The Geometry of Concepts: Sparse Autoencoder Feature Structure

• URL: http://arxiv.org/abs/2410.19750v2
• Date: Sun, 30 Mar 2025 23:55:03 GMT
• Title: The Geometry of Concepts: Sparse Autoencoder Feature Structure
• Authors: Yuxiao Li, Eric J. Michaud, David D. Baek, Joshua Engels, Xiaoqing Sun, Max Tegmark,
• Abstract summary: We find that the concept universe has interesting structure at three levels.<n>The "brain" intermediate-scale structure has significant spatial modularity.<n>The "galaxy" scale large-scale structure of the feature point cloud is not isotropic, but instead has a power law of eigenvalues with steepest slope in middle layers.
• Score: 10.95343312207608
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Sparse autoencoders have recently produced dictionaries of high-dimensional vectors corresponding to the universe of concepts represented by large language models. We find that this concept universe has interesting structure at three levels: 1) The "atomic" small-scale structure contains "crystals" whose faces are parallelograms or trapezoids, generalizing well-known examples such as (man-woman-king-queen). We find that the quality of such parallelograms and associated function vectors improves greatly when projecting out global distractor directions such as word length, which is efficiently done with linear discriminant analysis. 2) The "brain" intermediate-scale structure has significant spatial modularity; for example, math and code features form a "lobe" akin to functional lobes seen in neural fMRI images. We quantify the spatial locality of these lobes with multiple metrics and find that clusters of co-occurring features, at coarse enough scale, also cluster together spatially far more than one would expect if feature geometry were random. 3) The "galaxy" scale large-scale structure of the feature point cloud is not isotropic, but instead has a power law of eigenvalues with steepest slope in middle layers. We also quantify how the clustering entropy depends on the layer.

Related papers

• How high is `high'? Rethinking the roles of dimensionality in topological data analysis and manifold learning [8.397730500554047]
  We present a generalised Hanson-Wright inequality and use it to establish new statistical insights into the geometry of data point-clouds.<n>We revisit the ground-breaking neuroscience discovery of isometric toroidal structure in grid-cell activity made by Gardner et al.<n>Our findings reveal, for the first time, evidence that this structure is in fact to physical space, meaning that grid cell activity conveys a geometrically faithful representation of the real world.
  arXiv  Detail & Related papers  (2025-05-22T16:34:15Z)
• The Geometry of Meaning: Perfect Spacetime Representations of Hierarchical Structures [0.0]
  We show that there is a fast algorithm that embeds hierarchical structures in three-dimensional Minkowski spacetime.<n>Our results seem to indicate that all discrete data has a perfect geometrical representation that is three-dimensional.
  arXiv  Detail & Related papers  (2025-05-07T20:41:06Z)
• Understanding and Mitigating Hyperbolic Dimensional Collapse in Graph Contrastive Learning [70.0681902472251]
  We propose a novel contrastive learning framework to learn high-quality graph embeddings in hyperbolic space. Specifically, we design the alignment metric that effectively captures the hierarchical data-invariant information. We show that in the hyperbolic space one has to address the leaf- and height-level uniformity related to properties of trees.
  arXiv  Detail & Related papers  (2023-10-27T15:31:42Z)
• Bayes Complexity of Learners vs Overfitting [4.873362301533825]
  We show that a new notion of complexity of functions governs a PAC Bayes-like generalization bound. In contrast to previous works, our notion naturally generalizes to neural networks with several layers. An upper-bound we derive allows to show a separation in the number of samples needed for good generalization between 2 and 4-layer neural networks.
  arXiv  Detail & Related papers  (2023-03-13T13:07:02Z)
• Geometry Interaction Knowledge Graph Embeddings [153.69745042757066]
  We propose Geometry Interaction knowledge graph Embeddings (GIE), which learns spatial structures interactively between the Euclidean, hyperbolic and hyperspherical spaces. Our proposed GIE can capture a richer set of relational information, model key inference patterns, and enable expressive semantic matching across entities.
  arXiv  Detail & Related papers  (2022-06-24T08:33:43Z)
• A Scalable Combinatorial Solver for Elastic Geometrically Consistent 3D Shape Matching [69.14632473279651]
  We present a scalable algorithm for globally optimizing over the space of geometrically consistent mappings between 3D shapes. We propose a novel primal coupled with a Lagrange dual problem that is several orders of magnitudes faster than previous solvers.
  arXiv  Detail & Related papers  (2022-04-27T09:47:47Z)
• Neural Convolutional Surfaces [59.172308741945336]
  This work is concerned with a representation of shapes that disentangles fine, local and possibly repeating geometry, from global, coarse structures. We show that this approach achieves better neural shape compression than the state of the art, as well as enabling manipulation and transfer of shape details.
  arXiv  Detail & Related papers  (2022-04-05T15:40:11Z)
• Orientation-Aware Graph Neural Networks for Protein Structure Representation Learning [29.366321002562373]
  We propose the Orientation-Aware Graph Neural Networks (OAGNNs) to better sense the geometric characteristics in protein structure. Extending a single weight from a scalar to a 3D vector, we construct a rich set of geometric-meaningful operations. OAGNNs have a remarkable ability to sense geometric orientational features compared to classical networks.
  arXiv  Detail & Related papers  (2022-01-28T13:41:56Z)
• Highly Scalable and Provably Accurate Classification in Poincare Balls [40.82908295137667]
  We establish a unified framework for learning scalable and simple hyperbolic linear classifiers with provable performance guarantees. Our results include a new hyperbolic and second-order perceptron algorithm as well as an efficient and highly accurate convex optimization setup for hyperbolic support vector machine classifiers. Their performance accuracies on synthetic data sets comprising millions of points, as well as on complex real-world data sets such as single-cell RNA-seq expression measurements, CIFAR10, Fashion-MNIST and mini-ImageNet.
  arXiv  Detail & Related papers  (2021-09-08T16:59:39Z)
• Vector Neurons: A General Framework for SO(3)-Equivariant Networks [32.81671803104126]
  In this paper, we introduce a general framework built on top of what we call Vector Neuron representations. Our vector neurons enable a simple mapping of SO(3) actions to latent spaces. We also show for the first time a rotation equivariant reconstruction network.
  arXiv  Detail & Related papers  (2021-04-25T18:48:15Z)
• Learning from Protein Structure with Geometric Vector Perceptrons [6.5360079597553025]
  We introduce geometric vector perceptrons, which extend standard dense layers to operate on collections of Euclidean vectors. We demonstrate our approach on two important problems in learning from protein structure: model quality assessment and computational protein design.
  arXiv  Detail & Related papers  (2020-09-03T01:54:25Z)
• DSG-Net: Learning Disentangled Structure and Geometry for 3D Shape Generation [98.96086261213578]
  We introduce DSG-Net, a deep neural network that learns a disentangled structured and geometric mesh representation for 3D shapes. This supports a range of novel shape generation applications with disentangled control, such as of structure (geometry) while keeping geometry (structure) unchanged. Our method not only supports controllable generation applications but also produces high-quality synthesized shapes, outperforming state-of-the-art methods.
  arXiv  Detail & Related papers  (2020-08-12T17:06:51Z)
• Dense Non-Rigid Structure from Motion: A Manifold Viewpoint [162.88686222340962]
  Non-Rigid Structure-from-Motion (NRSfM) problem aims to recover 3D geometry of a deforming object from its 2D feature correspondences across multiple frames. We show that our approach significantly improves accuracy, scalability, and robustness against noise.
  arXiv  Detail & Related papers  (2020-06-15T09:15:54Z)
• Convolutional Occupancy Networks [88.48287716452002]
  We propose Convolutional Occupancy Networks, a more flexible implicit representation for detailed reconstruction of objects and 3D scenes. By combining convolutional encoders with implicit occupancy decoders, our model incorporates inductive biases, enabling structured reasoning in 3D space. We empirically find that our method enables the fine-grained implicit 3D reconstruction of single objects, scales to large indoor scenes, and generalizes well from synthetic to real data.
  arXiv  Detail & Related papers  (2020-03-10T10:17:07Z)

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.