MaRN: PyTorch library cuts MNIST CNN params 57.7x via low-dim parameter mappings
Less_Dream_6331 · reddit · 2026-10-09
The author released MaRN (Mapping Networks), a PyTorch library that optimizes a compact latent representation instead of directly training every model parameter, restoring parameters through low-dimensional mappings.
Early benchmark results:
- MNIST CNN: trainable params cut from 107,998 to 1,872 (57.7x reduction) at 91.80% accuracy
- LSTM forecasting: 12,051 → 2,048 params with validation MSE of 0.00006
- CNN2 + pruning: only 204 trainable params at 81.25% accuracy
The author is upfront about trade-offs: mapped models can train substantially slower, performance varies by task, and some benchmarks use synthetic data — no evidence of general superiority over direct training. The library includes global and layer-wise mappings, regularization options, and pruning/LRD integrations. Code and docs are open; feedback welcome on the approach, benchmark design, and use cases.
More from Research
- Apple Paper: SOTA Harnesses Offer No Edge Over Minimal Single-Session Coding Agent — himanshustwts · 2026-10-09
- LangChain Founder: Evals Work for Narrow Tasks but Break Down for Autonomous Agents — hwchase17 · 2026-10-09
- Hugging Face launches Robotic Episodes Viewer for 24k+ LeRobot datasets — mishig25 · 2026-10-09
- Blind humanoid walks, plays soccer and lifts suitcases with joint encoders only — accepted at Humanoids 2026 — Jan_R_Peters · 2026-10-09
- Delete object info from observations and PPO learns to search anyway — TU Darmstadt on its Humanoids 2026 paper — Jan_R_Peters · 2026-10-09
- U-Space finds an interpretable subspace for LLM uncertainty, no training needed — Tobias Braun · 2026-10-09