The human visual system is not a general-purpose object recognition machine: Evidence from neuroimaging and artificial neural networks
By | Hans Op de Beeck | Group leader of Human Brain Imaging and Rodent Visual Cognition | Laboratory of Biological Psychology | Belgium
Venue | Senior Common Room, Level 2 (2D17), Priory Road Complex
Humans are able to parse the complex flow of visual information into meaningful objects that can be recognized and acted upon. Recently, computer models such as deep convolutional neural networks (convNets) have been proposed that can reach unprecedented performance levels in object recognition. I will illustrate some striking examples of the similarities between information processing in these convNets and in the human visual brain, using data from published work (e.g., Bracci et al., 2016, J. Neurosci.; Kubilius et al., 2016, PLOS Comput. Biol.) and ongoing studies. The underlying multidimensional representations can be understood as a feature-based categorical code (Bracci et al., 2017, Neuropsychologia). However, information processing can be influenced by the learning history and goals of the network, be it human or artificial. These factors might underlie a striking discrepancy between convNets and human representations in which the human object vision pathway is biased to wrongly categorize some inanimate objects as being animate. For example, for human object representations, in contrast to convNets, a cow-shaped mug is a cow and not a mug. As a consequence, human object vision cannot be seen as a general-purpose and objective recognition system. To the contrary, it is a subjective system shaped and biased through its learning history and goals. Mimicking these dimensions in machine learning will be important to increase the similarity between humans and machines.
All Welcome | Tea, coffee and biscuits will be available after the seminar.