Three challenges for computational models of cognition Charles Kemp CMU
Humans vs machines Performance Outstanding Not so good Machine Human First order of business is to close this gap: Nagging worry for non-Bayesian models: maybe the model does badly because the learning component isn’t optimal. We have a good theoretical understanding of inductive inference. Bayesian approach puts you in a good position to explore what’s left.
✓ Three challenges ✓ Composition Generativity Putting it all together Structured Models Neural network/ continuous space models ✓ ✓ Composition Generativity Putting it all together ✓ ✓ ✓ ✓
Composition: sentences Given a database of geography facts, answer questions like: “how many rivers run through the states bordering Colorado?” “how many states border the state that borders the most states?” (Mooney, 1997)
Liang et al, Learning dependency based compositional semantics “A major focus of this work is our semantic representation, DCS, which offers a new perspective on compositional semantics.”
Socher et al, Semantic compositionality through recursive matrix vector spaces
Opportunities/Challenges Compositional systems that work with fuzzy concepts.
Generativity “Mr. and Mrs. Dursley, of number four Privet Drive, were proud to say that they were perfectly normal, thank you very much.”
Computational models (Hofstadter et al, Letter Spirit) (Cohen, AARON)
Hinton et al, A fast learning algorithm for deep belief nets
Jern & Kemp, A probabilistic account of exemplar and category generation Training: X X D Z N Q J Q J … M M B Test: Generate another Z N D B Human Model
Fleuret et al, Synthetic Visual Reasoning Test Category 1 Category 2
Opportunities/Challenges Compositional systems that work with fuzzy concepts. Avoid “cargo cult” science via benchmark engineering.
One problem, many settings (Salakhutdinov, Tenenbaum, Torralba) Psychological data: categorization (Canini et al) causal learning (Kemp et al)
One setting, many problems Generalization, Categorization, Identification, Recognition … (Shepard; Nosofsky; Ashby; Kemp & Jern…)
Many settings, many problems Cognitive architectures (ACT-R, SOAR) Artificial general intelligence
Opportunities/Challenges Compositional systems that work with fuzzy concepts. Avoid “cargo cult” science via benchmark engineering Systems that solve many different problems in many different settings
✓ Three challenges ✓ Composition Generativity Putting it all together Structured Models Neural network/ continuous space models ✓ ✓ Composition Generativity Putting it all together ✓ ✓ ✓ ✓