Wickelphones

• Pro: capture just enough info about the context that determines irregular past-tense Verb forms ense verbs, e.g.sing -sang, ring -rang
• 1 wickelphone = 1 input unit and 1 output unit, connection matrix of 42875 * 42875!!
• Instead, reduce wickelphones to wickelfeatures (1210), where each wickelphone becomes 16 Features
• Verbs with same stem and past tense
• English has many verbs where the past and stem are the same:
  • put, fit, spread
• Usually verbs ending with -t or -d are likely have no-change
• Model also started to let the past be the same as the stem
• Verbs with vowel change
• Model made two errors older children have been shown to make: 1. stem + ed, e.g. comed, singed 2. past-form + ed, e.g. camed, sanged
• Does not explain differences in response times between irregular and regular verbs
• Models production only, not comprehension
• You can’t reverse the model like you can reverse a rule
• Model can’t generalize
• Computational models (Rule-based, Connectionist and MBL) all use adult vocabulary as input to simulate children’s learning
• Children don’t show vocabulary burst between State 2 and 3 (needed to produce U-shaped learning with Connectionist Model)
• Thousands of exposures
• No distinction between tokens and types, each Verb simply included once