US 12,293,281 B2
Training DNN by updating an array using a chopper
Tayfun Gokmen, Briarcliff Manor, NY (US)
Assigned to International Business Machines Corporation, Armonk, NY (US)
Filed by International Business Machines Corporation, Armonk, NY (US)
Filed on Apr. 9, 2021, as Appl. No. 17/226,416.
Prior Publication US 2022/0327375 A1, Oct. 13, 2022
Int. Cl. G06N 3/065 (2023.01); G06F 11/14 (2006.01); G06F 17/16 (2006.01); G06N 3/08 (2023.01); G11C 5/06 (2006.01)
CPC G06N 3/065 (2023.01) [G05B 2219/21008 (2013.01); G05B 2219/25255 (2013.01); G05B 2219/32335 (2013.01); G05B 2219/33027 (2013.01); G05B 2219/33033 (2013.01); G05B 2219/33186 (2013.01); G05B 2219/41397 (2013.01); G06F 11/1476 (2013.01); G06F 17/16 (2013.01); G06N 3/08 (2013.01); G11C 5/063 (2013.01)] 21 Claims
OG exemplary drawing
 
1. A computer-implemented method of training a deep neural network (DNN), the method comprising:
determining incremental weight updates by updating an element of an A matrix with activation values and error values from a weight matrix multiplied by a chopper value comprising a random selection from the group consisting of: a positive one (+1) and a negative one (−1), wherein the element comprises an analog resistive processing unit (RPU);
reading an update voltage from the element;
determining a chopper product by multiplying the update voltage by the chopper value;
storing an element of a hidden matrix, wherein the element of the hidden matrix comprises a summation of continuous iterations of the chopper product; and
updating a corresponding element of a weight matrix based on the element of the hidden matrix reaching a threshold state, wherein the hidden matrix produces a low-pass filter to mitigate noise of the analog RPU, and the chopper value mitigates bias.