Deep networks: artificial vs. biological
The last decade has seen an explosive development in the application of artificial intelligence (AI), in particular deep networks. Convolutional neural networks (CNNs) now outperform humans on many visual classification tasks. AI has clearly been inspired by the functioning of real brains, but AI may also help in unraveling how ‘biological intelligence’, that is, real brains work.
With the advent of modern supercomputers, the mathematical exploration of large-scale biological network models mimicking ‘biological intelligence’ is now becoming feasible. Recently, the Allen Brain Institute in Seattle has developed a state-of-the-art biophysically detailed mathematical model of mouse primary visual cortex comprising 230.000 neurons (Billeh et al., Neuron, 2020).
The tuning of the vast number of model parameters to make the model behave as real mouse brains, is challenging. However, intriguing findings on the similarity of neural representations in biological and artificial networks (figure) suggest the possibility that tuning of the Allen mouse model can be aided by detailed analysis of CNNs trained on image recognition.
In the project, which will be done in collaboration with Allen Institute, the connections between biological networks of mouse visual cortex and CNNs trained for image recognition will be explored, with the aim to improve both.
- MSc in physics with a large computational component will be preferred, but candidates with a MSc in another mathematically oriented subject may be considered.
- Candidates with documented experience in computational neuroscience and scientific programming will be prioritized.
Call 2: Project start autumn 2022
This project is in call 2, starting autumn 2022.