Nikolaus Kriegeskorte, PhD

Understanding brain-computational mechanisms by testing deep neural network models with massively multivariate brain-activity data

Recent advances in neural network modelling have enabled major strides in computer vision and other artificial intelligence applications. Artificial neural networks are inspired by the brain and their computations could be implemented in biological neurons. Although designed with engineering goals, this technology provides the basis for tomorrow’s computational neuroscience, engaging complex cognitive tasks and high-level cortical representations. We are entering an exciting new era, in which we will be able to build neurobiologically faithful feedforward and recurrent computational models of how biological brains perform high-level feats of intelligence.

The objective of the lab is to understand the brain information processing that enables visual perception, object recognition, and scene understanding. Vision is of interest in its own right, but also provides a model for understanding, more generally, how the brain computes and how it might perform probabilistic inference through parallel and recurrent computations.

The lab uses massively multivariate measurements of brain activity along with behavioural data to test models of brain information processing that perform visual tasks. To explain visual processing, the models must meet computational challenges comparable to those biological visual systems face in the real world. The models therefore need to contain rich visual knowledge about the world and have substantial computational power. Building such models requires the methods of machine learning and artificial intelligence. We take a top-down approach to modelling, starting with models that perform the task, but abstract from much of the biological detail. We then attempt to reveal the aspects of human task performance and brain activity that these models fail to explain. This motivates adjustments to the architecture and the design of the units. Architectures and units must be plausibly implementable with biological neurons. Their design is chosen as required by function and inspired by biology, so as to better explain brain and behavioural data. The lab develops neural net models, statistical inference and visualisation techniques, and visual stimuli and tasks, and measures brain activity with fMRI and MEG in humans and with array recordings in nonhuman primates.