190 likes | 452 Views
Bootstrapping. Tom Griffiths. Bootstrapping. How to learn words without knowing words Various proposals: “semantic bootstrapping” (Pinker, 1984) “syntactic bootstrapping” (Gleitman, 1990) Characterized by accelerated learning (e.g. Regier, 2004) Question:
E N D
Bootstrapping Tom Griffiths
Bootstrapping • How to learn words without knowing words • Various proposals: • “semantic bootstrapping” (Pinker, 1984) • “syntactic bootstrapping” (Gleitman, 1990) • Characterized by accelerated learning (e.g. Regier, 2004) • Question: • when is bootstrapping possible?
Word learning “blicket” “blicket” “blicket”
Likelihood Prior probability Posterior probability Sum over space of hypotheses Bayes’ theorem h: hypothesis d: data
Bayesian word learning (Tenenbaum, 1999; Tenenbaum & Xu, 2002) • Data • scene-word pairs • Hypotheses • functions labeling scenes • Likelihood • weak sampling • strong sampling x h w
“blicket” p(d|h) = 0
“blicket” p(d|h) = 1/3
“blicket” “blicket” “blicket” p(d|h) = (1/3)3
“blicket” p(d|h) = 1/12
“blicket” “blicket” “blicket” p(d|h) = (1/12)3
Bootstrapping • Bayesian word learning is a form of semantic bootstrapping (Niyogi, 2002) • What about accelerated learning? • non-linear* increase in probability of correct answer for a random scene and word • When can it occur? • not when hypotheses independent and all equally likely, when using weak sampling • speculation: hypotheses are dependent
Forms of dependency • Hierarchical priors • unknowns across learning events • Compositional priors • unknowns within learning events
Hierarchical priors x x x x h h h h w w w w “blicket” “toma” “dax” “wug”
“dax” “blicket” “toma” “wug”?
Hierarchical priors • What is contained in a hierarchical prior? • Any learned information that constrains scene-word mappings • typical referents (whole object) • dimensions of stimuli (shape/substance) • pragmatic dependencies (mutual exclusivity) • sound and meaning (morphology)
h G x h1 x h2 h1 x h2 w1 w2 w1 w2 w1 w2 holistic independent compositional Compositional hypotheses “blicket toma”
Compositional hypotheses • Good news: • express syntactic bootstrapping • model referential uncertainty • Bad news • requires complete linguistic theory
Bootstrapping • When do we see accelerated learning? • speculation: dependent hypotheses • Sources of dependency in language • hierarchical priors • compositional hypotheses • Bootstrapping goes beyond language • learning causal theories aids learn causal relationships, learning concepts…