acceptodds
Under review as a conference paper at ICLR 2027

Lateral Inhibition Induces Adaptive Sparsity Resulting in Formation of Interpretable Representations in Autoencoders

Abstract

Lateral inhibition in neurobiological systems achieves sparse activity through recurrent competitive dynamics. Research on sparse autoencoders (SAEs) shows that imposing sparsity facilitates the formation of interpretable features in neural representations. Conventional SAEs, however, typically enforce sparsity through externally specified regularization objectives or activation constraints. To investigate how recurrent lateral competition shapes the learning of interpretable features, we introduce the Lateral-Inhibition Sparse Autoencoder (LISAE), which uses lateral inhibition to generate sparse activity while learning interpretable features in its decoder. Across controlled MNIST datasets and Gemma-2 representations, we show that LISAE recovers interpretable features while enabling input-dependent sparsity, with the number of active neurons adapting to the input structure. Furthermore, we show that anti-Hebbian-trained lateral inhibitory connections capture the relational structure of the feature space, which reflects the underlying similarities among features. Together, these findings establish local lateral inhibition as an alternative mechanism for learning adaptive sparse and relational representations in autoencoders.

Primary area:
interpretability and explainable AI
Keywords:
Lateral Inhibition, Sparse Coding, Interpretable Representation, Autoencoder

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_7d3bOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces LISAE, an autoencoder that selects sparse latent activity through recurrent lateral inhibition. Encoder and decoder weights learn through reconstruction gradients, while inhibitory connections and thresholds follow local anti-Hebbian and homeostatic updates. The implemented reconstruction uses an amplitude-preserving gate: the equilibrium activity selects which feedforward activations reach the decoder. Experiments on MNIST, superposed digits, bar stimuli, and one layer of Gemma-2-9b-it show recognizable decoder features, input-dependent numbers of active latents, and structure in the learned inhibitory matrix. The language-model validation includes qualitative steering examples and semantic-group visualizations. The paper also proposes a theorem about realizable equilibrium supports.

Critical strengths

Interesting integration of local competition with reconstruction learning. The combination of anti-Hebbian inhibition, homeostatic thresholds, and gradient-trained encoder/decoder weights provides a concrete mechanism for sparse feature selection.

Critical weaknesses
  1. It is unclear what LISAE adds over existing sparse autoencoders. Many SAE architectures (Vanilla [1], JumpReLU [2], BatchTopK [3], etc.) already allow the number of active features to vary with the input. The paper does not show that LISAE uses this flexibility better at similar sparsity and reconstruction quality. It also does not compare the semantic relationships in the inhibitory matrix with simpler measures such as activation co-occurrence or decoder similarity. These comparisons are needed to establish what the inhibition adds.

  2. The language-model evidence is too limited to justify the cost. The paper trains one 7,168-latent SAE on one layer using 500,000 tokens, reused for 20 epochs, which is unusually small scale. Two steering examples do not establish monosemanticity or show how reliably steering works while preserving fluency. There is no quantitative comparison of feature quality or downstream fidelity with baseline SAEs. The dense inhibitory matrix also grows quadratically: at width 65,536 it would have about 4.3 billion entries, plus the cost of recurrent settling. The paper acknowledges this limitation, but reports no runtime or memory measurements and shows no advantage that would justify the cost.

[1] Bricken et al. (2023). Towards Monosemanticity: Decomposing Language Models With Dictionary Learning. Transformer Circuits Thread, Anthropic.

[2] Rajamanoharan et al. (2024). Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoders. arXiv:2407.14435.

[3] Bussmann, Leask, and Nanda (2024). BatchTopK Sparse Autoencoders. arXiv:2412.06410.

Questions for authors
  1. Does the inhibitory matrix reveal semantic relationships better than activation co-occurrence or decoder similarity?
  2. How many settling steps are used, and what is the runtime and memory cost compared with a feedforward SAE?
  3. Could you provide a quantitative comparison of feature quality or steering with a baseline SAE?
3 comments
AuthorsOpenReview ↗

Response to Reviewer 7d3b - Part 1

We thank Reviewer 7d3b for your constructive feedback. As there were shared concerns regarding scalability and quantitative benchmark reporting across the reviews, we kindly refer the reviewer to our "General response to common concerns raised by reviewers" for an overarching discussion of our motivation. Here, we will focus directly on addressing the specific issues raised in your review.

Part 1

Regarding specific weakness:

  1. Issue on Adaptive Sparsity: The reviewer points out that existing methods (Vanilla, JumpReLU, BatchTopK, etc.) already allow the number of active features to vary with the input. To address this, we have added new experiments in Section 3.2 and Appendix F, training Vanilla, JumpReLU, and BatchTopK SAEs on our controlled, spatially superposed MNIST dataset.

    Indeed, soft L1-penalty methods (Vanilla and JumpReLU) exhibit seemingly adaptive capabilities, as their activation counts do grow proportionally with input complexity. However, when inspecting the underlying learned decoders, their features are often uninterpretable, even though their reconstruction loss remains similarly low. Conversely, hard L0-penalty methods like BatchTopK (and standard Top-K in Figure 1a) lack this adaptive scaling, yet the learned features seem more interpretable. This stark contrast indicates that L0 regularization plays a fundamentally different role in shaping the network's inductive bias. LISAE distinguishes itself from all these baselines because it successfully imposes an adaptive L0 regularization. It achieves markedly better results when the dataset underlying structural complexity varies greatly, imposing a distinct inductive bias on what can be learned [1], the theoretical implications of which require further investigation.

  2. Issue on "uses this flexibility better at similar sparsity and reconstruction quality": Again, our theoretical analysis and experiments in Section 3.2 and Appendices E, F could serve as an intial evidence that lateral inhibition imposes adaptive L0 penalty, achieving markedly better results when the dataset underlying structural complexity varies greatly, which may affect what utimately can be learned. Besides, we have shown LISAE has significantly lower number of inactive neuron in Section 3.1 due to the homeostatic regulation of firing thresholds, which is also an extra benefits of our model.

  3. Issue on "no comparison of the semantic relationships in the inhibitory matrix with simpler measures such as activation co-occurrence or decoder similarity":

    1. First, activation co-occurrence is exactly what anti-hebbian training has learned. Consequently, the lateral inhibitory connections grows proportional to the co-occurence and feature similarity.
    2. Second, if we compare the learned lateral matrix against direct decoder weight similarity, the learned connections will not quantitatively outperform a direct calculation. Because the decoder features in Figure 1a already cleanly resemble distinct digits, calculating their direct similarity is naturally optimal. However, we want to emphasize a conceptual breakthrough: to the best of our knowledge, this is the first work to demonstrate that different sets of connections trained by independent loss functions—in this case, anti-Hebbian trained lateral inhibitory connections—can synergistically shape one another's learning to ultimately yield two complementary, interpretable representations.
  4. Issue on "The language-model evidence is too limited to justify the cost.":

    1. First, as we mentioned to reviewer qxUE, we believe that demonstrating the ability to learn the relational structure in Figure 3b with limited data is actually a strength of our approach. If the decoder were not already largely monosemantic, the inhibitory connections would not be able to capture the semantic distances, and vice versa. I would argue that demonstrating the emergence of such a map under limited data conditions is arguably more persuasive than doing so after potentially overfitting the decoder on a massive dataset.

    2. "no quantitative comparison" We have added an initial SAEBench comparison in the Discussion and Appendix G to provide a quantitative baseline, training a 16k-width LISAE for 250M tokens on Gemma-2-2b representations. While this evaluation shows that our current implementation does not achieve state-of-the-art performance compared to standard SAEs, we emphasize once again the initial evidence presented in Section 3.2 and Appendices E, F: because of its adaptive L0 characteristics, LISAE excels when the dataset contains a large variation in the number of underlying features. This structural adaptivity falls outside the scope of current benchmarks, yet it may ultimately have a much deeper impact on representation learning.

AuthorsOpenReview ↗

Response to Reviewer 7d3b - Part 2

Part 2

  1. Issue on "The language-model evidence is too limited to justify the cost." (continue): 3. "reports no runtime or memory measurements and shows no advantage that would justify the cost." We acknowledge the computational complexity limitations of LISAE and have added a new paragraph in the Discussion detailing the runtime and memory requirements, noting that the current memory footprint is not ideal for industrial scaling. However, this complexity is what allows for qualitatively different characteristics.

         First, as shown in Section 3.2 and Appendix F, lateral inhibition serves as an adaptive L0 regularization that functions fundamentally differently from L1 penalty or fixed Top-K L0 penalties.
    
         Furthermore, this complexity allows the network to co-learn relational representations simultanously. All in all, this architecture does not merely add computational overhead; it introduces genuinely new information processing capabilities to the network.
    
         If one would really want to scale it up, since the $W_{LI}$ does not go through the backpropagation, its anti-hebbian training can be potentially moved to CPU to save GPU resources; but yes, that will be even slower. So a better engineering solution is definitely needed for industrial implementation.
    

Regarding specific questions

  • Does the inhibitory matrix reveal semantic relationships better than activation co-occurrence or decoder similarity? Please refer to our response in Issue 3. The short answer is no—it will not quantitatively outperform direct decoder similarity. Because the learned decoder features already clearly represent the distinct digits, directly calculating their similarity naturally yields an optimal result.

    However, our primary purpose here is not to beat a direct mathematical calculation, but rather to demonstrate a novel mechanistic possibility: that this relational structure can be obtained dynamically through simultaneous co-learning within a single neural network. Our goal is to show that a network can autonomously learn both the individual features and the semantic relationships between them as a map at the same time. This capability may have broader implications. For example, computational neuroscientists may be highly interested in this architecture for learning internal "cognitive maps," which could potentially serve as a model for abstract planning in the hippocampus [2].

  • How many settling steps are used, and what is the runtime and memory cost compared with a feedforward SAE? Please refer to our response in Issue 4.3. We have added a thorough analysis to the Discussion section. As noted in the paper, the recurrent dynamics take approximately 20 settling steps to converge.

    Currently, LISAE requires about 4 times more memory than a standard feedforward SAE. The runtime overhead is also significant and can be calculated as follows: while a standard feedforward SAE scales at O(D⋅N)O(D⋅N) (where DD is the input dimension and NN is the hidden neuron number), LISAE requires O(D⋅N+T⋅N2)O(D⋅N+T⋅N^2)for inference,where T≈20T≈20 represents the settling steps multiplying the N×NN×N lateral matrix. During training, the anti-Hebbian update for the lateral connections adds another O(N2)O(N^2) outer-product calculation, meaning the total training runtime scales proportionally to O(T⋅N2)O(T⋅N^2).

    Because NN is typically much larger than DD, this quadratic scaling is computationally heavy, and we agree it is not yet suitable for industrial scaling. However, as emphasized in Issue 4.3, this complexity is what fundamentally enables the qualitatively different computational capabilities of the network.

  • Could you provide a quantitative comparison of feature quality or steering with a baseline SAE? Please refer to our detailed response for Issue 4.2. We have added the corresponding comparison in Appendix G.

Lastly, I hope the impact of our work could also extend beyond conventional mechanistic interpretability community. For example, leveraging this L0 adaptivity, future work may achieve compositional and continual learning by inhibiting the right groups of neurons in real time via recurrent dynamics. It also highlights an interesting paradigm for algorithm design: allowing different learning and homeostatic rules to work synergistically to achieve complementary representational capabilities in one neural network.

[1]: Hindupur, Sai Sumedh R., et al. "Projecting assumptions: The duality between sparse autoencoders and concept geometry." Advances in Neural Information Processing Systems 38 (2026): 11649-11699.

[2]: Whittington, James CR, et al. "How to build a cognitive map." Nature neuroscience 25.10 (2022): 1257-1272.

Thanks for running the additional experiments and reporting the costs. LISAE underperforms Top-K on the reported Gemma quality metrics, and I don’t see enough evidence of an interpretability gain to justify the extra cost, therefore I decided to keep my rating.

Reviewer_qxUEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Authors propose a new approach for new SAE where a recurrent lateral-inhibition layer chooses which latents fire, instead of strict TopK or an L1 penalty. They train it on a MNIST and gemma-2-9b dataset and show that LISAE recover interpretable features.

Critical strengths

I generally like the paper. It is good written and easy to follow, despite unfamiliar topic (lateral inhibition autoencoders).

The approach is promising.

Critical weaknesses

Hovewer, training a SAE with 7168 features only for 500k tokens - isn't it too small for the SAE? Usually there are 32-64k of features, and at least 100m tokens, tending to 1b tokens. See gemma scope [1,2] as an example of modern scale SAE.

No reports of traditional SAE metrics - FVE, CE loss recovered.

The approach is not scalable. The authors should find a way to scale their approach to full size SAEs (at least 30k, better 50-100k features) and compare the proposed method with modern architectures on SAEBench. Also, please include compute comparison (SAEs are already compute intensive enough, no need to push it further).

[1] Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2

[2] Gemma Scope 2: helping the AI safety community deepen understanding of complex language model behavior

Questions for authors
1 comment
AuthorsOpenReview ↗

Response to Reviewer qxUE

We thank Reviewer qxUE for your constructive feedback. As there were shared concerns regarding scalability and quantitative benchmark reporting across the reviews, we kindly refer the reviewer to our "General response to common concerns raised by reviewers" for an overarching discussion of our motivation. Here, we will focus directly on addressing the specific issues raised in your review.

Regarding specific issues:

  • Issue on Limited Demonstration and No Traditional Score Reporting:

    • First, we believe that demonstrating the ability to learn the relational structure in Figure 3b with limited data is actually a strength of our approach. If the decoder were not already largely monosemantic, the inhibitory connections would not be able to capture the semantic distances, and vice versa. I would argue that demonstrating the emergence of such a map under limited data conditions is arguably more persuasive than doing so after potentially overfitting the decoder on a massive dataset.
    • Second, we have added an initial SAEBench comparison in the Discussion and Appendix G to provide a quantitative baseline, training a 16k-width LISAE for 250M tokens on Gemma-2-2b representations. While this evaluation shows that our current implementation does not achieve state-of-the-art performance compared to standard SAEs, we want to emphasize the new comparison in Section 3.2 and Appendix F: LISAE may achieve fundamentally different feature learning. Because of its adaptive L0 characteristics, LISAE excels when the dataset contains a large variation in the number of underlying features. This structural adaptivity falls outside the scope of current benchmarks, yet it may ultimately have a much deeper impact on representation learning.
  • Issue on Non-scalability: We acknowledge this limitation and have added an analysis paragraph in the Discussion noting that the current memory requirement is not ideal for industrial scaling. However, this computational complexity is what allows for qualitatively distinct capabilities. Again, as shown in Section 2 and Appendix F, lateral inhibition serves as an adaptive L0 regularization that functions fundamentally differently from L1 penalty or fixed Top-K L0 penalties. It achieves markedly better results when the dataset underlying structural complexity varies greatly, imposing a distinct inductive bias on what can be learned [1], the theoretical implications of which require further investigation.

    Furthermore, this complexity allows the network to co-learn relational representations. The resulting lateral matrix could potentially serve as an internal cognitive map for potential planning, similar to hypothesized functions of the hippocampus [2]. All in all, this architecture does not merely add computational overhead; it introduces genuinely new information processing capabilities to the network.

    Lastly, since the WLIW_{LI} does not go through the backpropagation, its anti-hebbian training can be potentially moved to CPU to save GPU resources; but yes, that will be even slower. So a better engineering solution is definitely needed for industrial implementation.

I hope the impact of our work could also extend beyond conventional mechanistic interpretability community. For example, leveraging this L0 adaptivity, future work may achieve compositional and continual learning by inhibiting the right groups of neurons in real time via recurrent dynamics. It also highlights an interesting paradigm for algorithm design: allowing different learning and homeostatic rules to work synergistically to achieve complementary representational capabilities in one neural network.

[1]: Hindupur, Sai Sumedh R., et al. "Projecting assumptions: The duality between sparse autoencoders and concept geometry." Advances in Neural Information Processing Systems 38 (2026): 11649-11699.

[2]: Whittington, James CR, et al. "How to build a cognitive map." Nature neuroscience 25.10 (2022): 1257-1272.

Reviewer_znxJOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on the adaptive sparsity in SAE and proposes Lateral-Inhibition Sparse Autoencoder (LISAE), which introduces biologically inspired recurrent competition into standard autoencoders. The anti-hebbian learning rule updates inhibitory connections to reduce neuron co-activation, while homeostatic threshold regulation prevents dead neurons. Results on MNIST and Gemma-2-9B show the nterpretable feature learning and input-dependent sparsity.

Critical strengths

1), Lateral Inhibition driven Sparsity Mechanism. The most interesting part of this paper is to introduce a new sparsity rule for the SAE community, enabling competitive feature selection without a predefined activation budget.

2), daptive Sparsity. The number of active neurons dynamically adjusts to input complexity, allowing the model to represent different numbers of semantic concepts.

3), Hybrid Learning Strategy. Unlike previous SAE training, the learning process of the proposed model is complex, while the authors addressed this by combining global gradient-based optimization with local anti-Hebbian learning and homeostatic regulation, offering an interesting alternative to purely gradient-based SAE training.

Critical weaknesses

1), Quadratic Computational Complexity. The inhibitory matrix introduces O(N2)O(N^2) parameters and computational overhead, making the approach difficult to scale to large-width SAEs commonly used in LLM interpretability.

2), Potential Over-Suppression of Correlated Features. Anti-Hebbian learning penalizes neuron co-activation, but frequently co-occurring features are not necessarily redundant. Semantically related yet complementary features may therefore suppress each other. Imaging the "dog" and "grass" concepts, they often occurs but show not necessarily redundant.

3), Limited Experimental Evaluation. Experiments primarily focus on MNIST and Gemma-2 representations. The paper lacks comprehensive evaluations on standardized SAE benchmarks, such as SAEBench [1], and does not establish superiority over existing SAE variants [2].

[1] SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretability. https://arxiv.org/abs/2503.09532

[2] Improving Sparse Autoencoder with Dynamic Attention. https://arxiv.org/abs/2604.14925.

Questions for authors

Please refer to the Weakness section.

1 comment
AuthorsOpenReview ↗

Response to Reviewer znxJ

We thank Reviewer znxJ for your constructive feedback. As there were shared concerns regarding scalability and quantitative benchmark reporting across the reviews, we kindly refer the reviewer to our "General response to common concerns raised by reviewers" for an overarching discussion of our motivation. Here, we will focus directly on addressing the specific issues raised in your review.

Regarding specific issues:

  • Issue on Quadratic Computational Complexity: We acknowledge this limitation and have added a thorough analysis in the Discussion noting that the current memory requirement is not ideal for industrial scaling. However, this complexity is what allows for qualitatively different characteristics. As shown in Section 3.2 and Appendix F, lateral inhibition serves as an adaptive L0 regularization that functions fundamentally differently from L1 penalty or fixed Top-K L0 penalties. It achieves markedly better results when the structural complexity of the input varies greatly, imposing a distinct inductive bias on what can be learned [1], the theoretical implications of which require further investigation.

    Furthermore, this complexity allows the network to co-learn relational representations. The resulting lateral matrix could potentially serve as an internal cognitive map for potential planning, similar to hypothesized functions of the hippocampus [2]. All in all, this architecture does not merely add computational overhead; it introduces genuinely new information processing capabilities to the network.

  • Issue on Potential Over-Suppression of Correlated Features: No, the model does not suffer from over-suppression, and this is explicitly prevented by the "drive" term in the recurrent dynamics. This drive term constantly pushes back against the lateral inhibition, and this exact push-and-pull dynamic is where the adaptive sparsity originates.

    What can be observed in Figure 2b and Figure 9 illustrates this intuitively. For example, in the spatial MNIST task, the network certainly encounters combinations of different digits during training, and inhibitory connections naturally form between them. However, when you test the model with a composite image containing 8 distinct digits, they do not over-suppress each other down to a single surviving digit. Instead, the network adaptively scales its active capacity to represent all 8 digits based on the input's complexity. The greater the input complexity, the greater the number of latent neurons activated.

    This happens specifically because the drive term operates continuously within the recurrent loop. It acts as a constant force compelling the network to retain as much of the underlying input structure as possible. Rather than allowing only the single strongest feature to survive a strict winner-takes-all competition, the continuous drive ensures that most genuinely relevant, correlated features survive the lateral inhibition.

  • Issue on No State-of-art Benchmark Comparison: We have added an initial SAEBench comparison in the Discussion and Appendix G to provide a quantitative baseline, training a 16k-width LISAE for 250M tokens on Gemma-2-2b representations. While this evaluation shows that our current implementation does not achieve state-of-the-art performance compared to standard SAEs, we emphasize again the comparison in Appendix F: LISAE may achieve fundamentally different feature learning. Because of its adaptive L0 characteristics, LISAE excels when the dataset contains a large variation in the number of underlying features. This structural adaptivity falls outside the scope of current benchmarks, yet it may ultimately have a much deeper impact on representation learning.

Lastly, I hope the impact of our work could extend beyond mechanistic interpretability community. For example, leveraging this L0 adaptivity, future work may achieve compositional and continual learning by inhibiting the right groups of neurons in real time via recurrent dynamics. It also highlights an interesting paradigm for algorithm design: allowing different learning and homeostatic rules to work synergistically to achieve complementary representational capabilities in one neural network.

[1]: Hindupur, Sai Sumedh R., et al. "Projecting assumptions: The duality between sparse autoencoders and concept geometry." Advances in Neural Information Processing Systems 38 (2026): 11649-11699.

[2]: Whittington, James CR, et al. "How to build a cognitive map." Nature neuroscience 25.10 (2022): 1257-1272.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to common concerns raised by reviewers

We sincerely thank all the reviewers for their constructive feedback. We acknowledge that all reviewers have mentioned the unscalability of the current approach due to heavy memory requirements and the lack of comparison with state-of-the-art SAE benchmarks.

The motivation behind this paper is not to build a better tool as a replacement for current SAE tools. Instead, we were intrigued by the fact that in our brain, or even in many simpler animals, many different learning rules and homeostatic regulations coexist in one neural system. They seem to work locally, but in the end, highly intelligent behavior emerges as a whole.

Here, our main contribution is demonstrating how these coexisting learning rules work together: backpropagation minimizes the global task, anti-Hebbian learning-trained inhibitory connections create adaptive L0 regularization, and homeostatic regulation of thresholds ensure all neurons remain active. By synergistically affecting what each component encodes, they give rise to interpretable representations that reflect the underlying structure of the world—capturing not only the individual source information, but also the relational structure among the underlying features. Our work draw inspiration from biological mechanisms, exploring how such mechanisms work and interact, rather than building new tools to improve performace. While most of our current AI systems are governed by a single global loss function, maybe those small local rules are the missing components to achieve greater information processing abilities.

To address specific issues, we have added two Appendix sections (F & G) to enrich Section 3.2 and the Discussion: a follow-up experiment on the spatial MNIST dataset comparing different SAEs, and an initial quantitative evaluation using SAEBench.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.