Sign Lock-In: Randomly Initialized Weight Signs Persist and Bottleneck Sub-Bit Model Compression
Researchers have identified a phenomenon called "sign lock-in" in neural networks, where the initial random signs of weights tend to persist throughout training. This persistence acts as a bottleneck for sub-bit model compression, limiting storage efficiency to below one bit per weight. The study formalizes this behavior with a stopping-time analysis and proposes a new training method using low-rank sign templates to overcome this limitation. AI
IMPACT Identifies a fundamental limitation in model compression and proposes a method to improve efficiency, potentially impacting deployment of large models on resource-constrained devices.