UCL Psychology and Language Sciences


Experimental Psychology Seminar - Deep supervised neural network models of visual cortex: Moderate depth and category-specialised supervision are best

10 January 2017, 4:00 pm–5:00 pm

Event Information


Room 305, 26 Bedford Way London WC1H 0AP

Speaker: Katherine Storrs, University of Cambridge

Over the past four years, advances in machine learning have created computer programs roughly as good as humans at naming objects in images. These deep convolutional neural networks (DCNNs) are loosely inspired by the mammalian visual system, but do they learn similar solutions to the problems of visual object recognition? Evidence suggests they may; representations in a 7-layer DCNN trained to categorise diverse objects predicted representations in human inferior temporal (IT) cortex better than any of 27 alternative models (Khaligh-Razavi & Kriegeskorte, 2014). I will present recent work building on these results, investigating the effect of depth, object classification accuracy, and task training on the degree to which DCNNs predict brain representations. In earlier studies using shallower architectures with poorer object classification accuracy, greater depth and higher task performance were associated with improved explanation of IT (Yamins et al., 2014). We find that this is not the case for newer, deeper architectures that near or surpass human performance. We further find that training task and stimuli have a large impact on how well a model can explain representations in category-selective subregions of IT, the fusiform face area (FFA) and parahippocampal place area (PPA). A 16-layer DCNN trained exclusively to identify faces best explained FFA, while the same architecture trained to categorise scenes best explained PPA. Results suggest (1) that the deepest state-of-the-art engineering solutions to object recognition may be diverging from biological solutions, and (2) that to fully explain IT representations, networks may need to perform well on multiple tasks and/or explicitly model the functional organisation of human ventral visual cortex.

Time: 4pm, 10th January 2017

Venue: Room 305, 26 Bedford Way London WC1H 0AP