Weight Pair Encoding reduces neural network grammar size

Researchers have developed a novel technique called Weight Pair Encoding (WeightPE) that aims to reduce the size of neural network weights by inducing a smaller grammar. This method incorporates a Re-Pair compressor within a straight-through estimator, allowing network weights to be explicitly fine-tuned for grammar compression. Experiments on Vision Transformer models fine-tuned on CIFAR-10 demonstrated that WeightPE achieved significantly smaller Re-Pair grammars compared to standard quantization methods, albeit with a slight decrease in accuracy. AI

IMPACT This research could lead to more efficient neural network models by reducing their storage footprint.

RANK_REASON The cluster contains a research paper detailing a new method for neural network weight compression. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• CIFAR-10
• LZ78
• Re-Pair
• sequitur algorithm
• ViT-B/16
• ViT-L/16
• Weight Pair Encoding

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →