Mirman Et Al.¶

• train syllables in words, predicting the next syllable
• use network to train on different types of individual words, matching them with one of five objects, simulating word learning
• 75 epocs 1000 syllable sequence, then it predicted almost perfectly the next syllable (teaching phonotactics of the language)
• Model trained to recognize one of five objects for each of five different two-syllable input patterns of three types 1. words (100% transitional probability) 2. partwords (25% probability transitions) 3. nonwords (0% transitions)
• Model is better at mapping two-syllable sequences to words when it has already been exposed to those sequences and they had high probabilities
• Novel-sequence non-word labels initially learned nearly as fast as word up to intermediate point.
• exposure to familiarization input allowed network to created distinct hidden representations for each syllable
• SRN can show how statistical learning supports word learning, showing a link
• Humans are good at learning sequences, even when the data is presented implicitly and even when the relationships are non-adjacent
• We aren't just sensitive to frequency: we are sensitive to actual transitional probabilities
• SRNs with very simple assumptions model non-adjacent learning and transitional probabilities
• Biological arguments for distributed representations
• Makes more sense that neurons get randomly assigned to be active for different inputs
• We can start with randomness and with learning it will become structured
• concepts are just bundles of features, that together become something
• Prevents catastrophic failures