Improving Sparse Autoencoder with Dynamic Attention

• URL: http://arxiv.org/abs/2604.14925v1
• Date: Thu, 16 Apr 2026 12:10:08 GMT
• Title: Improving Sparse Autoencoder with Dynamic Attention
• Abstract summary: We propose a new class of SAEs based on the cross-attention architecture with the latent features as queries and the learnable dictionary as the key and value matrices.<n>To encourage sparse pattern learning, we employ a sparsemax-based attention strategy that automatically infers a sparse set of elements according to the complexity of each neuron.
• Score: 8.50973639447688
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Recently, sparse autoencoders (SAEs) have emerged as a promising technique for interpreting activations in foundation models by disentangling features into a sparse set of concepts. However, identifying the optimal level of sparsity for each neuron remains challenging in practice: excessive sparsity can lead to poor reconstruction, whereas insufficient sparsity may harm interpretability. While existing activation functions such as ReLU and TopK provide certain sparsity guarantees, they typically require additional sparsity regularization or cherry-picked hyperparameters. We show in this paper that dynamically sparse attention mechanisms using sparsemax can bridge this trade-off, due to their ability to determine the activation numbers in a data-dependent manner. Specifically, we first explore a new class of SAEs based on the cross-attention architecture with the latent features as queries and the learnable dictionary as the key and value matrices. To encourage sparse pattern learning, we employ a sparsemax-based attention strategy that automatically infers a sparse set of elements according to the complexity of each neuron, resulting in a more flexible and general activation function. Through comprehensive evaluation and visualization, we show that our approach successfully achieves lower reconstruction loss while producing high-quality concepts, particularly in top-n classification tasks.

Related papers

• CLAIRE: Compressed Latent Autoencoder for Industrial Representation and Evaluation -- A Deep Learning Framework for Smart Manufacturing [51.56484100374058]
  CLAIRE is a hybrid end-to-end learning framework that integrates unsupervised deep representation learning with supervised classification for intelligent quality control in smart manufacturing systems.<n>It employs an optimized deep autoencoder to transform raw input into a compact latent space, effectively capturing the intrinsic data structure while suppressing irrelevant or noisy features.<n>The proposed framework highlights the potential of integrating explainable AI with feature-aware regularization for robust fault detection.
  arXiv  Detail & Related papers  (2026-03-06T15:11:58Z)
• Unlocking Prototype Potential: An Efficient Tuning Framework for Few-Shot Class-Incremental Learning [69.28860905525057]
  Few-shot class-incremental learning (FSCIL) seeks to continuously learn new classes from very limited samples.<n>We introduce an efficient prototype fine-tuning framework that evolves static centroids into dynamic, learnable components.
  arXiv  Detail & Related papers  (2026-02-05T03:50:53Z)
• Constrained Auto-Regressive Decoding Constrains Generative Retrieval [71.71161220261655]
  Generative retrieval seeks to replace traditional search index data structures with a single large-scale neural network.<n>In this paper, we examine the inherent limitations of constrained auto-regressive generation from two essential perspectives: constraints and beam search.
  arXiv  Detail & Related papers  (2025-04-14T06:54:49Z)
• Features that Make a Difference: Leveraging Gradients for Improved Dictionary Learning [4.051777802443125]
  Sparse Autoencoders (SAEs) are a promising approach for extracting neural network representations.<n>We introduce Gradient SAEs, which modify the $k$-sparse autoencoder architecture by augmenting the TopK activation function.<n>We find evidence that g-SAEs learn latents that are on average more effective at steering models in arbitrary contexts.
  arXiv  Detail & Related papers  (2024-11-15T18:03:52Z)
• Anchor Attention, Small Cache: Code Generation with Large Language Models [15.94784908771546]
  Current practices in NLP often use sparse attention which may, unfortunately, lead to substantial inaccuracies, or hallucinations, in code generation tasks. We propose a novel approach, AnchorCoder, which features token-wise anchor attention designed to extract and compress contextual information. It can consistently achieve a significant (at least 70%) reduction in KV cache requirements, while preserving the majority of model's performance.
  arXiv  Detail & Related papers  (2024-11-11T02:47:05Z)
• Enhancing Fine-Grained Visual Recognition in the Low-Data Regime Through Feature Magnitude Regularization [23.78498670529746]
  We introduce a regularization technique to ensure that the magnitudes of the extracted features are evenly distributed. Despite its apparent simplicity, our approach has demonstrated significant performance improvements across various fine-grained visual recognition datasets.
  arXiv  Detail & Related papers  (2024-09-03T07:32:46Z)
• LoRA-Ensemble: Efficient Uncertainty Modelling for Self-Attention Networks [52.46420522934253]
  We introduce LoRA-Ensemble, a parameter-efficient ensembling method for self-attention networks.<n>The method not only outperforms state-of-the-art implicit techniques like BatchEnsemble, but even matches or exceeds the accuracy of an Explicit Ensemble.
  arXiv  Detail & Related papers  (2024-05-23T11:10:32Z)
• Self-STORM: Deep Unrolled Self-Supervised Learning for Super-Resolution Microscopy [55.2480439325792]
  We introduce deep unrolled self-supervised learning, which alleviates the need for such data by training a sequence-specific, model-based autoencoder. Our proposed method exceeds the performance of its supervised counterparts.
  arXiv  Detail & Related papers  (2024-03-25T17:40:32Z)
• Semantics-Aware Dynamic Localization and Refinement for Referring Image Segmentation [102.25240608024063]
  Referring image segments an image from a language expression. We develop an algorithm that shifts from being localization-centric to segmentation-language. Compared to its counterparts, our method is more versatile yet effective.
  arXiv  Detail & Related papers  (2023-03-11T08:42:40Z)
• The Impact of Activation Sparsity on Overfitting in Convolutional Neural Networks [1.9424280683610138]
  Overfitting is one of the fundamental challenges when training convolutional neural networks. In this study we introduce a perplexity-based sparsity definition to derive and visualise layer-wise activation measures.
  arXiv  Detail & Related papers  (2021-04-13T12:55:37Z)
• Phase Retrieval using Expectation Consistent Signal Recovery Algorithm based on Hypernetwork [73.94896986868146]
  Phase retrieval is an important component in modern computational imaging systems. Recent advances in deep learning have opened up a new possibility for robust and fast PR. We develop a novel framework for deep unfolding to overcome the existing limitations.
  arXiv  Detail & Related papers  (2021-01-12T08:36:23Z)
• Exploiting the Full Capacity of Deep Neural Networks while Avoiding Overfitting by Targeted Sparsity Regularization [1.3764085113103217]
  Overfitting is one of the most common problems when training deep neural networks on comparatively small datasets. We propose novel targeted sparsity visualization and regularization strategies to counteract overfitting.
  arXiv  Detail & Related papers  (2020-02-21T11:38:17Z)

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.