Modeling artificial category learning from pixels: Revisiting Shepard, Hovland, and Jenkins (1961) with deep neural networks

Abstract

Recent work has paired classic category learning models with convolutional neural networks (CNNs), allowing researchers to study categorization behavior from raw image inputs. However, this research typically uses naturalistic images, which assess participant responses to existing categories; yet, much of traditional category learning research has focused on using novel, artificial stimuli to examine the learning process behind how people acquire categories. In this work, we pair a CNN with ALCOVE (Kruschke, 1992), a well-known exemplar model of categorization, and attempt to examine whether this model can reproduce the classic type ordering effect from Shepard, Hovland, and Jenkins (1961) on raw images rather than abstract features. We examine this question with a variety of CNN architectures and image datasets and compare ALCOVE-CNN to two other models that lacked certain key features of ALCOVE. We found that our ALCOVE-CNN model could reproduce the type ordering effect more often than the other models we tested, but in limited situations. Our results showed that success varied greatly across the various configurations we tested, suggesting that the feature representations from CNNs provide strong constraints in properly capturing this effect.

A controller-peripheral architecture and costly energy principle for learning

“…Specifically, Type I was the easiest to master, followed by Type II, followed by Types III–V, and Type VI was the hardest. ( 53 ) is a challenging human category learning dataset to fit and has proven difficult for models that take images as inputs ( 27 ).…”

“…Given the complementary roles cognitive and DNN models play in capturing cognition and perception, one obvious path to integration is using the outputs of DNN models as inputs to cognitive models (e.g., (26)(27)(28)). Although appealing straightforward, this approach does not address how different cognitive processes and their underlying brain regions interact to create intelligent behavior.…”

Human visual cortex and deep convolutional neural network care deeply about object background

“…The criticism towards DCNNs become pointed as studies revealed a divergence between humans and DCNNs categorization strategies -humans and DCNNs make mistakes on different images (16)(17)(18), DCNNs have an inherent texture bias while humans have an inherent shape bias (19)(20)(21)(22), and DCNNs are susceptible to adversarial attacks imperceptible to humans (23,24). While these studies point to differences in categorization strategies, they do not negate the fact that DCNNs can still produce representations which align with human visual processing (25), as reflected in its high predictive performance of brain dynamics.…”