Home // SIGNAL 2023, The Eighth International Conference on Advances in Signal, Image and Video Processing // View article
Authors:
Florent Crozet
Stéphane Mancini
Marina Nicolas
Keywords: Convolutional Neural Network Compression, pseudo-random number generators
Abstract:
With the proliferation of convolutional neural network (CNN)-based computer vision solutions, computing inference on smart sensors has become a challenge. The inference of CNN is difficult to embed in such tiny devices due to the constraints on memory. To address this challenge, we propose a compression method able to reduce the number of weights to store in a structured way, so that the gain in the number of weights comes with a gain in the number of computations at inference. Our solution is based on the replacement of the convolutional filters by a linear combination of some stored filters and a set of seeds corresponding to pseudo-random generated filters. During the inference, pseudo-random number generators are used to compute the non-stored filters, thanks to the associated seeds. On the other side, the linear combination allows mutualizing partly the cost of convolutions. We show that further exchanging memory for a small logic cost to generate the pseudo random filters allows to decrease the number of weights significantly, on several state-of-the-art networks without sacrificing the accuracy. For example, applying this method to CNNs like ResNet50 leads to a compression factor of 2.5 for less than 5% accuracy drop. Furthermore, our method is compatible with compression methods targeting the precision of the weights to store, namely quantization. This gives room to further in
Pages: 24 to 30
Copyright: Copyright (c) IARIA, 2023
Publication date: March 13, 2023
Published in: conference
ISSN: 2519-8432
ISBN: 978-1-68558-057-5
Location: Barcelona, Spain
Dates: from March 13, 2023 to March 17, 2023