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

AI summary

Overview

Research area: Representation learning and computational neuroscience — specifically, information-theoretic measures of emergence in neural systems (both biological and artificial). Technical level: Intermediate. The core argument is conceptually accessible, but the paper relies on information theory (mutual information, Fisher information), causal emergence formalism, and reservoir computing. Scope: The paper proposes the Emergent Adaptive Representation Learning (EARL) framework, which distinguishes two modes of emergent compression — state-space and parameter-space — and measures both across training in a reservoir–bottleneck model.

What This Paper Is About

Different brains, and different artificial networks trained from different initialisations, often end up representing the same stimuli in strikingly similar ways, even though their underlying components (neurons, weights) differ wildly. The paper asks how a learned representation becomes independent of the "substrate" that produces it, and argues that this happens through two distinct forms of compression: the network's state space becoming more self-predictive, and its parameter space collapsing so that many different microscopic settings yield the same representation. The goal is to formalise both, build measurable proxies for each, and track when they appear during learning.

Key Contributions

1. The EARL framework, which separates two modes of emergence in learned representations: state-space compression (a representation predicts its own future better than its individual units do) and parameter-space compression (networks with different micro-level parameters converge to the same representation).
2. A Fisher-information formalisation of both modes (Appendix B), including a total derivative for parameter-space compression that accounts for how trained weights adapt when the substrate parameters change.
3. Two ensemble-level, gradient-free information-theoretic estimators: Ψ (an adapted causal-emergence measure with a second-order redundancy correction) and Λ (a degeneracy measure based on a held-out Barber–Agakov lower bound, with shuffled-label nulls).
4. An empirical demonstration that the two modes occur in distinct phases: parameter-space compression rises early and largely independently of the task, while state-space compression rises roughly 30 epochs later and only when the network learns the task's temporal structure.

Main Findings

• Learning works as expected on the task: one-step prediction error falls monotonically on all three evaluation sets (training attractors, unseen rotations, held-out attractors), and the learned representation becomes increasingly aligned with the true latent.
• State-space compression is non-monotonic: Ψ first decreases, reaching its minimum by about epoch 10, then rises for the rest of training, reaching half its rise by about epoch 40 on all three evaluation sets. Ψ is negative throughout, which the authors attribute to the 500 units' shared redundancy.
• Parameter-space compression rises immediately: Λ climbs from about 0.78 to 0.96 on the training sequences (ΔΛ = 0.177 ± 0.020, mean ± s.d.), with similar changes on the other sets, and increases on every set (two-sided sign test, p = 1.9 × 10⁻⁶).
• Parameter-space compression leads state-space compression by about 32 epochs: Λ reaches half its total change by about epoch 8, versus about epoch 40 for Ψ; Λ is never later than Ψ (sign test, p ≤ 1.5 × 10⁻⁵).
• Controls separate task-independent from task-dependent phases: with a time-permuted or cross-sequence target, Ψ still dips and Λ still rises to about 0.85, but Ψ never recovers and Λ plateaus. Models trained on the true target gain more than their matched controls (paired sign test, p = 0.002 for each control and set): by about 115 nats in Ψ and 0.09–0.10 in Λ on the training sequences.
• Noisy input preserves the pattern: with σ_y = 0.1, Λ approaches 1 in trained runs and plateaus near 0.95 in controls, while the ordering of the two modes and the differences from controls are preserved.
• Robustness to estimator choices: varying the number of random projections or the estimator (Gaussian Barber–Agakov vs. multinomial-logistic bound) shifts the level of Λ but not its rise, and the retained information about θ stays above the shuffled-label null.

Methodology in Plain English

The authors use a deliberately simple model so that "substrate" and "representation" can be cleanly separated. A fixed, random recurrent network (a reservoir of 500 leaky-tanh units) receives inputs; only a small 10-dimensional linear bottleneck on top of it is trained, via backpropagation through a ridge-regression readout that predicts the next input. The learned representation is taken as the top three principal components of the bottleneck activity.

The inputs are 3D chaotic attractors (Sprott B, Rössler, Lorenz for training) embedded into 10 dimensions, with held-out attractors (Chen, Sprott A) and rotated versions used for testing.

To measure parameter-space compression, they train an ensemble of networks that share the same reservoir draw but vary five reservoir hyperparameters (spectral radius, input scaling, bias scaling, leak rate, ridge penalty) across 100 settings drawn log-uniformly within ±30% of an Optuna optimum. Λ then asks: how much information does the learned representation still carry about which hyperparameter setting produced it, compared with how much the raw reservoir activity carries? A value of 1 means the representation has become blind to the substrate.

To measure state-space compression, they compute Ψ: the information the representation carries about its own next state, minus the sum of information each individual unit carries about that future — with a correction adding back pairwise redundancy. Both measures are estimated from recorded activity using Gaussian mutual information, without needing gradients.

Why This Matters

The paper offers a quantitative handle on a puzzle that spans neuroscience and machine learning: why do structurally different systems produce the same representations? It suggests a specific ordering — substrate invariance comes first, and task-specific structure is built on top of it — which reframes how we think about what training does in its early epochs.

Potential applications implied by the work (the paper itself reports no applied experiments):

• Neural data analysis: measuring parameter-space compression across animals or across sessions could quantify representational stability in the face of drift and synaptic turnover.
• Model comparison and reproducibility: Λ compares models by what their representations encode rather than by aligning representations directly, offering a substrate-agnostic similarity metric.
• Robustness and generalisation assessment: emergence measures have previously been linked (in the authors' earlier work) to generalisation to unseen rotations and dynamical systems.
• Training diagnostics: because the two measures diverge in time, they could act as early indicators of whether a network is learning genuine temporal structure or merely fitting.

Industry relevance: for teams building recurrent or sequence models, the framework offers a principled diagnostic for distinguishing "the model has become stable" from "the model has learned the task" — useful for early stopping, transfer, and evaluating whether an ensemble of differently initialised models is genuinely converging.

Future Directions

• Extend the measures to other tasks by replacing V_{t+1} in Ψ and F(V) in Λ with task-appropriate quantities; the paper notes this is possible but does not demonstrate it.
• Measure parameter-space compression in biological data, including datasets exhibiting representational drift — a natural follow-up given the authors' prior observation of Ψ's non-monotonic trajectory in hippocampal and prefrontal recordings.
• Address the known limitations of the measures: Ψ is negative throughout and its trend depends on temporal sampling of the data; Λ requires a task-relevant latent and measures invariance across models rather than robustness of a single model. Higher-order redundancies beyond the pairwise correction are also not accounted for.
• Connect the information-theoretic proxies more tightly to the Fisher formulation, since the main-text estimators are analogues of the Fisher sensitivities derived in Appendix B rather than direct estimates of them.

Target Audience

Researchers in representation learning, computational neuroscience, and complex-systems/emergence research who are interested in why different networks converge to similar representations. It will be most useful to readers comfortable with mutual information and linear readouts; those seeking a purely empirical benchmark comparison will find the single reservoir architecture limiting, while those interested in measurable definitions of emergence will find the operationalisation of Ψ and Λ the paper's main value.

Authors’ 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.