Do Mice Grok? Glimpses of Hidden Progress During Overtraining in Sensory Cortex (2411.03541v2)

Abstract: Does learning of task-relevant representations stop when behavior stops changing? Motivated by recent theoretical advances in machine learning and the intuitive observation that human experts continue to learn from practice even after mastery, we hypothesize that task-specific representation learning can continue, even when behavior plateaus. In a novel reanalysis of recently published neural data, we find evidence for such learning in posterior piriform cortex of mice following continued training on a task, long after behavior saturates at near-ceiling performance ("overtraining"). This learning is marked by an increase in decoding accuracy from piriform neural populations and improved performance on held-out generalization tests. We demonstrate that class representations in cortex continue to separate during overtraining, so that examples that were incorrectly classified at the beginning of overtraining can abruptly be correctly classified later on, despite no changes in behavior during that time. We hypothesize this hidden yet rich learning takes the form of approximate margin maximization; we validate this and other predictions in the neural data, as well as build and interpret a simple synthetic model that recapitulates these phenomena. We conclude by showing how this model of late-time feature learning implies an explanation for the empirical puzzle of overtraining reversal in animal learning, where task-specific representations are more robust to particular task changes because the learned features can be reused.

• The paper reveals that overtraining yields hidden progress in sensory cortex as mice refine odor representations even when behavior appears to plateau.
• A margin maximization mechanism, demonstrated through increased decoding accuracy, drives the continued neural differentiation observed during overtraining.
• A biologically-plausible model bridges these findings to deep learning grokking, suggesting new directions for accelerating adaptive learning in reversal tasks.

Insightful Overview of "Do Mice Grok? Glimpses of Hidden Progress During Overtraining in Sensory Cortex"

The paper "Do Mice Grok? Glimpses of Hidden Progress During Overtraining in Sensory Cortex" investigates the phenomenon of continued task-specific representational learning in mice, specifically within the posterior piriform cortex (PPC), despite behavioral performance plateauing. This concept is inspired by the grokking discovery in deep learning—a process wherein neural networks achieve successful generalization during an overtraining phase after reaching ceiling training accuracy. The authors explore the theoretical and empirical parallels between representational learning in biological systems and deep learning models, drawing insights from both neuroscience and machine learning frameworks.

Key Findings

Through a comprehensive reanalysis of neural data originally recorded by Berners-Lee et al., 2023, the paper presents compelling evidence suggesting substantial representational shifts during overtraining in mice. Key findings include:

• Separation of Odor Representations: The paper finds that as mice undergo overtraining, the neural representations of target and non-target odors within the piriform cortex continue to separate, even without observable changes in mouse behavior. This separation is inferred from increased decoding accuracy over the overtraining period.
• Margin Maximization Dynamics: The maximum margin classifier's margin increases during overtraining, implying that representations evolve in a way that incorrect classifications at training start become correct at the end, reinforcing the importance of margin maximization as a mechanism for improving task performance.
• Biologically Plausible Model: The authors develop a synthetic model replicating mouse PPC's dynamics, which accounts for the grokking phenomenon observed in deep learning. This model suggests margin maximization as a pivotal driver of task-specific representational changes during overtraining.
• Overtraining Reversal Effect: The paper suggests that features learned during overtraining can contribute to faster adaptation to reversed tasks, providing neural insights into cognitive phenomena studied in psychology.

Theoretical and Practical Implications

The implications of these findings are notable both theoretically and practically in multiple domains:

• Theoretical Insights into Grokking: The results support theories linking latent feature learning with grokking, emphasizing the significance of margin maximization as a universal mechanism driving learning dynamics during overtraining periods.
• Neuroscientific Advances: This paper highlights the potential for continued representational learning in sensory cortical areas despite behavioral performance appearing stable, suggesting that our understanding of how the brain encodes and refines sensory information could be significantly expanded.
• Machine Learning Parallel: The insights derived from biological systems could inform deep learning research, particularly in developing models that mimic the biologically observed benefits of representational drift and margin maximization.

Future Directions

The authors speculate on several promising avenues for future research:

• Experimental Validation: Rigorous experimental frameworks to validate these findings across different sensory systems and task difficulties could enhance the robustness of conclusions drawn regarding rich learning dynamics.
• Generalization of Late-Time Learning: Understanding whether such overtraining benefits exist universally across different species and task configurations remains a critical question requiring further empirical scrutiny.
• Modeling Biological Learning Mechanisms: Deep learning models might benefit from integrating biological principles of continuous feature evolution and implicitly biasing towards task-generalizing solutions.

Conclusion

The paper presents intriguing insights into the parallel developments of biological and artificial learning systems, providing a nuanced understanding of how representational learning can continue even after behavioral mastery. By bridging theories from machine learning with experimental neuroscience, the paper suggests a shared underlying mechanics in learning dynamics, potentially offering a holistic framework for interpreting complex learning phenomena across domains.