Search Papers | Poster Sessions | All Posters
Poster B59 in Poster Session B - Thursday, August 8, 2024, 1:30 – 3:30 pm, Johnson Ice Rink
Cognitive abstractions for data-efficient learning, systematic generalization, and latent factor inference
Felipe Del Rio1,2 (), Eugenio Herrera2, Julio Hurtado4, Alvaro Soto1,2, Ali Hummos3, Cristian B. Calderon2; 1Universidad Catolica de Chile, 2Centro Nacional de Inteligencia Artificial, 3Massachusetts Institute of Technology, 4University of Warwick
Humans excel at generalization. Recent studies suggest that the latter emerges from systematic compositionality: The ability to adapt to (or understand) novel contexts based on the flexible recombination of previously learned (sub)concepts (or "cognitive abstractions"). Here, we propose a framework to study the effects of cognitive abstractions by leveraging standard generative neural networks. As predicted by empirical human studies, neural networks with cognitive abstractions, learn faster and generate systematic out-of-distribution (OOD). Moreover, we show that these cognitive abstractions can be used to infer the underlying latent factor that generate the images. Our framework can be used to analyze the representational basis that allow for the emergence of these properties.
Keywords: compositional learning systematic generalization auto-encoder gradient-based inference