From random to trained: Investigating the application of convolutional neural networks as models for infants’ visual brain development

PhD Seminar

July 10th, 2020
Versione stampabile
Venue: Online Seminar offered through Zoom Platform
Time: 11.00 - 12.00
 

Doctoral Course of Cognitive Science - PhD Seminar

Speaker

  •  Anna Truzzi - Trinity College Dublin (Ireland)
Scientific coordinator: Gianluca Esposito
 

Abstract

To understand the development of the infant brain and mind in the first year of life, it would be valuable to have computational models of the learning process. In adults, deep neural networks (DNNs), which have revolutionised artificial intelligence, are currently the best available computational models for a number of brain systems. For instance, convolutional neural networks (CNNs) trained on object classification are currently the best model of visual responses in adults’ inferior temporal cortex (IT). It has been argued that this is because the CNNs, during training, learn features that allow them to represent object classes using features similar to the ones used by the adult brain. This similarity could make them good models for human development as well. However, recent results challenge this perspective: untrained random-weight CNNs contain representations that correlate with IT representations as well as, or even better than, trained CNNs. Although apparently in conflict, IT contains rich representations of both perceptual and semantic features, and it could be that random and trained networks are actually capturing different aspects of IT representations. In our project we applied regression models to investigate whether combining a trained CNN with a random network, which showed a good correlation with IT, could explain IT better than the trained network alone. Furthermore, to characterise which aspects of IT variability were explained by each network, we visualised their representations and then applied mediation models to quantify whether and how much perceptual and semantic features of the visual input mediated the correlation between trained vs random CNNs and IT. Results across analysis consistently showed that random and trained CNNs capture different aspects of IT and that the similarity of the random CNN and IT was mostly mediated by perceptual features, while the similarity of the trained CNN and IT was mostly mediated by semantic features. Findings from this project highlight that it is not enough to know how similar a given CNN is to the brain: we also need to know why. Understanding which characteristics of the learning process makes CNNs’ representation similar to the adult brain will suggest new exciting testable hypothesis on how the infant brain develops and learns.