Enzo Tartaglione, visiting scholar from University of Turin, will give a talk about Neural Network compression, on Wednesday 12 Feb in room 5A126 at 10am.
Title: Introducing sparsity in artificial neural networks: a sensitivity-based approach.
Abstract: The ever-increasing number of parameters in deep neural networks poses challenges for memory-limited applications. Recently many approaches have been proposed to efficiently solve this issue, ranging from proposing new architectures to quantization-based techniques. Regularize-and-prune methods aim at sparsifying the neural network weights.
In this talk we are going to introduce sensitivity-based regularization, a regularize-and-prune strategy whose objective is to recognize the parameters not useful for the given learning task. Once these are detected, their value is driven towards zero: thus, a very large fraction of the parameters can be pruned by thresholding. The sparsity can be structured or un-structured: we are going to see variants of sensitivity-based approach for both un-structured and structured (SE-RE-NE) regularization.
The proposed methods allow to significantly reduce the size of the trained models: in particular, SE-RE-NE allows headless effective size reduction and faster inference time.
Speaker’s Short Bio: Enzo Tartaglione received a summa cum laude degree in Electronic Engineering in 2015 from Politecnico di Torino, Italy. In the same year he received a magna cum laude Master of Science in Electrical and Computer Engineering from University of Illinois at Chicago, USA. In 2019 he received a PhD cum laude from Politecnico di Torino, with the thesis “From Statistical Physics to Algorithms in Deep Neural Systems”. He is currently a member of the EidosLAB at University of Torino as a postdoctoral researcher, working in the DeepHealth project. His main research interests are in the area of deep learning, with a focus on pruning and regularization strategies.