How do deep neural nets really work?
By Prof. Bart ter Haar Romeny (Eindhoven University of Technology, Biomedical Engineering)
Deep neural nets are ubiquitous, and have seen spectacular breakthroughs in performance in many fields. Deep learning is in great majority approached with heuristic optimization. However, its internal workings are still for a large part a black box. Explainability of DL, needed to build trust, is a large research topic today, approached from many directions, like visualization, layer-wise relevance propagation, perturbation, local approximation etc. But no good answers yet …
In this extended tutorial talk I will focus on a geometric approach inspired by recent findings in the visual system. One of the key strategies of the brain is efficiency, so we will discuss extensively several methods to accomplish this, such as approximate representation learning, updating by differences only, and attention.
In particular, we will study how perceptual grouping might work (‘visual binding’), i.e. contextual processing. Contrary to data/dimension reduction, we see an explosion of data in the visual system. The retina measures with 26 different mosaics simultaneously, and (as in deep networks) in higher levels the local descriptions get far richer, though at lower resolution, i.e. deeper in the graph hierarchy. We start in a single pixel, develop differential operators to study direct neighborhoods, and study affinities between the tensors in higher levels. Current DL approaches are primarily static, but motion of the scene or observer is a crucial binding factor. Another topic is a new model to understand generative adversarial networks (GANs), based on classical techniques in computer vision.
This lecture is special in the sense that -everything- is illustrated with life coding, proving the feasibility and understandability of the presented method.
Prof. Romeny has spent his whole career in geometrical biomedical image analysis, exploiting ‘brain-inspired computing’. The methods developed in that computer vision community turn out to be today eminently suitable for explainable AI (XAI). Prof. Romeny is a awarded speaker, and will give this extended lecture as his farewell as president of our society for 9 years.
We have three more at this event:
• Alex Frangi (Leeds Univ.), an outstanding computer vision researcher from England tells about research at the crossroads of image analysis and modeling with emphasis on machine learning.
• Efstratios Gavves(UvA), will teach us on the so crucial (and rewarding) dynamics of temporal machine learning.
• Veronika Cheplygina(IT Univ. of Copenhagen), also leaving the board, has an exciting lecture about shortcomings of machine learning and recommendations to overcome these.
Share this Post