Elliot Crowley
Tue 04 Dec 2018, 11:00 - 12:00
IF 4.31/4.33

If you have a question about this talk, please contact: Gareth Beedham (gbeedham)

Title: pruning and squeezing deep neural networks


Abstract: Recently there has been an increasing demand for efficient Deep Learning methods to be used at different scales. However, such a need is in contrast with the standard assumptions of modern machine learning techniques that are trained with soft constraints on both computational power and storage. Although massive computation is available at server level, this is not always the case at user level, where limits in hardware and performance impact the deployment of expensive models on personal devices. We are investigating possible ways to solve these issues following two main directions: (i) pruning unnecessary parameters, (ii) dynamically adapting the network structure based on the characteristics of the input. This presentation introduces what we have done so far and the future line of research we are currently considering.