Variable precision representation for efficient VQ codebook storage

In vector quantization (VQ) with fast search techniques, the storage available limits the number of codevectors used in VQ. Variable precision representation (VPR) is a simple codebook compression scheme. VPR for each vector y stores the number e(y), the number of leading bits which are zero in all elements, and avoids storing those leading bits. When storing the difference of codevectors in a binary tree structured VQ codebook, VPR can save from 24% to 44% in storage. Storing the codevector difference removes the redundancy between similar codevectors. Also as the mean square error of the VQ encoder is lowered, on the average, the difference becomes smaller and yields to better compression. To process vectors in VPR format, the operator uses a bit-serial, element-parallel scheme to evaluate the inner product. The operator's throughput can be increased by replicating its core.<
>