Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Download Towards General AI: What we can learn from Human Learning
Document related concepts
Transcript
Towards General AI: What we can learn from Human Learning Nicole Beckage & Dongkyu Choi University of Kansas Child Language as a case study Study one of the earliest cognitively intense tasks that young children quickly master Knowledge Intelligent and Invisible Computing 3 Longitudinal Data: Experimental psychology lab collected longitudinal data Followed a specific child for 12 consecutive months. 114 children with longitudinal CDIs. Around 40000 learning examples Intelligent and Invisible Computing 4 Network Growth modeling Embed cognitive theory into network growth algorithms Algorithmic assumptions formalize cognitive, learning and attentional mechanisms window head noodles fork balloon plate giraffe cake house trash moon dog tree bird star face fish sheep horse cat water coat snow bottle slide donkey telephone pajamas bunny leg pizza tomorrow jeans deer rain window duck trash door fork plate butter penguin puppy house bird home crib egg goose tree airplane cake cheese noodles pizza bunny rain snow water slide bug coat fish garden jeans tomorrow turtle cow pony bug face bottle moon telephone balloon squirrel hair toe turtle pajamas star child crib home giraffe candy girl child boots cow sheep pony horse dog cat shoe donkey monkey tiger squirrel owl diaper leg sister candy Intelligent and Invisible Computing 5 window noodles candy fork balloon Edges: Predefined plate relationship between words. Here, free home associations. Nodes: words that the child produces. child crib giraffe cake squirrel house trash moon dog tree bird star face fish sheep horse cat water coat turtle cow snow bottle pony slide donkey telephone pajamas bunny leg bug pizza tomorrow jeans rain window noodles candy fork balloon plate child crib home giraffe cake squirrel house trash moon dog tree bird star face fish sheep horse cat water coat turtle cow snow bottle pony bug rain slide donkey telephone pajamas bunny leg pizza jeans tomorrow Age 18mo Intelligent and Invisible Computing 7 window head face telephone coat turtle fish rain candy bottle boots giraffe moon horse cow dog leg jeans tomorrow trash pony donkey cake puppy bug child snow water egg slide noodles balloon bunny house sheep cat bird crib pizza pajamas fork plate star home tree airplane squirrel Age 20mo Intelligent and Invisible Computing 8 head giraffe toe turtle pajamas star deer face bottle moon telephone balloon window duck trash door fork plate butter penguin puppy house bird home crib egg goose tree airplane cake cheese noodles pizza bunny rain snow water slide bug coat fish garden jeans tomorrow hair girl child boots cow sheep pony horse dog cat shoe donkey monkey tiger squirrel owl diaper leg sister candy Age 22mo Intelligent and Invisible Computing 9 Model A: Preferential Attachment Attractor model on subset of known words. Known words with a high number of edges are attractors. Unknown words are learned if they attach to highly connected words. Subset of known words (the child’s current vocabulary knowledge) plays largest role in acquisition. Intelligent and Invisible Computing 10 Model A: Preferential Attachment fish mouth tongue cow juice grandpa water bus keys truck book car park rock bicycle outside box cookie light airplane train garage motorcycle grandma monkey Attractor model on subset of known words. nose eye brother bath cracker hat bee block shoe hair Unknown words are learned if they attach to highly connected words. Subset of known words (the child’s current vocabulary knowledge) plays largest role in acquisition. foot sock head ball Known words with a high number of edges are attractors. dog Intelligent and Invisible Computing 11 Model B: Preferential Acquisition Attractor model on subset of unknown words. Unknown words with a high number of connections are attractors. Unknown words are learned if they are highly connected in the learning environment. Independent of the words a child knows; language environment drives learning. Intelligent and Invisible Computing 12 Model B: Preferential Acquisition Attractor model on subset of unknown words. Unknown words with a high number of connections are attractors. Unknown words are learned if they are highly connected in the learning environment. Independent of the words a child knows; language environment drives learning. Intelligent and Invisible Computing 13 Experimental setup: Predicting acquisition trajectories Is semantic relationships or phonological relationships more important to early language acquisition? Compare predictive accuracy assuming free association norms Or phonological similarity network. Run network growth model assuming each of the growth processes. Preferential Attachment (current vocabulary knowledge) Preferential Acquisition (structure of the language environment) Predict the words to be learned next. Performance informs our understanding of acquisition Intelligent and Invisible Computing 14 Results: Predictive modeling Proportion of children best fit 80 70 60 50 40 30 20 10 0 Pref Att Pref Acq semantics Lure phono Best fitting model for semantics is preferential acquisition. Best fitting model for phonology is preferential attachment. Intelligent and Invisible Computing 15 Results: Predictive modeling Proportion of children best fit 80 70 60 50 40 30 20 10 0 Pref Att Pref Acq semantics Lure phono The dominant mechanism for semantic learning is connectivity in the learning environment. The dominant mechanism for phonological learning is connectivity of known vocabulary words. Intelligent and Invisible Computing 16 Moving beyond incremental insights into robust learning. We can study learning in young children. But this is very challenging and requires complex models How can we use our understanding of human intelligence to build robust cognitive systems? Cognitive Architectures: ICARUS Characteristics: Long and short term memories Concepts, problems, and goals Hierarchical knowledge structure Goal reasoning (goal driven but reactive) Intelligent and Invisible Computing 18 ICARUS operation Operates in cycles 1) Receive sensory input 2) Infer beliefs about current situation 3) Nominates top-level goals relevant to current situation 4) Evaluates skills to achieve the top-level goals 5) If a skill ‘path’ exists, execute the implied actions 6) Otherwise, invoke problem solver Intelligent and Invisible Computing 19 Perceptual Buffer Long-term Conceptual Memory Inference Perception Belief Memory Long-term Goal Memory Goal Reasoning Skill Retrieval Environment Long-term Skill Memory Skill Learning Short-term Goal Memory Motor Buffer Execution Intelligent and Invisible Computing 20 The future of AI Integration of cognitive science and computer science Models of AI requires expert knowledge of humans Execution of AI requires expert knowledge of computation Use insights of human acquisition, to extend ICARUS Exploration robust learning and generalization Development within Malmo platform Intelligent and Invisible Computing 21 Thanks! Email contacts: Nicole Beckage [email protected] Dongkyu Choi [email protected] Intelligent and Invisible Computing 22 Intelligent and Invisible Computing 23 Intelligent and Invisible Computing 24 Focus on predicting words learned next What are the cognitive mechanisms and developmental processes that support future language learning? Can we construct computational models that predict learning of individual words for individual learners? Intelligent and Invisible Computing 28 Skill evaluation Belief inference top-level goals higher-level (non-primitive) concepts skill hierarchy primitive concepts direct actions sensory input Intelligent and Invisible Computing 29 ICARUS: recent extensions Numeric processing Ability to perform numeric calculations in concepts and skills Ability to plan with numeric effects Compositional beliefs Can create new objects with numeric attributes Percepts-beliefs unification Unified memory for sensory input and beliefs Episodic memory New memory and processing module for storing and retrieving previous events Intelligent and Invisible Computing 30 Intelligent and Invisible Computing 31 General Artificial Intelligence?
