Age of Acquisition in Connectionist Networks

Recently, there has been a resurgence of interest in the role of the Age of Acquisition (AoA) of an item in determining subjects’ reaction time in naming words, objects, and faces. Using the number of epochs required to learn an item as a direct measure of AoA in connectionist networks, Smith, Cottrell & Anderson (in press) have shown that AoA is a stronger predictor of final Sum Squared Error than frequency. In this paper, we replicate Smith et al. using more realistic frequency distributions for the items, and examine why some patterns may be learned earlier than others. First, we have found that the same patterns tend to be learned early and late by networks differing in their initial random weights; hence, the issue is, what property of the patterns determines AoA? We have found that even very weak pattern similarity structure is a strong predictor of AoA when frequency is controlled for. Also, we have found evidence that such a similarity structure may still be an important factor in determining AoA even when pattern frequency is varied.