Computational Vision & Neuroscience

(Matthias Bethge)

Perceiving Neural Networks

Perception is a computational feat. The conversion of high-dimensional sensory input to meaning relies on the ability to solve complex pattern recognition problems. Natural tasks like object recognition or visual search are good examples of this process revealing the computational challenges underlying perception. To tackle these challenges complex neural networks have developed in the brain that perform surprisingly well. At the interface between artificial intelligence and neuroscience we focus on uncovering the algorithms and neuro-computational design principles of perceiving neural networks. A practical example of the outcome of this research is a new method for creating artistic images. More specifically, we want to explain how characteristic properties of neural systems originate from the computational requirements of specific perceptual skills:

Neural systems ← Computation → Perceptual skills

and we use Machine Learning and Computational Neuroscience methods to study the problem of perceptual inference and its neural basis at different levels:

• Neurobiology
• Theory Perception
• Neural data analysis
• Deep neural networks
• Psychophysics
• Generative image modeling
• Population coding