Implications of Noise in Resistive Memory on Deep Neural Networks for Image Classification

TL;DR

A special noisy operator is introduced that mimics the noise in an exemplary resistive memory unit, and the resilience of convolutional neural networks on the CIFAR-10 classification task is explored, and a couple of countermeasures to improve this resilience are discussed.

Abstract

Resistive memory is a promising alternative to SRAM, but is also an inherently unstable device that requires substantial effort to ensure correct read and write operations. To avoid the associated costs in terms of area, time and energy, the present work is concerned with exploring how much noise in memory operations can be tolerated by image classification tasks based on neural networks. We introduce a special noisy operator that mimics the noise in an exemplary resistive memory unit, explore the resilience of convolutional neural networks on the CIFAR-10 classification task, and discuss a couple of countermeasures to improve this resilience.

Figures (10)

• Figure 1: Cycle-to-cycle variability for LRS (upper subfigure) and HRS (lower subfigure) measured on an HfOx test device for different forming voltages.
• Figure 2: Exemplary application of the fliptensors $f_1$ and $f_2$ to a vector with three 4-bit integer values.
• Figure 3: Validation accuracy for VGGs of different depths for float32 data type.
• Figure 4: Midpoint noise level $\mu$ (multiplied by $10^{-6}$) for different VGG depths and data types.
• Figure 5: Sweep over noisy bits for VGG-A on CIFAR-10. Explanation of the legend: the number is the position of the first noisy bit in big-endian format starting from 0. For instance, case "1" means that all but the sign bit are noisy.