The neural mechanisms of early categorization: a connectionist model of infant ERP

Westermann, G. 1 , Pauen, S. . 2 , Wahl, S. 2 & Hoehl, S. 2

1 Department of Psychology, Lancaster University, UK
2 Department of Psychology, Heidelberg University, Germany

An important question in infant categorization is how bottom-up perceptual knowledge interacts with top-down prior knowledge in forming early categories. To investigate this question, in a new category-oddball task infants were presented with a total of 100 objects in random order. Objects came from one category (animals or furniture) in 80% of cases (standard) and from the other category in 20% of cases (oddball). ERP was used to measure differences in cortical responses to standards and oddballs in 4- and 8-month-old infants. Results showed characteristic differences between standard and oddball ERPs in both age groups; however, in 4-month-olds this difference disappeared in the second half of the experiment.

We present a neural network model of infant ERP to evaluate different hypotheses of the mechanisms underlying these observed results. The model consists of two linked auto-encoders, one simulating fast, hippocampal processing and the other, slower cortical processing. Different age groups were modeled by training the networks on small or large amounts of ‘background knowledge’ prior to simulating the ERP experiment. Output error on the cortical part of the model was used to account for ERP data. Varying the amount of background knowledge and the representations of objects modeled four hypotheses about the interaction between bottom-up and top-down knowledge and the nature and robustness of the top-down knowledge.

Results of these simulations suggest that the ERP responses observed in the infants result from interactions between bottom-up and top-down processing even in the 4-month-old infants. However, whereas in younger infants the experience during the experimental session begins to override the not-yet stable background knowledge, in 8-month-olds background knowledge has become robust enough to dominate the category response throughout the experiment.

This is the first study to combine ERP with computational modeling to reveal the mechanisms underlying cortical responses in infant category learning.