Differentiable product quantization compresses embedding layers by replacing them with learned codebooks, achieving 14-238x compression with negligible performance loss on language tasks.
All models were trained with a batch size of 2048 sentences for 250k steps, and with the SM3 optimizer (Anil et al.,
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2019 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Differentiable Product Quantization for End-to-End Embedding Compression
Differentiable product quantization compresses embedding layers by replacing them with learned codebooks, achieving 14-238x compression with negligible performance loss on language tasks.