re-bottleneck-latent-restructuring

Restructure latent representations in pretrained audio autoencoders without full retraining. Apply three variants—ordered, semantic, and equivariant—to enforce structure like channel ordering, semantic alignment, or filter correspondence. Achieves 20-60% semantic gains in under 48 GPU hours versus 14.5K hours for full retraining.

Independent GitHub snapshot · Automated classification · Original documentation governs