Abstract: . . .  between them in proportion to the number of other wickelfeatures they account for uniquely (i.e., a “the rich gets richer” competitive 3 Wickelfeatures are generated in a similar way to wickelphones. The latter involves decomposing a phoneme strings into consecutive triples. Thus, the phoneme string /kæt/ ( cat ) is decomposed into the / kæ/, /kæt/, and /æt /. Notice that the triples are position-independent, but that the overall string can be pieced together again from the triples (in general, as Pinker & Prince (1988) have noted, this piecing together process cannot always be carried out successfully, but in this context it is adequate). Wickelfeatures correspond to triples of phonetic features rather than triples of entire phonemes. Page 18 CONNECTIONIST PSYCHOLINGUISTICS IN PERSPECTIVE 19 approach). The end result of this competition is a set of more or less non-overlapping wick- elphones which correspond to as many as possible of the wickelfeatures in the input to the decoder network. By employing a particular training regime, Rumelhart and McClelland were able to obtain the U-shaped learning profile, characteristic of children’s acquisition of the English past tense. First, the network was trained on a set of 10 high-frequency verbs (8 irregular and 2 regular) for 10 epochs. At this point the network reached a satisfactory performance, treating both regular and irregular verbs in the same way (as also observed in the first stage of human acquisition of past tense). Next, 420 . . .
. . .  proportion to the number of other wickelfeatures they account for uniquely (i.e., a “the rich gets richer” competitive 3 Wickelfeatures are generated in a similar way to wickelphones. The latter involves decomposing a phoneme strings into consecutive triples. Thus, the phoneme string /kæt/ ( cat ) is decomposed into the / kæ/, /kæt/, and /æt /. Notice that the triples are position-independent, but that the overall string can be pieced together again from the triples (in general, as Pinker & Prince (1988) have noted, this piecing together process cannot always be carried out successfully, but in this context it is adequate). Wickelfeatures correspond to triples of phonetic features rather than triples of entire phonemes. Page 18 CONNECTIONIST PSYCHOLINGUISTICS IN PERSPECTIVE 19 approach). The end result of this competition is a set of more or less non-overlapping wick- elphones which correspond to as many as possible of the wickelfeatures in the input to the decoder network. By employing a particular training regime, Rumelhart and McClelland were able to obtain the U-shaped learning profile, characteristic of children’s acquisition of the English past tense. First, the network was trained on a set of 10 high-frequency verbs (8 irregular and 2 regular) for 10 epochs. At this point the network reached a satisfactory performance, treating both regular and irregular verbs in the same way (as also observed in the first stage of human acquisition of past tense). Next, 420 medium-frequency . . .
--3000,2,750,3169,59892