- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper investigates a novel form of feature learning in multilayer perceptrons (MLPs) trained on clustered data. The authors demonstrate that MLPs naturally develop monosemantic specialized neurons: individual neurons that align predominantly with a single cluster-specific predictive direction. This specialization enables MLPs to simultaneously discover cluster structure and learn local low-dimensional representations, effectively behaving like an implicit mixture-of-experts architecture without explicit routing mechanisms. The key contributions are: (1) empirical demonstration of neuron monosemanticity in trained MLPs across different activation functions; (2) theoretical analysis proving that randomly initialized neurons specialize to cluster-specific predictive directions under small initialization; (3) a provable sample-complexity advantage showing MLPs achieve polynomial-sample consistency while kernel methods and RFM fail under any fixed polynomial scaling as the number of clusters grows; (4) identification that gated activations (ReGLU, SwiGLU) further improve sample efficiency in clustered data settings.
Critical strengths
- The paper identifies a qualitatively different mechanism of feature learning: neuron specialization to locally predictive features. It goes beyond the classical view of global low-dimensional representation learning.
- The authors provide thorough experiments spanning different activation functions (ReLU, GELU, ReGLU, SwiGLU), different cluster counts, and multiple specialization metrics.
- The finding that MLPs can achieve near-local performance with implicit clustering has implications for understanding why neural networks succeed in settings where global low-dimensional structure is absent.
Critical weaknesses
- The paper focuses on synthetic Gaussian mixture data throughout. While this enables clean theoretical analysis, the paper does not clearly demonstrate its practical relevance. Real-world clustered data often have overlapping clusters, non-Gaussian distributions, and varying cluster sizes—none of these scenarios are explored.
- The paper compares against RFM as the primary baseline. However, it does not discuss or compare several recent adaptive kernel methods, mixture-of-experts models with learned routing, and deep learning approaches designed for clustered data
- Despite the title referencing "Revenge of Monosemanticity," the paper does not deeply engage with the monosemanticity literature in language models.
Questions for authors
- Can you provide experiments on real-world datasets where clustered structure naturally exists?
- How does neuron specialization behave when clusters are not well-separated or when the number of clusters is unknown?
- Can you provide intuition or theoretical analysis for why multiplicative gating improves specialization?
- Since the paper argues MLPs implicitly learn mixture-of-experts-like structures, a direct empirical comparison with explicitly trained MoE models would strengthen the paper.
- How does the monosemanticity you observe in this setting relate to the feature monosemanticity observed in language models?
2 comments
Dear Reviewer kKzL,
We are following up on our previous message. In response to your thoughtful questions, we have provided additional clarifications and included several new experiments in the appendices. If you find that these updates address your questions, we would greatly appreciate it if you consider raising your score.
Thank you for your time and thoughtful feedback.
We thank the reviewer for their feedback. We address the questions one by one as follows:
Yes, we added experiments on MNIST and Fashion-MNIST in Appendix D, where digit and clothing categories provide natural class structure. We evaluate settings with 2, 5, and 10 classes across ReLU, GELU, ReGLU, and SwiGLU. The histograms and tables show that monosemnoticity appears naturally on real data.
In our experiments, MLPs naturally develop specialized neurons without knowing the number of clusters or their identities.
For clusters that are not well separated, we added experiments in Appendix E. Table 15 shows that specialization persists even at partially overlapped clusters, although the fraction of specialized neurons decreases as overlap increases.
Multiplicative gating can separate where a feature is useful (gate branch) from what is predicted (value branch), whereas a ReLU/GeLU unit must combine both roles in one affine response. We added Appendix F, Figure 15 to illustrate this.
Our “MoE-like” interpretation provides a simple mental model of how a standard dense MLP learns from clustered data in our setting. It describes how the network can jointly identify cluster structure and learn local predictive functions, despite having no explicit router or separate expert modules. An explicit MoE introduces this router–expert decomposition architecturally, while learning the routing assignments and expert functions. Comparing it with a dense MLP primarily evaluates the potential benefits of imposing that decomposition, as explored by Kawata et al. (2025), whose work we cite and discuss in the paper.
Our work provides a simple, theoretically grounded setting in which we can understand how monosemantic neurons naturally emerge during training and how this can improve sample complexity. While semantic concepts in LLMs are more complex than the cluster-specific features we study, our results identify a fundamental capability of MLPs that may help explain similar phenomena in larger models. This is what we are currently working on, but the setting is much more complex that goes beyond the scope of this work.