Description
Talk: ‘Towards an efficient embedded vision through neural network compression’
Nowadays, deep learning plays a key role in many areas of computer science and related fields such as computer vision, speech recognition, natural language processing, robotics, and so on. In general, deep models require a large amount of training data, have a significant memory footprint, and often involve high computational complexity during the inference phase. To democratize deep learning on embedded devices with limited computational resources, it is essential to design lightweight models that are effective in energy consommation and computational complexity while maintaining strong performance, making them suitable for deployment on edge devices.
This talk presents our recent efforts in designing lightweight deep models or compressing pre-trained deep networks without significantly degrading their performance. We introduce efficient neural architectures specifically designed for deployment on embedded systems. Furthermore, we present various methods for effectively compressing pre-trained convolutional models, relying on neural network pruning techniques or tensor decomposition methods to represent convolutional layers with low-rank approximations.
