A new software based in-situ training (SBIST) method to achieve high accuracies is proposed for binarized neural networks inference accelerator chips in which measured offsets in sense amplifiers (activation binarizers) are transformed into biases in the training software. To expedite this individual training, the initial values for the weights are taken from results of a common forming training process which is conducted in advance by using the offset fluctuation distribution averaged over the fabrication line. SPICE simulation inference results for the accelerator predict that the accuracy recovers to higher than 90% even when the amplifier offset is as large as 40mV only after a few epochs of the individual training.
Zian CHEN
Waseda University
Takashi OHSAWA
Waseda University
The copyright of the original papers published on this site belongs to IEICE. Unauthorized use of the original or translated papers is prohibited. See IEICE Provisions on Copyright for details.
Copy
Zian CHEN, Takashi OHSAWA, "A Low-Cost Training Method of ReRAM Inference Accelerator Chips for Binarized Neural Networks to Recover Accuracy Degradation due to Statistical Variabilities" in IEICE TRANSACTIONS on Electronics,
vol. E105-C, no. 8, pp. 375-384, August 2022, doi: 10.1587/transele.2021ECP5040.
Abstract: A new software based in-situ training (SBIST) method to achieve high accuracies is proposed for binarized neural networks inference accelerator chips in which measured offsets in sense amplifiers (activation binarizers) are transformed into biases in the training software. To expedite this individual training, the initial values for the weights are taken from results of a common forming training process which is conducted in advance by using the offset fluctuation distribution averaged over the fabrication line. SPICE simulation inference results for the accelerator predict that the accuracy recovers to higher than 90% even when the amplifier offset is as large as 40mV only after a few epochs of the individual training.
URL: https://global.ieice.org/en_transactions/electronics/10.1587/transele.2021ECP5040/_p
Copy
@ARTICLE{e105-c_8_375,
author={Zian CHEN, Takashi OHSAWA, },
journal={IEICE TRANSACTIONS on Electronics},
title={A Low-Cost Training Method of ReRAM Inference Accelerator Chips for Binarized Neural Networks to Recover Accuracy Degradation due to Statistical Variabilities},
year={2022},
volume={E105-C},
number={8},
pages={375-384},
abstract={A new software based in-situ training (SBIST) method to achieve high accuracies is proposed for binarized neural networks inference accelerator chips in which measured offsets in sense amplifiers (activation binarizers) are transformed into biases in the training software. To expedite this individual training, the initial values for the weights are taken from results of a common forming training process which is conducted in advance by using the offset fluctuation distribution averaged over the fabrication line. SPICE simulation inference results for the accelerator predict that the accuracy recovers to higher than 90% even when the amplifier offset is as large as 40mV only after a few epochs of the individual training.},
keywords={},
doi={10.1587/transele.2021ECP5040},
ISSN={1745-1353},
month={August},}
Copy
TY - JOUR
TI - A Low-Cost Training Method of ReRAM Inference Accelerator Chips for Binarized Neural Networks to Recover Accuracy Degradation due to Statistical Variabilities
T2 - IEICE TRANSACTIONS on Electronics
SP - 375
EP - 384
AU - Zian CHEN
AU - Takashi OHSAWA
PY - 2022
DO - 10.1587/transele.2021ECP5040
JO - IEICE TRANSACTIONS on Electronics
SN - 1745-1353
VL - E105-C
IS - 8
JA - IEICE TRANSACTIONS on Electronics
Y1 - August 2022
AB - A new software based in-situ training (SBIST) method to achieve high accuracies is proposed for binarized neural networks inference accelerator chips in which measured offsets in sense amplifiers (activation binarizers) are transformed into biases in the training software. To expedite this individual training, the initial values for the weights are taken from results of a common forming training process which is conducted in advance by using the offset fluctuation distribution averaged over the fabrication line. SPICE simulation inference results for the accelerator predict that the accuracy recovers to higher than 90% even when the amplifier offset is as large as 40mV only after a few epochs of the individual training.
ER -