Retracing the Rational Analysis of Memory Justin Li Computer - PowerPoint PPT Presentation

Retracing the Rational Analysis of Memory Justin Li Computer Science and Engineering University of Michigan justinnh@umich.edu 2014-06-19

Introduction Problem Models Summary , What is this talk about? Goal: ◮ (re)examine and formalize the goal of memory mechanisms ◮ unify mechanisms such as cued and spontaneous retrieval, working and semantic memory activation, etc. 2014-06-19 Li. Retracing the Rational Analysis of Memory 2

Introduction Problem Models Summary , The Rational Analysis of Memory Anderson (1990) performed a rational analysis of memory: Goal Environment Constraints Optimization 2014-06-19 Li. Retracing the Rational Analysis of Memory 3

Introduction Problem Models Summary , The Rational Analysis of Memory Anderson (1990) performed a rational analysis of memory: Goal provide the agent with knowledge it is most likely to need Environment Constraints Optimization 2014-06-19 Li. Retracing the Rational Analysis of Memory 3

Introduction Problem Models Summary , The Rational Analysis of Memory Anderson (1990) performed a rational analysis of memory: Goal provide the agent with knowledge it is most likely to need Environment one where probability of need is a function of recency and frequency Constraints Optimization 2014-06-19 Li. Retracing the Rational Analysis of Memory 3

Introduction Problem Models Summary , The Rational Analysis of Memory Anderson (1990) performed a rational analysis of memory: Goal provide the agent with knowledge it is most likely to need Environment one where probability of need is a function of recency and frequency Constraints memories are accessed sequentially at fixed cost Optimization 2014-06-19 Li. Retracing the Rational Analysis of Memory 3

Introduction Problem Models Summary , The Rational Analysis of Memory Anderson (1990) performed a rational analysis of memory: Goal provide the agent with knowledge it is most likely to need Environment one where probability of need is a function of recency and frequency Constraints memories are accessed sequentially at fixed cost Optimization stop retrieval when cost > probability of need ∗ gain 2014-06-19 Li. Retracing the Rational Analysis of Memory 3

Introduction Problem Models Summary , Bayesian Memory Goal: return element m ∈ M with the highest probability of need P ( m ) Given: set of context elements C ⊂ M Find: P ( m ) P ( C | m ) arg max P ( m | C ) = arg max P ( C ) m ∈ M m ∈ M = arg max P ( m ) P ( C | m ) m ∈ M 2014-06-19 Li. Retracing the Rational Analysis of Memory 4

Introduction Problem Models Summary , Bayesian Memory arg max P ( m ) P ( C | m ) m ∈ M What does this mean? P ( m ) probability of need of element m (ie. the prior ) P ( C | m ) probability of need of the context C given that m is needed (ie. the likelihood ) 2014-06-19 Li. Retracing the Rational Analysis of Memory 5

Introduction Problem Models Summary , ACT-R’s Memory Mechanisms ◮ Cued Retrieval ◮ Partial Match ◮ Spreading Activation 2014-06-19 Li. Retracing the Rational Analysis of Memory 6

Introduction Problem Models Summary , Cued Retrieval Assuming the context C is the set of cues: P ( C | m ) P ( m ) arg max m ∈ M 2014-06-19 Li. Retracing the Rational Analysis of Memory 7

Introduction Problem Models Summary , Cued Retrieval Symbolic Long Term Memories Semantic Episodic Procedural Episodic Reinforce- Semantic Chunking Learning Learning ment Symbolic Short-Term Memory Procedure l a r Decision o s t i a c e r p t e p D A Clustering LT Visual Memory Action Perception ST Visual Memory Body 2014-06-19 Li. Retracing the Rational Analysis of Memory 8

Introduction Problem Models Summary , Cued Retrieval Symbolic Long Term Memories Semantic Episodic Procedural Episodic Reinforce- Semantic Chunking Learning Learning ment Symbolic Short-Term Memory Procedure l a r Decision o s t i a c e r p t e p D A Clustering LT Visual Memory Action Perception ST Visual Memory Body 2014-06-19 Li. Retracing the Rational Analysis of Memory 8

Introduction Problem Models Summary , Cued Retrieval Assuming the context C is the set of cues: P ( C | m ) P ( m ) arg max m ∈ M We want ∀ m , P ( C | m 1 ) = P ( C | m 2 ) Take  1 , if ∀ c ∈ C is a child of m   P ( c | m ) =   0 , otherwise   2014-06-19 Li. Retracing the Rational Analysis of Memory 9

Introduction Problem Models Summary , Partial Match Assuming the context C is the set of cues: P ( C | m ) P ( m ) arg max m ∈ M We want P ( C | m ) to be: ◮ proportional to the number of c ∈ C that is a child of m ◮ inversely proportional the number of children that m has 2014-06-19 Li. Retracing the Rational Analysis of Memory 10

Introduction Problem Models Summary , Spreading Activation Assuming the context C is the working memory: arg max P ( C | m ) P ( m ) m ∈ M 2014-06-19 Li. Retracing the Rational Analysis of Memory 11

Introduction Problem Models Summary , Spreading Activation Symbolic Long Term Memories Procedural Semantic Episodic Semantic Episodic Reinforce- Chunking Learning Learning ment Symbolic Short-Term Memory Procedure l Decision a r o s t i a c r e p t e p D A Clustering LT Visual Memory Perception ST Visual Memory Action Body 2014-06-19 Li. Retracing the Rational Analysis of Memory 12

Introduction Problem Models Summary , Spreading Activation Assuming the context C is the working memory: arg max P ( C | m ) P ( m ) m ∈ M Note there is no cue – this model could also spontaneous 2014-06-19 Li. Retracing the Rational Analysis of Memory 13

Introduction Problem Models Summary , Bayesian Networks Problems: 2014-06-19 Li. Retracing the Rational Analysis of Memory 14

Introduction Problem Models Summary , Bayesian Networks Problems: ◮ What is P ( m ) ? ◮ in ACT-R, base-level activation is ln ( P ( m )) ◮ other options? ◮ working memory activation or semantic memory activation? 2014-06-19 Li. Retracing the Rational Analysis of Memory 14

Introduction Problem Models Summary , Bayesian Networks Problems: ◮ What is P ( m ) ? ◮ in ACT-R, base-level activation is ln ( P ( m )) ◮ other options? ◮ working memory activation or semantic memory activation? ◮ What is P ( C | m ) ? ◮ in a Bayes net, all external factors ◮ inference is NP-hard ◮ semantic networks are not Bayesian networks (ie. acyclic) 2014-06-19 Li. Retracing the Rational Analysis of Memory 14

Introduction Problem Models Summary , Nuggets and Coal Nuggets Coal ◮ Memory retrieval can be ◮ Bayesian inference fails on cast in a Bayesian semantic networks framework ◮ Additional assumptions ◮ This framework provides needed to make inference explanations for multiple tractable and correct memory mechanisms 2014-06-19 Li. Retracing the Rational Analysis of Memory 15

Introduction Problem Models Summary , Questions? Symbolic Long Term Memories Procedural Semantic Episodic Semantic Episodic Reinforce- Chunking Learning Learning ment Symbolic Short-Term Memory Procedure l r Decision a o s i t c a r e p t e p D A Clustering LT Visual Memory Perception ST Visual Memory Action Body 2014-06-19 Li. Retracing the Rational Analysis of Memory 16