From Data Statistics to Feature Geometry: How Correlations Shape Superposition

• URL: http://arxiv.org/abs/2603.09972v1
• Date: Tue, 10 Mar 2026 17:59:02 GMT
• Title: From Data Statistics to Feature Geometry: How Correlations Shape Superposition
• Abstract summary: A central idea in mechanistic interpretability is that neural networks represent more features than they have dimensions.<n>Superposition has mostly been studied in idealized settings where features are sparse and uncorrelated.<n>We show that superposition is typically understood as introducing interference that must be minimized geometrically and filtered out by non-linearities.
• Score: 23.393264014989615
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: A central idea in mechanistic interpretability is that neural networks represent more features than they have dimensions, arranging them in superposition to form an over-complete basis. This framing has been influential, motivating dictionary learning approaches such as sparse autoencoders. However, superposition has mostly been studied in idealized settings where features are sparse and uncorrelated. In these settings, superposition is typically understood as introducing interference that must be minimized geometrically and filtered out by non-linearities such as ReLUs, yielding local structures like regular polytopes. We show that this account is incomplete for realistic data by introducing Bag-of-Words Superposition (BOWS), a controlled setting to encode binary bag-of-words representations of internet text in superposition. Using BOWS, we find that when features are correlated, interference can be constructive rather than just noise to be filtered out. This is achieved by arranging features according to their co-activation patterns, making interference between active features constructive, while still using ReLUs to avoid false positives. We show that this kind of arrangement is more prevalent in models trained with weight decay and naturally gives rise to semantic clusters and cyclical structures which have been observed in real language models yet were not explained by the standard picture of superposition. Code for this paper can be found at https://github.com/LucasPrietoAl/correlations-feature-geometry.

Related papers

• Neural Point-Forms [1.0262304700896199]
  We introduce a new family of principled learnable geometric features for point clouds called neural point-forms (NPFs)<n>NPFs provide a competitive, and interpretable representation, with the strongest benefits appearing when labels depend on sampling density, manifold-like structure, or response-relevant population geometry.
  arXiv  Detail & Related papers  (2026-05-15T01:44:31Z)
• Shape-of-You: Fused Gromov-Wasserstein Optimal Transport for Semantic Correspondence in-the-Wild [7.4746628878102195]
  Shape-of-You is an unsupervised learning framework for semantic correspondence.<n>It achieves state-of-the-art performance on SPair-71k and AP-10k datasets.
  arXiv  Detail & Related papers  (2026-03-12T07:22:03Z)
• Sparse Semantic Dimension as a Generalization Certificate for LLMs [53.681678236115836]
  We introduce the Sparse Semantic Dimension (SSD), a complexity measure derived from the active feature vocabulary of a Sparse Autoencoder (SAE) trained on the model's layers.<n>We validate this framework on GPT-2 Small and Gemma-2B, demonstrating that our bound provides non-vacuous certificates at realistic sample sizes.
  arXiv  Detail & Related papers  (2026-02-11T21:45:18Z)
• Generative Adversarial Gumbel MCTS for Abstract Visual Composition Generation [29.755551944026738]
  We study abstract visual composition in which identity is determined by the configuration and relations among a small set of geometric primitives.<n>An AlphaGo-style search enforces feasibility, while a fine-tuned vision-language model scores semantic alignment as reward signals.<n>Inspired by the Generative Adversarial Network, we use the generated instances for adversarial reward refinement.
  arXiv  Detail & Related papers  (2025-12-01T03:38:44Z)
• Clustering Inductive Biases with Unrolled Networks [4.47196217712431]
  We propose an autoencoder architecture (WLSC) whose latent representations are implicitly, locally organized for spectral clustering through a Laplacian quadratic form of a bipartite graph. We show that our regularization can be interpreted as early-stage specialization of receptive fields to certain classes of stimuli.
  arXiv  Detail & Related papers  (2023-11-30T02:02:30Z)
• Disentanglement via Latent Quantization [60.37109712033694]
  In this work, we construct an inductive bias towards encoding to and decoding from an organized latent space. We demonstrate the broad applicability of this approach by adding it to both basic data-re (vanilla autoencoder) and latent-reconstructing (InfoGAN) generative models.
  arXiv  Detail & Related papers  (2023-05-28T06:30:29Z)
• Sparse, Geometric Autoencoder Models of V1 [2.491226380993217]
  We propose an autoencoder architecture whose latent representations are implicitly, locally organized for spectral clustering. We show that the autoencoder objective function maintains core ideas of the sparse coding framework, yet also offers a promising path to describe the differentiation of receptive fields.
  arXiv  Detail & Related papers  (2023-02-22T06:07:20Z)
• Equivariance with Learned Canonicalization Functions [77.32483958400282]
  We show that learning a small neural network to perform canonicalization is better than using predefineds. Our experiments show that learning the canonicalization function is competitive with existing techniques for learning equivariant functions across many tasks.
  arXiv  Detail & Related papers  (2022-11-11T21:58:15Z)
• Robust Change Detection Based on Neural Descriptor Fields [53.111397800478294]
  We develop an object-level online change detection approach that is robust to partially overlapping observations and noisy localization results. By associating objects via shape code similarity and comparing local object-neighbor spatial layout, our proposed approach demonstrates robustness to low observation overlap and localization noises.
  arXiv  Detail & Related papers  (2022-08-01T17:45:36Z)
• Deep Parametric Continuous Convolutional Neural Networks [92.87547731907176]
  Parametric Continuous Convolution is a new learnable operator that operates over non-grid structured data. Our experiments show significant improvement over the state-of-the-art in point cloud segmentation of indoor and outdoor scenes.
  arXiv  Detail & Related papers  (2021-01-17T18:28:23Z)
• A Boundary Regression Model for Nested Named Entity Recognition [17.968819067122418]
  Recognizing named entities (NEs) is commonly conducted as a classification problem that predicts a class tag for an NE candidate in a sentence. Recent developments in neural networks have adopted deep structures that map categorized features into continuous representations. In this paper, the regression operation is introduced to locate NEs in a sentence.
  arXiv  Detail & Related papers  (2020-11-29T10:04:38Z)
• Anchor & Transform: Learning Sparse Embeddings for Large Vocabularies [60.285091454321055]
  We design a simple and efficient embedding algorithm that learns a small set of anchor embeddings and a sparse transformation matrix. On text classification, language modeling, and movie recommendation benchmarks, we show that ANT is particularly suitable for large vocabulary sizes.
  arXiv  Detail & Related papers  (2020-03-18T13:07:51Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.