Online Processing is Essential for Learning: Understanding Fast Mapping and Word Learning in a Dynamic Connectionist Architecture

摘要: Online Processing is Essential for Learning: Understanding Fast Mapping and Word Learning in a Dynamic Connectionist Architecture Jessica S. Horst (jessica-horst@uiowa.edu) Bob McMurray (bob-mcmurray@uiowa.edu) Larissa K. Samuelson (larissa-samuelson@uiowa.edu) Department of Psychology, E11 Seashore Hall Iowa City, IA 52242 USA Abstract independent processes (Horst & Samuelson, submitted), the present computational model suggests subtle integration. Specifically, confirms that fast mapping word learning represent two distinct, but related, time scales can, fact, lead to learning. Thus, we argue are related not coextensive, as previously argued literature. In making initial mapping, child faced with very specific problem: given series objects name, must determine which object should be associated name. other words, referent To solve this problem, engages probabilistic constraint satisfaction. That is, determines most probable optimal solution constraints. case constraints include both input, name presented, child’s own developmental history, includes vocabulary known names. problem solving required done real-time likely (object) (unknown referent). Further, aspect does necessarily require on part child. However, will show, repetition these dynamics change next novel presented eventually, enough repetitions, can process quickly emerges moment, while encoding robust representation link unfolds over longer scale (for similar argument see also Capone McGregor, press; Carey, 1978). Recent empirical data underscores hypothesized distinction between The acquisition meaning often partially attributed mapping. recent research may distinct components language acquisition. Here examine their interaction Hebbian Normalized Recurrence Network, connectionist architecture captures online processing long-term statistical After training small lexicon, performed well above chance task. Careful analyses weight changes, however, suggest task solved minimal only processes, provides unique insights regarding relationship carefully distinguished. early childhood, children learn words at impressive rates; typically saying first 10-14- months-of-age, around 300 by second birthday 60,000 fifth (Carey, Young problems remarkable ease. When familiar objects, one correctly pick 1978; Carey Bartlett, This ability, “fast mapping” cited evidence children’s proficiency. findings although important learning, selecting response coextensive literature, discussed if they same thing (see, example, Behrend, Scofield, Kleinknecht, 2001). We argue, though subtly recognition First, select moment (fast mapping). Second, encode name-object such it recalled after delay (complete learning). extant literature reinforces studies either test ability make or remember mappings delay, some additional review submitted). While have potentially logically Supporting Empirical Data (submitted) found twenty-four- month-old were able map many eight new single session, did retain five minute delay. On each trial, object. half trials asked get (e.g., “can you car?”), asked, blicket?”). Overall, excellent referents (see Figure 1A). Five minutes later tested retention fast-