Observable Propagation: Uncovering Feature Vectors in Transformers

• URL: http://arxiv.org/abs/2312.16291v2
• Date: Tue, 4 Jun 2024 01:26:01 GMT
• Title: Observable Propagation: Uncovering Feature Vectors in Transformers
• Authors: Jacob Dunefsky, Arman Cohan,
• Abstract summary: We introduce "observable propagation" (in short: ObProp) for finding linear features used by transformer language models in computing a given task. Our paradigm centers on the concept of "observables", linear functionals corresponding to given tasks. We use ObProp to perform extensive qualitative investigations into several tasks, including gendered occupational bias, political party prediction, and programming language detection.
• Score: 21.2431937128876
• License: http://creativecommons.org/licenses/by-sa/4.0/
• Abstract: A key goal of current mechanistic interpretability research in NLP is to find linear features (also called "feature vectors") for transformers: directions in activation space corresponding to concepts that are used by a given model in its computation. Present state-of-the-art methods for finding linear features require large amounts of labelled data -- both laborious to acquire and computationally expensive to utilize. In this work, we introduce a novel method, called "observable propagation" (in short: ObProp), for finding linear features used by transformer language models in computing a given task -- using almost no data. Our paradigm centers on the concept of "observables", linear functionals corresponding to given tasks. We then introduce a mathematical theory for the analysis of feature vectors, including a similarity metric between feature vectors called the coupling coefficient which estimates the degree to which one feature's output correlates with another's. We use ObProp to perform extensive qualitative investigations into several tasks, including gendered occupational bias, political party prediction, and programming language detection. Our results suggest that ObProp surpasses traditional approaches for finding feature vectors in the low-data regime, and that ObProp can be used to better understand the mechanisms responsible for bias in large language models.

Related papers

• Knowledge Composition using Task Vectors with Learned Anisotropic Scaling [51.4661186662329]
  We introduce aTLAS, an algorithm that linearly combines parameter blocks with different learned coefficients, resulting in anisotropic scaling at the task vector level. We show that such linear combinations explicitly exploit the low intrinsicity of pre-trained models, with only a few coefficients being the learnable parameters. We demonstrate the effectiveness of our method in task arithmetic, few-shot recognition and test-time adaptation, with supervised or unsupervised objectives.
  arXiv  Detail & Related papers  (2024-07-03T07:54:08Z)
• Learning Active Subspaces and Discovering Important Features with Gaussian Radial Basis Functions Neural Networks [0.0]
  We show that precious information is contained in the spectrum of the precision matrix that can be extracted once the training of the model is completed. We conducted numerical experiments for regression, classification, and feature selection tasks. Our results demonstrate that the proposed model does not only yield an attractive prediction performance compared to the competitors.
  arXiv  Detail & Related papers  (2023-07-11T09:54:30Z)
• Generalizing Backpropagation for Gradient-Based Interpretability [103.2998254573497]
  We show that the gradient of a model is a special case of a more general formulation using semirings. This observation allows us to generalize the backpropagation algorithm to efficiently compute other interpretable statistics.
  arXiv  Detail & Related papers  (2023-07-06T15:19:53Z)
• Scalable variable selection for two-view learning tasks with projection operators [0.0]
  We propose a novel variable selection method for two-view settings, or for vector-valued supervised learning problems. Our framework is able to handle extremely large scale selection tasks, where number of data samples could be even millions.
  arXiv  Detail & Related papers  (2023-07-04T08:22:05Z)
• Inexact iterative numerical linear algebra for neural network-based spectral estimation and rare-event prediction [0.0]
  Leading eigenfunctions of the transition operator are useful for visualization. We develop inexact iterative linear algebra methods for computing these eigenfunctions.
  arXiv  Detail & Related papers  (2023-03-22T13:07:03Z)
• EqMotion: Equivariant Multi-agent Motion Prediction with Invariant Interaction Reasoning [83.11657818251447]
  We propose EqMotion, an efficient equivariant motion prediction model with invariant interaction reasoning. We conduct experiments for the proposed model on four distinct scenarios: particle dynamics, molecule dynamics, human skeleton motion prediction and pedestrian trajectory prediction. Our method achieves state-of-the-art prediction performances on all the four tasks, improving by 24.0/30.1/8.6/9.2%.
  arXiv  Detail & Related papers  (2023-03-20T05:23:46Z)
• 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)
• Functional Nonlinear Learning [0.0]
  We propose a functional nonlinear learning (FunNoL) method to represent multivariate functional data in a lower-dimensional feature space. We show that FunNoL provides satisfactory curve classification and reconstruction regardless of data sparsity.
  arXiv  Detail & Related papers  (2022-06-22T23:47:45Z)
• Path Integrals for the Attribution of Model Uncertainties [0.18899300124593643]
  We present a novel algorithm that relies on in-distribution curves connecting a feature vector to some counterfactual counterpart. We validate our approach on benchmark image data sets with varying resolution, and show that it significantly simplifies interpretability.
  arXiv  Detail & Related papers  (2021-07-19T11:07:34Z)
• A Differential Geometry Perspective on Orthogonal Recurrent Models [56.09491978954866]
  We employ tools and insights from differential geometry to offer a novel perspective on orthogonal RNNs. We show that orthogonal RNNs may be viewed as optimizing in the space of divergence-free vector fields. Motivated by this observation, we study a new recurrent model, which spans the entire space of vector fields.
  arXiv  Detail & Related papers  (2021-02-18T19:39:22Z)
• Supervised Quantile Normalization for Low-rank Matrix Approximation [50.445371939523305]
  We learn the parameters of quantile normalization operators that can operate row-wise on the values of $X$ and/or of its factorization $UV$ to improve the quality of the low-rank representation of $X$ itself. We demonstrate the applicability of these techniques on synthetic and genomics datasets.
  arXiv  Detail & Related papers  (2020-02-08T21:06:02Z)

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.