Technical Program

Paper Detail

Paper: PS-1A.11
Session: Poster Session 1A
Location: Symphony/Overture
Session Time: Thursday, September 6, 16:30 - 18:30
Presentation Time:Thursday, September 6, 16:30 - 18:30
Presentation: Poster
Paper Title: Predicting human perceived similarity across a wide range of object categories via sparse positive embeddings
Manuscript:  Click here to view manuscript
Authors: Martin Hebart, Laboratory of Brain & Cognition, National Institute of Mental Health, United States; Charles Zheng, Francisco Pereira, Section on Functional Imaging Methods, National Institute of Mental Health, United States; Chris Baker, Laboratory of Brain & Cognition, National Institute of Mental Health, United States
Abstract: How do humans represent behaviorally-relevant dimensions of real-world objects? To address this question, we recently used a triplet odd-one-out task to collect >800,000 behavioral judgments on images of 1,854 diverse basic-level object categories. To explain human behavior and characterize the similarity between pairs of objects, we developed a simple cognitive model that yielded sparse, interpretable perceptual and conceptual dimensions. To determine the utility of those dimensions, here we investigate two questions. First, to what degree can we predict those dimensions from a semantic embedding (Pilehvar & Collier, 2016) and activations in a deep convolutional neural network (CNN)? Second, can we use those predicted dimensions to reconstruct human behavioral similarity? To address these questions, we applied Ridge and Elastic Net regression to semantic embeddings and the activations in fully-connected layer 7 of the CNN VGG-16, respectively. We related the performance to two baseline models: The computational models alone, and a recently proposed method that transforms model features (Peterson et al., 2016). Our results demonstrate excellent prediction of many dimensions and strongly improved predictions of behavioral similarity using our model as compared to both baseline models. These results represent an important step towards both predictive and interpretable models of human cognitive representations.