Novice and Expertise with Supervised VAE Model

Latent representation of Novice and Expertise using SVAE

One long-standing question in vision science and cognitive psychology is how people learn and recognize categories of objects. Objects can be categorized in various levels of abstraction. As individuals gain expertise within a certain category domain, they become adept at differentiating exemplars at the subordinate levels. Whether such perceptual expertise is a result of changes in representations and/or other cognitive processes like decision making remains inconclusive. In this study, we simulate how visual experience in a certain category modulates the object category representations across basic- and subordinate-levels by using a Supervised Variational Autoencoder (SVAE), which isa variance of VAE with a classification component. To simulate a visual experience of a novice, we initially trained the model to learn the basic categories of objects with the CIFAR-10 dataset. After that, we simulated the visual experience of a bird expert by subsequently training the model to learn subordinate-level categories of birds. To examine the influence of supervision signals on the development of expertise, we tested several different methods to test the development of subordinate-level category representations of a bird with a subset of CUB-200-2011 dataset. We succeeded in training both novice and expert models to reconstruct the images, but the latent representations that models learned did not show clear evidence of categorical representation. Possible reasons why it was difficult to learn the meaningful latent representations while preserving decent reconstruction performance and how the current framework could be improved are discussed.