Drift regularization to compensate for variations in the drift coefficient of the simulated accelerator
By introducing a drift regularization function during training, the variance of neuron values is reduced, thus solving the problem of classification accuracy degradation caused by conductance drift in analog accelerators and improving the performance of analog accelerators on edge devices.
CN115699031BActive Publication Date: 2026-05-26INTERNATIONAL BUSINESS MACHINE CORPORATION
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- INTERNATIONAL BUSINESS MACHINE CORPORATION
- Filing Date
- 2021-06-04
- Publication Date
- 2026-05-26
AI Technical Summary
Technical Problem
The inaccuracy of deep neural network weights caused by conductance drift during the training process of phase-change memory-based analog accelerators leads to a degradation in classification accuracy, which has become a major obstacle limiting their application in edge computing.
Method used
By introducing a drift regularization function during training, the variance of neuron values is reduced, the network's noise robustness is improved, and the classification accuracy degradation caused by conductance noise is reduced.
Benefits of technology
It effectively reduces the degradation of classification accuracy caused by conductivity noise, improves the performance of the analog accelerator on edge devices, and maintains high classification accuracy.
✦ Generated by Eureka AI based on patent content.
Smart Images
Figure CN115699031B_ABST
Abstract
A drift regularization is provided to counteract variations in drift coefficients in an analog neural network. A method for training an artificial neural network is illustrated. Multiple weights are randomly initialized. Each of these weights corresponds to a synapse in the artificial neural network. At least one input array is fed into the artificial neural network. The artificial neural network determines at least one output array based on the at least one input array and the multiple weights. The at least one output array is compared to ground truth data to determine a first loss. A second loss is determined by adding drift regularization to the first loss. The drift regularization is positively correlated with the variance of the at least one output array. The multiple weights are updated based on the second loss via backpropagation.
Need to check novelty before this filing date? Find Prior Art