Ferenc Huszar

Monday 18th March 2019

Time: 4.00pm

Ground Floor Seminar Room

25 Howland Street, London, W1T 4JG

Deep Learning for Visual Data Compression

Deep learning has proven remarkably successful at modeling the statistics of dense, high-dimensional data, and dethroned traditional signal processing-based approaches as the state-of-the-art in in several domains from speech processing to object segmentation. It is likely that deep learning will eventually revolutionize video and image compression as well. 

Before this is possible, a number of key challenges remain: 
Convolutional neural networks are computationally demanding; 
Deep learning does not work well with the discrete, non-differentiable nature of data compression; We lack good inductive biases and priors for dealing with temporal aspects of video; 
Finally, we need more robust tools to measure, predict and directly optimize for subjective perceptual quality. 

In this talk I provide an overview of the work our team has done over the past couple years to address each of these challenges.

Biography:
Ferenc Huszar (fhuszar@twitter.com) is a Senior Machine Learning Researcher at Twitter. He currently works on real-time recommender systems, and is also involved in Twitter’s ethical machine learning working group. Prior to this Ferenc joined Magic Pony Technology, a startup developing deep learning-based video processing technology, acquired by Twitter in 2016. Ferenc holds a PhD in Machine Learning from the University of Cambridge, and is an active contributor of his research blog inference.vc.