Neural network simplification using a progressive barrier based approach

Neural networks are indispensable to state-of-the-art artificial intelligence algorithms. However, its high accuracy comes at the cost of high computational complexity. This leads to the high operating cost of data centers and also hinders its deployment on mobile devices. In this thesis, we propose an algorithm to address this problem. The proposed algorithm uses progressive barriers to automatically and progressively simplify a pre-trained neural network until the target complexity is met while maximizing the accuracy. Along with the neural network that meets the target complexity, the algorithm also generates a family of simplified networks with different accuracy-complexity trade-offs, which allows for dynamic network selection and further study. Experiment results show that the algorithm achieves better accuracy-complexity trade-offs on a highly compact MobileNet architecture, compared with state-of-the-art automated network simplification approaches. For image classification on the ImageNet dataset, the algorithm reduces the number of multiply-accumulate operations by 1.68x while achieving 0.9% higher accuracy.