Deep Neural Networks(DNNs) have demonstrated state-of-art performance on many problems in different fields such as computer vision, natural language processing, and others. Training of (DNNs) is computationally an expansive because of the backpropagation algorithm. In the inference stage,(DNNs) model requires an expansive computation and high storage cost, which makes them a challenging problem for devices that have constrained resources on the edge. In order to mitigate this issue, a plethora of methods has been applied and published in the literature in recent years. In this paper, we present FPTT, a two-stage pipeline: filter pruning, and Tensor Train Decomposition, that work together to reduce the storage and computation requirements of (DNNs) in order to deploy them easily on the edge. Our work demonstrates the same accuracy or tiny degradation of accuracy with retraining after applying both stages. In particular, for the VGG-16 on CIFAR-10, CIFAR-100 and ImageNet we report compression reduction of 85%, 84.03%, and 93.69% respectively and for Alexnet on CIFAR-10, CIFAR-100 we report compression reduction of 96.5% and 94.5%, and for LeNet-5 on MNIST and CIFAR-10 we report compression rate of 93.2% and 98.71%.
Paper
Full text
Efficient Deep Neural Networks for Edge Computing
Semantic Scholar · Computer Science · 2019
Abstract
Deep Neural Networks(DNNs) have demonstrated state-of-art performance on many problems in different fields such as computer vision, natural language processing, and others. Training of (DNNs) is computationally an expansive because of the backpropagation algorithm. In the inference stage,(DNNs) model requires an expansive computation and high storage cost, which makes them a challenging problem for devices that have constrained resources on the edge. In order to mitigate this issue, a plethora of methods has been applied and published in the literature in recent years. In this paper, we present FPTT, a two-stage pipeline: filter pruning, and Tensor Train Decomposition, that work together to reduce the storage and computation requirements of (DNNs) in order to deploy them easily on the edge. Our work demonstrates the same accuracy or tiny degradation of accuracy with retraining after applying both stages. In particular, for the VGG-16 on CIFAR-10, CIFAR-100 and ImageNet we report compression reduction of 85%, 84.03%, and 93.69% respectively and for Alexnet on CIFAR-10, CIFAR-100 we report compression reduction of 96.5% and 94.5%, and for LeNet-5 on MNIST and CIFAR-10 we report compression rate of 93.2% and 98.71%.