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

Hardik Rajpal, Dan GoodmanarXiv 2026 · arXiv

TL;DR

How do different neural networks—whether biological or artificial—arrive at the same internal representations despite having completely different "hardware" and initializations? This paper introduces the Emergent Adaptive Representation Learning (EARL) framework, revealing that networks first become invariant to their own internal parameters before they eventually learn to compress their state space to capture a task's specific structure.

What Problem Does This Paper Address?

In both neuroscience and AI, we observe a phenomenon called substrate invariance: different brains or differently initialized neural networks often converge on similar representations when performing the same task. In the brain, this happens despite "representational drift" (synapses constantly changing), and in AI, it happens across various architectures.

The authors seek to quantify how and when this invariance emerges during training. They tackle the gap between "micro-level" activity (individual neurons) and "macro-level" representations (coordinated population activity), asking whether the network first learns to ignore its own internal noise/parameters or first learns the structure of the data.

Methods

The authors propose the EARL framework, which distinguishes between two modes of emergent compression:

  1. State-space compression ($\Psi$): This occurs when a coarse-grained description of the system (the representation) predicts its own future better than the individual units can. It’s a measure of "causal emergence."
  2. Parameter-space compression ($\Lambda$): This occurs when many different internal configurations (parameters) lead to the same macroscopic behavior. It’s a measure of "degeneracy" or "universality."
Figure 1: The two modes of compression. A State space: a stable macro configuration emerges as microscopic fluctuations are compressed away. B Parameter space: the same configuration emerges from different models of microscopic interactions.

To test this, they use a reservoir-bottleneck architecture. A fixed, random recurrent reservoir (the "substrate") processes input, while a small, trainable "bottleneck" layer learns to represent the underlying dynamics to predict the next step. They formalize both modes using information-theoretic measures (Gaussian mutual-information estimates) that can be calculated from recorded activity without needing access to the network's gradients.

Figure 3: Reservoir–bottleneck architecture, adapted from our previous work [31]. Random orthogo- nal embeddings of the attractor latent Lt drive a fixed random reservoir with states Xt; a trainable linear bottleneck maps Xt to Zt, whose top three principal components form the learned representa- tion Vt; a ridge readout predicts the next input. The bottleneck is the only trained component.

Datasets

The study uses synthetic datasets consisting of three-dimensional chaotic attractors (Sprott B, Rössler, Lorenz, Chen, and Sprott A).

  • Training: 12 sequences from three attractors.
  • Evaluation: Unseen rotations of training attractors and entirely "held-out" attractors to test generalization.
  • Controls: "Time-permuted" and "cross-sequence" targets were used to see if the compression modes were truly learning the task or just reacting to the input statistics.

Key Insights and Ideas

  • Invariance Before Structure: The most striking finding is the timeline. Parameter-space compression ($\Lambda$) rises almost immediately, often regardless of whether the network is actually learning the task correctly. State-space compression ($\Psi$), however, only begins to rise much later, once the network starts mastering the temporal structure of the data.
  • Two-Phase Learning: Learning isn't a monolithic process. There is an initial "fitting" phase where the representation becomes independent of the specific reservoir parameters, followed by a "discovery" phase where the network extracts the task-relevant latent variables.
  • Task-Independent Invariance: The early rise in parameter-space compression happens even with "garbage" targets (permuted data), suggesting that the pressure to compress through a bottleneck naturally forces a degree of universality before any real "learning" happens.

Key Results

  • The Timing Gap: Parameter-space compression reaches its half-maximum by roughly epoch 8, while state-space compression takes until roughly epoch 40 to reach the same milestone—a gap of about 32 epochs.
  • Generalization: Both measures of compression were found to be consistent across training sets, unseen rotations, and held-out attractors, suggesting these are robust properties of the learning process.
  • Quantitative Gains: Over 100 epochs, models trained on true targets gained significantly more compression than control models (approx. 115 nats in $\Psi$ and 0.10 in $\Lambda$).
Learning dynamics showing the early rise of $\Lambda$ (C) compared to the later recovery of $\Psi$ (B).

Limitations and Future Work

  • Relative Measures: The authors note that $\Psi$ is typically negative in large networks because the sum of individual unit information over-counts shared redundancy. Therefore, the trend is more important than the absolute value.
  • Requirement of Latents: Calculating $\Lambda$ currently requires knowledge of the "true" task-relevant latent variables ($L_t$), which may not always be available in complex real-world datasets.
  • Future Directions: The authors suggest applying these measures to biological data (like hippocampal or prefrontal cortex recordings) to see if the same "invariance before structure" pattern holds true in living brains during learning. They also aim to explore datasets exhibiting "representational drift" to see if $\Lambda$ can track how brains maintain stable meanings despite changing neurons.

Summary updated

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.
BibTeX
@misc{rajpal2026invariance,
  author = {Hardik Rajpal and Dan Goodman},
  title = {{Invariance before structure: two modes of emergent compression in representation learning}},
  year = {2026},
  eprint = {2611.07952},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.07952},
}
Figure