MaRN: a PyTorch library for low-dimensional parameter mapping

The author released MaRN, a PyTorch library that optimizes a compact latent representation instead of all model parameters. Benchmarks show an MNIST CNN cut from 107,998 to 1,872 trainable parameters (57.7x) at 91.80% accuracy, though mapped models train much slower; the author calls results exploratory, not evidence of general superiority.

• Developer releases manual transformer autograd in PyTorch2026-10-28
• Practical notes for GPU kernel dev setup2026-10-27
• Deepity releases C++/CUDA predictive coding library, edges JAX baseline on MNIST2026-10-27
• Raschka releases from-scratch reasoning model video course2026-10-26
• Developer proposes speculative decoding using 2-bit draft plus RAM residual2026-10-22