Mitigate IR-Drop Effect by Modulating Neuron Activation Functions for Implementing Neural Networks on Memristor Crossbar Arrays

The line resistance (LR) in a large-scale memristor crossbar array can cause serious IR-drop problem, degrading the hardware deployment capability of neural networks (NNs). In this work, two innovation schemes from the level of software are proposed to mitigate the hardware IR-drop problem by intentionally modulating the NN activation function before deploying. The methods are evaluated over typical activation functions and various line resistances on MLP and LeNet-5 for MNIST recognition. Results show the methods can significantly improve the tolerance of NNs to IR-drop and recover the accuracy in some extent. The methods require no extra hardware overhead and reduce the complexity of peripheral circuits, which make them more achievable and attractive.

Automated summary

In this work, two innovative schemes from the level of software are proposed to mitigate the hardware IR-drop problem caused by line resistance in a large-scale memristor crossbar array. The methods are tested on MLP and LeNet-5 neural networks for MNIST recognition using typical activation functions and various line resistances. Results demonstrate that the methods can significantly improve the tolerance of neural networks to IR-drop and recover accuracy to some extent. These methods require no additional hardware overhead and reduce the complexity of peripheral circuits, making them more achievable and attractive.