re-bottleneck-latent-restructuring

【开发与工程】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.

独立 GitHub 快照 · 自动分类 · 以原作者文档为准