Do Mice Grok? Glimpses of Hidden Progress in Sensory Cortex

Tanishq Kumar, Blake Bordelon, Cengiz Pehlevan, Venkatesh N Murthy, Samuel J. GershmanICLR 2025 · OpenReview

When an animal—or an artificial intelligence—reaches "perfect" performance on a task, does its brain stop learning? A recent paper bridges a fascinating gap between deep learning theory and systems neuroscience, suggesting that the sensory cortex continues to refine its internal representations long after a subject's outward behavior has plateaued.

By reanalyzing neural data from mice, researchers found evidence of "hidden progress" that closely mirrors the AI phenomenon known as "grokking." This ongoing neural refinement provides a compelling biological explanation for why overtraining helps animals adapt more quickly to new, related challenges.

The Mystery of the Behavioral Plateau

In machine learning, researchers have recently documented a phenomenon called grokking: a model might reach 100% accuracy on its training data, but then, after thousands of additional "useless" training steps, it suddenly learns to generalize to unseen data. The model's internal representations continue to restructure themselves even when the training error is zero.

In psychology, it is well known that human experts continue to improve their internal "models" of a task even after they stop making overt mistakes. However, in neuroscience, researchers typically stop recording once an animal "masters" a task. Because of this, it has remained unclear whether the brain's internal representations remain static during a behavioral plateau, or if they undergo a rich, hidden transformation similar to grokking in artificial neural networks.

Searching for Biological Grokking

To find evidence of this hidden learning in the biological brain, the researchers took a two-pronged approach:

  1. Neural Data Reanalysis: They revisited a 2023 dataset from the Murthy Lab (Berners-Lee et al.) involving tetrode recordings from the posterior piriform cortex (PPC)—the region of the mouse brain responsible for processing smells. Mice were trained to distinguish one "target" odor from hundreds of others. Crucially, the recordings continued for 18 days after the mice reached near-perfect behavioral performance (a phase known as "overtraining").
  2. Synthetic Modeling: They built a simple Multi-Layer Perceptron (MLP) and a more biologically plausible recurrent model to simulate the odor task. They used these models to test if specific mathematical loss functions could explain the representational shifts observed in the mice.

Key Findings: Hidden Progress in the Cortex

The reanalysis revealed that even when the mice's licking behavior was identical on Day 8 and Day 26, their neural activity was not. The brain was still hard at work.

Decoding Accuracy Continues to Rise

Linear decoders trained on the neural population activity showed increasing accuracy throughout the overtraining period. Even though the mouse's actual behavior was already at a ceiling, the neural signals encoding the task were becoming clearer and more robust.

Decoding accuracy continues to rise during the overtraining period.

Representational Separation and Margin Maximization

In AI, models often learn to generalize better by finding the "widest" possible gap—or margin—between different classes of data. The authors found evidence of a similar biological margin maximization. The "margin" of a linear classifier trained on the mouse's neurons increased significantly during overtraining. Essentially, the brain was building a "buffer zone" for its decisions.

This buffer zone is visible in how the neural representations of different odors drift apart. Pairwise correlations between target and non-target odor responses became significantly more negative (more distinct) over the overtraining period.

Representational similarity matrices show target and non-target representations separating over time.

Solving the "Overtraining Reversal" Puzzle

Psychologists have long noted a counterintuitive phenomenon: overtrained animals learn to "reverse" a task (e.g., learning to lick for the opposite smell) faster than animals that were barely trained to proficiency.

The authors' synthetic model demonstrated that overtraining on easy odors led to a sudden "grokking" jump in accuracy on difficult, unseen probe odors. Because overtraining builds richer, more separated features, the brain doesn't have to relearn the features from scratch when the rules change; it can simply "flip" its readout.

Furthermore, when the researchers replaced the Cross-Entropy loss in their model with a Hinge loss—which stops learning once a "good enough" margin is reached—the grokking effect vanished. This suggests the biological brain might be optimizing an objective function that never truly "settles," constantly pushing for better representations.

Limitations and Future Directions

While these findings offer a compelling bridge between AI theory and neuroscience, the study is observational. The strong correlations observed do not definitively prove that these representational changes cause the improved generalization seen in mice. Additionally, the sample size of recorded neurons per day (~20) is relatively small; modern techniques like Neuropixels could provide a much higher-resolution view of this process in the future.

Finally, the biological "loss function" driving this change remains a mystery. Is the brain driven by metabolic efficiency, noise reduction, or an inherent drive to minimize uncertainty? And does this "grokking" occur in other areas, like the visual or motor cortex, during expert-level training? Answering these questions could fundamentally change how we understand learning, mastery, and the hidden life of the brain.

Summary updated

Abstract

Does learning of task-relevant representations stop when behavior stops changing? Motivated by recent work in machine learning and the intuitive observation that human experts continue to learn after mastery, we hypothesize that task-specific representation learning in cortex can continue, even when behavior saturates. 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"). 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 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 demonstrating 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.
BibTeX
@inproceedings{kumar2025do,
  author = {Tanishq Kumar and Blake Bordelon and Cengiz Pehlevan and Venkatesh N Murthy and Samuel J. Gershman},
  title = {{Do Mice Grok? Glimpses of Hidden Progress in Sensory Cortex}},
  booktitle = {International Conference on Learning Representations},
  year = {2025},
  url = {https://openreview.net/forum?id=oYemKnlIrO},
}
Figure