Efficient Sneak Path-aware Training of Binarized Neural Networks for RRAM Crossbar Arrays
Cited 0 times inCited 0 times in
- Efficient Sneak Path-aware Training of Binarized Neural Networks for RRAM Crossbar Arrays
- Lee, Sugil
- Lee, Jongeun
- Issue Date
- Graduate School of UNIST
- Although RRAM crossbar arrays have been suggested as an efficient way to implement MVM for DNNS, the sneak path problem of RRAM crossbar arrays due to wire resistance can distort the result of MVM quite significantly, resulting harsh performance degradation of the network. Therefore, a software solution that can predict the effect of sneak paths to mitigate the impact without permanent hardware cost or expensive SPICE simulations is very desirable. In this paper, a novel method to incorporate the sneak path problem during training with a negligible overhead is proposed. The test validation results, done through accurate SPICE simulations, show very high improvement in the performance close to the baseline BNNs on GPU, which demonstrates the efficiency of the proposed method to capture the sneak path problem.
- Department of Computer Science and Engineering
- Go to Link;
- Appears in Collections:
- Files in This Item:
Efficient Sneak Path-aware Training of Binarized Neural Networks for RRAM Crossbar Arrays.pdf
can give you direct access to the published full text of this article. (UNISTARs only)
Show full item record
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.