A Relevance-Based CNN Trimming Method for Low-Resources Embedded Vision

Abstract

A significant amount of Deep Learning research deals with the reduction of network complexity. In most scenarios the preservation of very high performance has priority over size reduction. However, when dealing with embedded systems, the limited amount of resources forces a switch in perspective. In fact, being able to dramatically reduce complexity could be a stronger requisite for overall feasibility than excellent performance. In this paper we propose a simple to implement yet effective method to largely reduce the size of Convolutional Neural Networks with minimal impact on their performance. The key idea is to assess the relevance of each kernel with respect to a representative dataset by computing the output of its activation function and to trim them accordingly. The resulting network becomes small enough to be adopted on embedded hardware, such as smart cameras or lightweight edge processing units. In order to assess the capability of our method with respect to real-world scenarios, we adopted it to shrink two different pre-trained networks to be hosted on general purpose low-end FPGA hardware to be found in embedded cameras. Our experiments demonstrated both the overall feasibility of the method and its superior performance when compared with similar size-reducing techniques introduced in recent literature.

Quantization-Aware NN Layers with High-throughput FPGA Implementation for Edge AI

“…The fully-connected layer produced an output tensor of size 3 (one-hot encoding of the 3 classes) through softmax activation. With a simple pruning procedure, similar to [ 53 ], we reduced the size of the network, removing unnecessary kernels. The algorithm ranks the filters based on their output (after the activation) when predicting a random sample of the training set.…”