mlsanityData to 14 Oct 2026

Fundamental Limits of Neural Network Sparsification: Evidence from Catastrophic Interpretability Collapse

arXiv 2603.18056 ↗First posted 18 Mar 2026cs.LGcited by 0

On the map: Mechanistic interpretability → Sparse autoencoder training and evaluation

Abstract (arXiv)

Extreme neural network sparsification (90% activation reduction) presents a critical challenge for mechanistic interpretability: understanding whether interpretable features survive aggressive compression. This work investigates feature survival under severe capacity constraints in hybrid Variational Autoencoder--Sparse Autoencoder (VAE-SAE) architectures. We introduce an adaptive sparsity scheduling framework that progressively reduces active neurons from 500 to 50 over 50 training epochs, and provide empirical evidence for fundamental limits of the sparsification-interpretability relationship. Testing across two benchmark datasets -- dSprites and Shapes3D -- with both Top-k and L1 sparsification methods, our key finding reveals a pervasive paradox: while global representation quality (measured by Mutual Information Gap) remains stable, local feature interpretability collapses systematically. Under Top-k sparsification, dead neuron rates reach 34.4±0.9% on dSprites and 62.7±1.3% on Shapes3D at k=50. L1 regularization -- a fundamentally different "soft constraint" paradigm -- produces equal or worse collapse: 41.7±4.4% on dSprites and 90.6±0.5% on Shapes3D. Extended training for 100 additional epochs fails to recover dead neurons, and the collapse pattern is robust across all tested threshold definitions. Critically, the collapse scales with dataset complexity: Shapes3D (RGB, 6 factors) shows 1.8× more dead neurons than dSprites (grayscale, 5 factors) under Top-k and 2.2× under L1. These findings establish that interpretability collapse under sparsification is intrinsic to the compression process rather than an artifact of any particular algorithm, training duration, or threshold choice.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Sparse autoencoder training and evaluation, one of 13 directions in Mechanistic interpretability. Builds on 15 of its 23 references that are in the map. First posted on arXiv 18 Mar 2026.

Builds on

15 of 23 references.

Closest ancestors · 2

Same direction, newest first

  1. 2025Interpretable and Steerable Concept Bottleneck Sparse Autoencoders11 Dec
  2. 2022Toy Models of Superposition21 Sep

Other references · 13

  1. 2018The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networkscited by 2,262
  2. 2015Learning both Weights and Connections for Efficient Neural Networkscited by 2,031
  3. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 1,420
  4. 2018Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representationscited by 941
  5. 2017Channel Pruning for Accelerating Very Deep Neural Networkscited by 702
  6. 2017Learning Efficient Convolutional Networks through Network Slimmingcited by 676
  7. 2018Disentangling by Factorisingcited by 616
  8. 2017Learning Sparse Neural Networks through L0 Regularizationcited by 583
  9. 2020What is the State of Neural Network Pruning?cited by 583
  10. 2024Scaling and evaluating sparse autoencoderscited by 553
  11. 2021Sparsity in Deep Learning: Pruning and growth for efficient inference and training in neural networkscited by 527
  12. 2019Rigging the Lottery: Making All Tickets Winnerscited by 495
1 more
  1. 2019Parameter Efficient Training of Deep Convolutional Neural Networks by Dynamic Sparse Reparameterizationcited by 189

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026Sparse Transformer: a Controlled Testbed for Mechanistic Interpretability
  2. 2025Train One Sparse Autoencoder Across Multiple Sparsity Budgets to Preserve Interpretability and Accuracy
  3. 2024Interpretability as Compression: Reconsidering SAE Explanations of Neural Activations with MDL-SAEs
  4. 2026Sparse but not Simpler: A Multi-Level Interpretability Analysis of Vision Transformers
  5. 2026When Pruning Meets Interpretability: Preserving Sparse Autoencoder Robustness in LLMs
  6. 2026Improving Robustness In Sparse Autoencoders via Masked Regularization
  7. 2025Understanding Generalization, Robustness, and Interpretability in Low-Capacity Neural Networks
  8. 2026Beyond the Hard Budget: Sparsity Regularizers for More Interpretable Top-k Sparse Autoencoders