Scaling Down: Efficient Inference for Convolutional Neural Networks

Convolutional neural networks achieve impressive results for image recognition tasks, but are often too large to be used efficiently for inference applications. In this paper, we explore several efficient architectures that satisfy a baseline accuracy on an image recognition task. For this task, accuracy is defined as the number of correctly identified images over total images. We train a NasNet-A convolutional neural network to an accuracy of 0.8034 that has 5.2M parameters, 662M multiplication operations, and 659M addition operations. When comparing this model against the baseline model WideResNet-28-10, it achieves a score of 0.1659 using the Micronet Challenge scoring scheme. The Micronet Challenge score is defined as the sum of the number of parameters and number of multiplications and additions, normalized by the number of parameters and multiplications and additions for the baseline model WideResNet-28-10