Low-resolution weight networks (1 to 4.08 bits) can roughly match 32-bit accuracy on CIFAR-10, with 2.32-bit weights identified as the best memory-performance trade-off.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Low-Resolution Neural Networks
Low-resolution weight networks (1 to 4.08 bits) can roughly match 32-bit accuracy on CIFAR-10, with 2.32-bit weights identified as the best memory-performance trade-off.