Reproduce “HyperVQ: MLR-based Vector Quantization in Hyperbolic Space” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Nabarun Goswami, Yusuke Mukuta, Tatsuya Harada
HyperVQ replaces nearest-vector lookup in vector quantization with classification in hyperbolic space, selecting a decision-region representative as the quantized vector. It maps encoder outputs to the Poincaré ball, uses hyperbolic multinomial logistic regression with Gumbel Softmax and a straight-through estimator, then maps the representative back to Euclidean space. On CIFAR-100 with 512 codes, accuracy was 31.59±0.91%, versus 28.28±0.88% for an embedding-matrix variant; this supports the design in that test only.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.