Image Classification Model for Edge Devices

Abstract

Machine learning is a potentially powerful tool for medical diagnoses, when empowered by very deep neural networks and very powerful hardware. However, that hardware is expensive and other solutions rely on a steady internet connection that is not guaranteed in most parts of the world. Skin cancer is especially prevalent in poorer, more rural areas and a lightweight tool for edge devices could save many lives. This thesis presents a convolutional neural network that is on-par with the performance of various control models at a fraction of the model size. We trained a series of potential solutions on the HAM10000 [1] dataset of skin lesions. This series includes SqueezeNet [2] and several deeper iterations of the SqueezeNet architecture, all of which weigh in less than 30 MiB. The solution with the best accuracy, however, turned out to be the original SqueezeNet design. We further tested and ensured that even an aggressive pruning schedule does not reduce accuracy, and that the model can be effectively quantized to run on a Google Edge Tensor Processing Unit [3]. Our proposed solutions were compared with AlexNet [4], to remain consistent with the original literature on SqueezeNet.