It could be of great difficulty and cost to directly apply complex deep neural network to mobile devices with limited computing and endurance abilities. This paper aims to solve such problem through improving the compactness of the model and efficiency of computing. On the basis of MobileNet, a mainstream lightweight neural network, we proposed an Adaptive Tensor-Train Decomposition (ATTD) algorithm to solve the cumbersome problem of finding optimal decomposition rank. For its non-obviousness in the forward acceleration of GPU side, our strategy of choosing to use lower decomposition dimensions and moderate decomposition rank, and the using of dynamic programming, have effectively reduced the number of parameters and amount of computation. And then, we have also set up a real-time target network for mobile devices. With the support of sufficient amount of experiment results, the method proposed in this paper can greatly reduce the number of parameters and amount of computation, improving the model’s speed in deducing on mobile devices.

Zengrui Zhao

Papers

Adaptive Tensor-Train Decomposition for Neural Network Compression