Poster B89, Sunday, March 26, 8:00 – 10:00 am, Pacific Concourse
Neural Responses Decrease While Performance Increases with Practice: A Neural Network Model
Milena Rabovsky1, Steven S. Hansen2, James L. McClelland2; 1Freie Universitaet Berlin, Germany, 2Stanford University
The observation that neural responses decrease as behavior becomes faster and more accurate with practice is ubiquitous in neuroscience. Our work provides an account of this finding in terms of a shift in the relative roles of activation and strength of connections. We used a neural network model to simulate neural responses during language understanding, and examined the model’s correlate of neural responses (specifically, the N400 component of the event-related brain potential), measured as the change in hidden layer activation induced by the current stimulus, at several time points during training of the network. We observed that the N400 magnitude first increased and then gradually decreased over the course of training while comprehension performance measured at the output layer showed a steady rise with additional practice. These results fit the empirical finding that N400 amplitudes first increase over the first few years of life and later decrease with age. Importantly, our results also speak to the issue of possible mechanisms underlying the reduction of neural activation with practice. In the model, the reduction in neural response is due to the continuous adaptation of connection weights over training. Specifically, as connection weights between the hidden and the output layer grow stronger, less activation at the hidden layer is necessary to efficiently modulate the output. This shift of labor from activation to connection weights might be an important mechanism contributing to the often observed reduction of neural activation with practice.
Topic Area: LONG-TERM MEMORY: Development & aging