acceptodds
Under review as a conference paper at ICLR 2027

State-Gated Learning Shapes the Geometry of Cooperation and Competition in Recurrent Networks

Abstract

Learning reshapes how units in a recurrent network cooperate and compete. When learning conditions change, can interactions formed earlier remain functionally relevant under the same present conditions? We address this question with State-Gated Explaining-Away Plasticity (SG-EAP), a normative recurrent model in which a supplied state changes activity regularization and inhibitory gain during reconstruction learning. Different learning states bias connectivity among similarly driven units toward cooperative or competitive geometry. These biases remain detectable during adaptation even when networks with different histories share the same final learning condition and evaluation inputs. We then intervene on the retained geometry within the same matched-history pair. Under a common low-gate condition, implanting the similarity-aligned component from the previously high-gate network improves classification with missing features, whereas replacing that component with the low-history value reduces it. Each edited representation is evaluated with a linear readout refitted on complete responses. The reciprocal effect exceeds all 16 tested equal-norm orthogonal exchanges in each of ten independent input batches. It explains part of the history-dependent performance difference and depends on the recurrent background. These results identify an interpretable structural trace through which earlier learning can partially shape computation under feature loss during ongoing adaptation.

Primary area:
applications to neuroscience & cognitive science
Keywords:
state-gated learning, recurrent neural networks, learning history, cooperation and competition, computational neuroscience

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 2 reviews, above 9% of the others.
Reviewer_5j3yOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The authors ask whether a trace of past learning is retained in networks. They focus on statistics of connections between neurons that receive similar inputs. By training networks in a manner that encourages either competition or cooperation in such cases, they test all four combinations of two sequential regimes. They find that the first regime leaves a trace even when performing an identical second regime. They also show that exchanging this component of connectivity between networks and refitting readouts can partially transfer their performance.

Critical strengths

The question of retention of learning history is an interesting one.

Critical weaknesses

The paper is written in a way that makes it extremely hard to make sense of the authors’ claims. The introduction is not very coherent. For instance, where do “pressures” come from? What is “clean-only”? These are defined later on, or in appendix, but this makes reading the paper harder. When referring to “The compact figure”, do the authors refer to “Figure 1”? The significance of the trace is unclear, as it nearly disappears with extended training, and can also reverse under some conditions.

Questions for authors

See above

Reviewer_b9ycOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper trains a small RNN to de-noise its own inputs, with a scalar value gg jointly determining an activity penalty and a sparsity penalty (both part of the loss) and an inhibitory gain (part of the model's dynamics) - this forms the SG-EAP architecture introduced by the paper. The RNN is split into 80 excitatory neurons and 20 inhibitory neurons. In this setup, the authors define a statistic mWm_W, the slope of excitatory-to-excitatory weights against correlation in the input, as a measure of cooperation (positive mWm_W) versus competition (negative mWm_W). The paper reports 3 findings: (a) training with low gg gives cooperative weights while training with high gg gives competitive weights; (b) even if the value of gg is changed to match for the final 600 training updates, two networks that were trained on different gg initially will still differ in mWm_W as per their original gg; (c) exchanging the component of excitatory-to-excitatory weights aligned with the input similarity between two models (one trained at low gg throughout and another trained at high gg initially, then low gg) demonstrates a shift in task accuracy in the direction predicted by the historical difference in gg.

Critical strengths

The paper makes significant efforts to isolate the effects of the historical discrepancy in gg. Each compared pair of networks has the same final value of gg, input noise stream and a re-initialised optimiser. In general, the experiments carefully control for confounding variables such that the results can be attributed to the difference in training history rather.

Similarly, the authors are careful to note that their observations prior to section 4.2 merely correlate mWm_W with the observed functional differences due to different training histories. In section 4.2 they move towards a causal relationship by exchanging the mWm_W coordinate between two models with different training histories and measuring the impact on task accuracy.

Critical weaknesses

It's unclear to me that the measured differences between histories are not just a result of incomplete training. None of the networks seem to converge, so networks that start the final 600 updates with different weights would probably differ anyway. As stated in the paper, in most cases, the effect also seems to fade with longer training (appendix C.12).

I think the paper could be presented with more intuition and a clearer final message. As it stands, I feel that the paper's contents are mostly just observations about the constructed setup, rather than a generalisable finding or some deeper understanding of learning dynamics, representational geometry etc.

Questions for authors
  1. Is the difference between HL and LL larger than the difference between two different LL runs that start the final phase with different weights?
  2. Similarly, do the history effects remain at all after convergence? I think appendix C.12 mostly answers this, but do you have an explanation for why the common-high structural trajectories diverge while the common-low trajectories converge?
  3. In the discussion, you mentioned "high-gate representations transfer better from complete to masked inputs"; this is an interesting finding, do you have any intuition on why this might be?
  4. More broadly, is there a conclusion that you can draw from these findings that extends beyond the particular setup of SG-EAP?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.