Disentangling Generativity in Visual Cognition
- Declan Campbell, Neuroscience, University of Wisconsin -- Madison, Madison, Wisconsin, United States
- Timothy Rogers, Psychology, University of Wisconsin- Madison, Madison, Wisconsin, United States
AbstractHuman knowledge is generative: from everyday learning people extract latent features that can recombine to produce new imagined forms. This ability is critical to cognition, but its computational bases remain elusive. Recent research with β-regularized Variational Autoencoders (β-VAE) suggests that generativity in visual cognition may depend on learning disentangled (localist) feature representations. We tested this proposal by training β-VAE’s and standard autoencoders to reconstruct bitmaps showing a single object varying in shape, size, location, and color, and manipulating hyperparameters to produce differentially-entangled feature representations. These models showed variable generativity, with some standard autoencoders capable of near-perfect reconstruction of ~43 trillion images after training on just 2000. However, constrained β-VAE’s were unable to reconstruct images reflecting feature combinations which were systematically withheld during training (e.g. all blue circles). Thus, deep auto-encoders may provide a promising tool for understanding visual generativity and potentially other aspects of visual cognition.