Deep neural networks enable state-of-the-art accuracy in visual recognition tasks such as image classification and object recognition. However, modern networks contain millions of learned connections, and the current trend is toward deeper and more tightly connected architectures. This poses a challenge for the deployment of advanced networks on resource-constrained systems such as smartphones or mobile applications. To make neural networks on embedded devices more usable, there are different techniques to compress the models.
In this Talk the most common compression algorithms will be presented and their functionality explained. Among the techniques presented will be pruning, quantization and others.