Get Started
Home
Topics
Search
Library
Research questionHow can we select essential hidden neurons from activations alone without labels or gradients while preserving discrimination?Overparameterized networks may contain redundant hidden units, but their essential computational content is difficult to identify without supervision or gradient information. Removing units can also eliminate distinctions needed for the task.
AI
Machine Learning
Mechanistic Interpretability
Neural and Evolutionary Computing
Statistical Machine Learning
Latest papersRecent research connected to this question, newest first.Coarse-Graining Hidden Representations: Unsupervised Neuron Selection via Mapping EntropyThe evidence concerns mapping-entropy-based selection in teacher–student networks, a nonlinear Gaussian-process task, and translation-augmented MNIST. Reported comparisons use random subsets of equal size, with the clearest differences under strong compression.research paper · Sep 4, 2026
Related questions
How does gradient training build hierarchical representations beyond the lazy kernel regime?Can pre-generalization interventions reveal when training constrains which equally fitting neural-network solutions will later generalize?How can we reduce fine-tuning-induced shortcut reliance on underrepresented groups without retraining or group labels?How can implicit neural representations capture high-frequency detail without memorizing noise or requiring fragile tuning?