A system-level framework that makes dynamic activation quantization practical for on-device DNN training, cutting activation memory up to 22.9x and training time up to 3.2x with under 1% accuracy loss.
Robust real-time multi-vehicle collaboration on asynchronous sensors
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.NI 1years
2025 1verdicts
CONDITIONAL 1roles
background 1polarities
unclear 1representative citing papers
citing papers explorer
-
DAF: An Efficient End-to-End Dynamic Activation Framework for on-Device DNN Training
A system-level framework that makes dynamic activation quantization practical for on-device DNN training, cutting activation memory up to 22.9x and training time up to 3.2x with under 1% accuracy loss.