Fast Trainable Multilinear Bases for Image Compression
arXiv:2608.00053
2026
Architecture
2 ideas extracted · analyzed Aug 31, 2026
What the math gives to ML
This paper provides a concrete way to learn dataset-adaptive orthogonal or unitary transforms without sacrificing exact invertibility, norm preservation, or near-linear application cost. The transferable asset is the parameterization: a fixed Fourier-like tensor-network topology whose small gates are optimized on unitary matrix manifolds. This can become a drop-in learned spectral mixer for image models, token mixers, or compression-aware neural pipelines, with guaranteed absence of activation-norm amplification and a cheap exact inverse. The most direct experiment is to replace fixed DCT/Fourier transforms in an image codec or neural image tokenizer with the trainable isometric network and measure rate-distortion, transform cost, and stability.
Ideas from this paper
△ Mechanism confirmed, baseline not beaten
2026
Replace a fixed DCT or Fourier transform in transform coding with a trainable isometric tensor-network transform whose local gates are learned once on a dataset. Retain the k coefficients with largest magnitude and reconstruct with the exact adjoint transform; the transform remains norm-preserving and fast while adapting its coefficient ordering to the data distribution.
Useful8/10
Difficulty5/10
Novelty7/10
△ Mechanism confirmed, baseline not beaten
2026
Use the relaxed QFT tensor-network topology as a trainable norm-preserving mixer inside a neural block, replacing a dense token-mixing matrix or an expensive global convolution. The network learns data-adapted global interactions while retaining structured O(N log^2 N) application and an exact cheap inverse, making it suitable for image tokens, long sequences, or reversible residual blocks.
Useful7/10
Difficulty6/10
Novelty6/10