Robustness of Memristor-Based Design of Accurate Binary Differential Convolutional Neural Network for Biomedical Image Classification

This paper presents a novel edge computing system named memristor-based accurate binary differential convolutional neural network (MABDNet) for biomedical image classification. The system utilizes precise binary differential algorithms to approximate full-precision weights, thereby maintaining accuracy while enhancing robustness, rendering it more suitable for hardware implementation. Experimental results demonstrate that when the yield rate of memristor arrays is greater than or equal to 50%, MABDNet achieves a classification accuracy of 85.15% on the MedMNIST v2 dataset, whereas traditional binary neural networks (BNNs) only attain 51.99%. Furthermore, it is observed that with an increase in noise intensity, the classification accuracy of MABDNet also improves, indicating that moderate introduction of Gaussian noise can enhance the robustness of MABDNet to a certain extent.