rapidly learn and make inferences based on sparse, delayed, and unreliable feedback. What processes enable naive biological and artificial networks to adapt so efficiently to new situations in a ‘few shot’ manner?

So far, work on learning algorithms in neuroscience and machine learning has focused mainly on the computations underlying adjustments to contingency changes once the structure of the task or environment is already known. These approaches have elucidated mechanisms underlying how behavior can be adapted and fine-tuned. However, we still lack generalizable models of how brains “learn from scratch,” particularly in the rich complexity of the real world – where perhaps the most powerful forms of learning occur.

This workshop will therefore focus on initial learning across a variety of domains that require naive agents to infer latent structure, identify relevant variables and goals, construct representations and predictions, bootstrap from limited feedback, and develop internal knowledge - sometimes even before clear behavioral markers of learning emerge. How cognitive and motor domains interact as part of this learning, and the roles of prior beliefs, inductive biases, and development in acting as a scaffold for naive learning, are key discussions this workshop will tackle. 

We believe this topic is incredibly timely; modern machine learning has made progress in meta-learning, self-supervised learning, and interpretable neural networks, offering theory and tools for formalizing how naive agents extract knowledge through experience. In parallel, there is a growing body of neuroscientific work focused on the rapidity with which biological organisms can learn from naive, particularly in complex environments. In combination with longitudinal neural recordings, this work is beginning to allow us to monitor fast inference of latent structure, development of internal representations, and updating knowledge from limited experience. By bringing these fields together, this workshop will provide a forum for dialogue and the exchange of inspiration between biological and computational approaches to learning in naive systems. 

To that end, we aim to unite computational and experimental leaders at the forefront of studying learning from scratch, who:

• Model learning from the earliest stages of naive exploration

• Leverage behavior for mechanistic insight into circuit computations and plasticity rules

• Expose latent learning processes that establish internal knowledge before overt behavioral expression

• Incorporate inductive biases and consider agents’ priors in models

• Bridge biology and theory by uniting clever experimental design with reinforcement learning, generative modeling, dynamical systems, and neural network approaches

Together, we will highlight the unique challenges - and advantages - of studying early learning, and identify new directions for this fundamental area of research. This workshop will be of interest to a range of community members interested in learning rules, plasticity, neural dynamics across brain regions, neuromodulation, flexible cognition, decision making, behavior, and intelligence broadly. Collectively, we aim to establish a roadmap for identifying the principles that enable brains to rapidly transform new experiences into structured knowledge for adaptive behavior.