An ART neural network model of discrimination learning

In a study of discrimination learning (DL) in human participants aged 4-20, two distinct learning modes were found by modeling the trial-by-trial learning process: (1) a discontinuous rational learning process by means of hypothesis testing; and (2) a slow, yet discontinuous learning process [1]. A neural network (adapted from [2]) was developed that implements attention-guided learning by selective sensory processing based on dimensional preferences mediated through reinforcement. The network is able to model developmental differences in DL by reducing the influence of negative reinforcement when modeling children's performance. Statistical analysis by fitting Markov models on the networks response sequences shows distinct modes of learning. The existence of two distinct learning modes is a consequence of uniformly distributed dimensional preferences in interaction with negative reinforcement. Networks that model childlike performance show both fast, rational and slow, discontinuous learning. Networks that model adult performance show fast learning only.