Current browse context:
cs.LG
Change to browse by:
References & Citations
Computer Science > Machine Learning
Title: A binary-activation, multi-level weight RNN and training algorithm for processing-in-memory inference with eNVM
(Submitted on 30 Nov 2019 (this version), latest version 12 Oct 2020 (v3))
Abstract: We present a new algorithm for training neural networks with binary activations and multi-level weights, which enables efficient processing-in-memory circuits with eNVM. Binary activations obviate costly DACs and ADCs. Multi-level weights leverage multi-level eNVM cells. Compared with previous quantization algorithms, our method not only works for feed-forward networks including fully-connected and convolutional, but also achieves higher accuracy and noise resilience for recurrent networks. In particular, we present a RNN trigger-word detection PIM accelerator, whose modeling results demonstrate high performance using our new training algorithm.
Submission history
From: Siming Ma [view email][v1] Sat, 30 Nov 2019 01:25:51 GMT (1093kb,D)
[v2] Tue, 3 Dec 2019 13:26:56 GMT (1094kb,D)
[v3] Mon, 12 Oct 2020 13:23:20 GMT (2073kb,D)
Link back to: arXiv, form interface, contact.