A cross-attention SAE with sparsemax attention achieves lower reconstruction loss and higher-quality concepts than fixed-sparsity baselines by making activation counts data-dependent.
Imagenet: A large-scale hierarchical image database
4 Pith papers cite this work. Polarity classification is still indexing.
4
Pith papers citing it
verdicts
UNVERDICTED 4representative citing papers
CylinderDepth uses cylindrical spatial attention with non-learned weights to enforce cross-view consistency in self-supervised surround depth estimation.
The paper reformulates absolute pose regression as regressing disentangled world-coordinate raymaps and pointmaps from images, then recovering pose via a differentiable solver, claiming SOTA results on 7-Scenes and Cambridge Landmarks.
DLC inserts lightweight classifier-proximal plugins into distillation-based continual learning to achieve 8% accuracy gains on large benchmarks with only 4% extra backbone parameters.
citing papers explorer
-
Improving Sparse Autoencoder with Dynamic Attention
A cross-attention SAE with sparsemax attention achieves lower reconstruction loss and higher-quality concepts than fixed-sparsity baselines by making activation counts data-dependent.
-
CylinderDepth: Cylindrical Spatial Attention for Multi-View Consistent Self-Supervised Surround Depth Estimation
CylinderDepth uses cylindrical spatial attention with non-learned weights to enforce cross-view consistency in self-supervised surround depth estimation.
-
GRLoc: Geometric Representation Regression for Visual Localization
The paper reformulates absolute pose regression as regressing disentangled world-coordinate raymaps and pointmaps from images, then recovering pose via a differentiable solver, claiming SOTA results on 7-Scenes and Cambridge Landmarks.
-
Pushing the Limits of Distillation-Based Continual Learning via Classifier-Proximal Lightweight Plugins
DLC inserts lightweight classifier-proximal plugins into distillation-based continual learning to achieve 8% accuracy gains on large benchmarks with only 4% extra backbone parameters.