Thesis defense Katja Müller (Donders series 482)
22 January 2021
Promotor: prof. dr. M. van Gerven
Co-promotor: dr. U. Güclu
Convolutional neural networks in vision neuroscience
How do you make sense of the light waves entering your eye right now? Neuroscience has come closer to the answer during the last decade. When an old information processing model of the visual system led to the sudden breakthrough in artificial intelligence, neuroscience took notice. Neuroscientists discovered that – when it comes to seeing and hearing – so-called deep neural networks are actually quite a bit like the brain. The thesis has been situated in this spectrum of research. We first confirmed that deep neural networks and our visual system process information in a similar temporal order. However one problem with such comparison is that neural networks from AI have very simple goals, like recognizing an arbitrarily small set of objects. This makes them less similar to brain information processing than they could be. So we set out to train deep neural networks directly on brain data. As neural networks need a lot of examples, for this we required a lot of data from one human brain. We found a willing master student and put him into the brain scanner for dozens of hours to acquire the largest dataset of a brain responding to what it sees (in this case, BBC's Doctor Who) recorded to date. After training a neural network from his real visual system, several properties of his visual system emerged, and we could use the artificial neural network to begin peeking into what its parts really respond to.