News & Events

Deep Learning Inference with Limited Resources

Vincent Gripon

Wegmans Hall 1400

Abstract:

Deep  learning  architectures  are  the  golden   standard   for   many   machine   learning problems. Thanks  to  their  large  number  of  trainable  parameters,  they  are  able   to absorb  complex dependencies  in  the  input  data  to  produce  correct  decisions,  when trained appropriately. However, this dependency on a very  large  number  of  parameters  is also  a  weakness:  their computation  and memory  footprints  are  considerable  and  it  is  hard — if not impossible — to guarantee their ability to perform well when dealing with corrupted and  noisy  inputs.  In  this  talk,  we shall  review  the main  strategies  that  have  been proposed  in  the  literature  to  reduce  computations  and  memory of  deep   learning systems, including quantization, factorization, and pruning. We shall also discuss how adequate are these systems to faulty implementations. In a last part, we will discuss the susceptibility of deep learning architectures to deviations of the inputs, what appears to have become a major open question.

Biography:

Vincent Gripon (S'10, M'12) is a permanent researcher with IMT Atlantique, a french top technical university. He obtained his M.S from École Normale Supérieure Paris-Saclay in 2008 and his PhD from IMT Atlantique in 2011. He spent one year as a visiting scientist at McGill University between 2011 and 2012 and he is currently an invited Professor at Mila and Université de Montréal. His research mainly focuses on efficient implementation of artificial neural networks, graph signal processing, deep learning robustness and  associative memories.  He  co-authored  more  than  70 papers in these domains in prestigious venues (AAAI, tPAMI, Statistical Physics, TNNLS, TSP...).

Refreshments will be provided