Vector Quantization
A quantization technique that maps a group of weight vectors collectively to the nearest codeword in a pre-learned codebook, achieving compression by storing codeword indices rather than original numerical values. It serves as an important technical precursor to sub-1-bit compression methods.
Timeline (last 90 days)
Vector quantization (VQ) maps a group of weights as a whole to the nearest codeword in a pre-learned codebook, storing only the codeword index.
A codebook containing 256 codewords requires only 8 bits for addressing; if each vector group contains 16 weights, then on average each weight consumes only 0.5 bits to store the index.
All Facts (2)
Vector quantization (VQ) maps a group of weights as a whole to the nearest codeword in a pre-learned codebook, storing only the codeword index.
50%UnverifiedA codebook containing 256 codewords requires only 8 bits for addressing; if each vector group contains 16 weights, then on average each weight consumes only 0.5 bits to store the index.
50%