Understanding Deep Learning and Human Visual Perception

The success of deep learning neural networks is spectacular. Research into the mechanisms of AI is ubiquitous, and mainly focused on huge datasets, cheap GPU parallel processing, and clever network topologies. But still it is considered largely to be a black box inside.

A relatively underestimated approach is to exploit how much we can learn from modern brain research. The brain is with 25 Watt and 6 KHz neurons also many orders more efficient. The biological and AI worlds speak different languages.

This talk likes to bridge this gap and will highlight some new optical and physiological recording techniques, especially from functional circuits in the layers of the retina and visual cortex. We discuss the ‘neuro-mathematics’ intuition of how the brain and CNNs do a geometrical analysis, and how receptive fields and first layers are formed. We pay special attention to the multi-scale and multi-orientation analysis discovered in human vision and how they relate to the key problem of perceptual grouping.

Throughout the lecture we will give practical applications with respect to computer vision and machine learning, especially in medical image analysis. The lecture will be highly visual and is aimed to a broad audience. The speaker is known for his excellent educational skills.