Description
Hello. Thank you for the excellent Python module. This issue potentially relates to #153.
The paper "Factorizers for Distributed Sparse Block Codes" describes the Generalized Sparse Block Code (GSBC). Instead of enforcing a single "on" bit per block, as in Binary Sparse Block Codes, GSBC segments are normalized such that they to sum to 1. See Fig 1 for an illustration.
The paper's proposed binding and unbinding operation is block-local circular convolution.
The paper's proposed bundling operation is element-wise summation followed by block-local normalization.
I don't believe a permutation operation is utilized.
The paper's proposed similarity measure is not cosine similarity nor hamming distance. I do not fully comprehend it or else I would provide a more insightful description.