Dawood Abbas Abdul Malik Zixiao Chen Audrey Chu Elena Georgieva
A final project for Computational Cognitive Modeling at NYU - Spring 2021
Abstract:
Understanding similarity is a central question in the field of cognitive science. Similarity is used in industry to tackle topics such as semantic segmentation, recommendation systems, scene understanding, and even text mining. Deep neural networks have been the popular choice in solving these perception problems and have even reach or surpassed human-level accuracy. In this paper, we examine and extend on the methodologies implemented in Adapting Deep Network Features to Capture Psychological Representations (J. Peterson and T.Griffiths, 2016). In addition to developing three neural networks, we collect similarity ratings data from 57 human subjects. We examine the connection between representations learned by the networks and human judgements, and also manipulate the deep features to align more with human judgements. We further analyze human judgements by comparing similarity judgements based on survey participant behavior such as time to complete and rating by stimuli sequence. By understanding how image representations differ for the same person, we can apply this extended human behavior knowledge to areas such as image recommendation platforms.
