Age of acquisition effects in adult lexical processing reflect loss of plasticity in maturing systems: Insights from connectionist networks.

Abstract

Early learned words are recognized and produced faster than later learned words. The authors showed that such age of acquisition effects are a natural property of connectionist models trained by backpropagation when patterns are introduced at different points into training and learning of early and late patterns is cumulative and interleaved. Analysis of hidden unit activations indicated that the age of acquisition effect reflects a gradual reduction in network plasticity and a consequent failure to differentiate late items as effectively as early ones. Further simulations examined the effects of vocabulary size, learning rate, sparseness of coding, use of a modified learning algorithm, loss of early items, acquisition of very late items, and lesioning the network. The relationship between age of acquisition and word frequency was explored, including analyses of how the relative influence of these factors is modulated by introducing weight decay.

Salvaging a childhood language

“…Its logic is as follows: A memory network organizes itself, through larger weight changes, into a certain efficient configuration early on to represent what it knows about (e.g., a childhood language); it cannot nimbly reorganize itself even if the nature of the input changes (e.g., a second language). Thus early-learned materials will have an advantage and lasting impact on the memory network (e.g., Ellis & Lambon-Ralph, 2000;Zevin & Seidenberg, 2004). While our study was not set up to evaluate this account, our findings nonetheless need to be reckoned with.…”

What does it take to evolve behaviorally complex organisms?

“…When it is time to acquire the more difficult What task, however, it turns out to be impossible to reach perfect performance in the later acquired What task. Many simulations using the backpropagation algorithm have shown that neural networks exhibit what is sometimes called the "age of acquisition" effect: at the end of the simulation performance on tasks that are learned earlier than other tasks is better than performance on later acquired tasks (Ellis & Lambon Ralph, 2000;Smith et al, in press). This "age of acquisition" effect is observed also in our simulations in which populations of neural networks evolve the ability to solve the What and Where tasks, as can be seen by plotting separately the fitness curves for the Where and the What tasks in the modular architecture ( Figure 7).…”

“…The financial support of the Italian National Research Council (Short Term Mobility Program 2000 to RC and Bilateral Project 1998-2000 and by the Yale Institute for Biospheric Studies to GPW is gratefully acknowledged. The authors thank the referees for valuable suggestions on the manuscript.…”

Developmental cognitive neuroscience: progress and potential

“…[41,42]). These models allow researchers complete control over simulated learning systems and their environments, to help identify factors contributing to enhanced learning during particular points in development.…”

“…Such simulations have demonstrated that entrenchment can occur as some units and connections become committed and unused connections are pruned [41,42]. The result is that stimuli encountered early are learned more robustly than stimuli encountered later.…”