Learned neural representations show key phases of compression and invariance

Invariance before structure: two modes of emergent compression in representation learning

Machine LearningNeural and Evolutionary Computing

Summary

Some neural networks and brains can learn useful patterns that stay stable despite differences in their tiny parts. The authors describe two ways this happens: one where the network's overall state becomes simpler while still predicting what will come next, and another where different networks end up with very similar overall patterns even if their internals differ. They study these in a setup where networks learn tasks over time and find that the phase of becoming similar across different networks happens early, while the phase of understanding the task structure happens later. This shows that stability and task-specific learning develop separately.

What this means in practice

  • For machine learning engineers: Design training schedules that recognize early substrate invariance and later task-specific structure emergence to improve neural network performance.
  • For cognitive modelers: Use findings on separate phases of representation emergence to build better models of how brains learn stable patterns over time.

Authors

Hardik Rajpal, Dan Goodman

Abstract

Learned representations show striking invariance across biological and artificial neural networks, with different substrates and variations driven by drift, heterogeneity and initialisation. To understand this invariance, we propose Emergent Adaptive Representation Learning (EARL), a framework that distinguishes two modes of emergence in learned representations: state-space compression, where a learned representation gains predictive power over its own future beyond what the network's units can predict individually, and parameter-space compression, where networks with different micro-level parameters converge to the same representation. We formalise both as Fisher sensitivities and introduce ensemble-level information-theoretic measures for each. In a recurrent reservoir with a trained bottleneck, parameter-space compression rises within the first few epochs of training and partly independently of the task, whereas state-space compression rises about 30 epochs later and only when the network begins to learn the temporal structure of the task. These results suggest that invariance to the substrate and task-specific emergent structure are acquired in distinct phases of learning.