The Role of Prior Knowledge in Human Reconstructive Memory Mark Steyvers Pernille Hemmer University of California, Irvine 1 Interaction between Prior Knowledge & Episodic Memory reconstructive memory Episodic Memory Prior Knowledge, Semantic Memory, “Schemas” “What objects do you remember from your hotel room?” “What coffee did you order from Starbucks last week?” 2 Research questions Machine learning Psychology How can we formalize prior knowledge (“schemas”)? How do humans integrate memory and prior knowledge? What errors do people make? Interface between human & machine learning Can we explain human error with rational inference principles? Can we build better information retrieval systems by taking cognitive processes into account? 3 Remembering Objects from a Graduate Office chair desk skull books (30% of subjects) Brewer & Treyens (1981) Experiments perception: list objects you see in a scene memory: list objects you remember from a scene guessing: list objects you might see in a kitchen/office/… Dining Hotel Kitchen 5 Human Data: Precision vs. Output Position 1 Precision Cumulative Accuracy 0.95 10 Secs. Time = 10secs Time = 2secs Scene Cue 0.9 0.85 0.8 2 Secs 0.75 0.7 0.65 0.6 0.55 2 4 6 8 10 12 14 16 Output Position 6 Bayesian Analysis of Reconstructive Memory p(objects | memory ) p(memory | objects ) p(objects ) Posterior Likelihood Prior What objects were studied given noisy memory contents? How likely is this memory content given these objects were studied? How likely are these objects a priori? 7 Problem 1: Infer prior knowledge Prior Knowledge kitchen1 kitchen2 ? kitchen3 Hierarchical Beta processes (Thibaux & Jordan, 2007); related to Indian Buffet process 8 Hierarchical Beta Process (Thibaux & Jordan, 2007) “Kitchen” General knowledge b b ~ Beta(c0b0 , c0 (1 b0 )) Object probs. for specific image aj Reported objects participants a j ~ Beta( c j b, c j (1 b)) y yij ~ Bernoulli( a j ) objects Kitchen65 Kitchen45 9 Problem 2: Recall using Prior + Episodic M. Prior Knowledge b semantic memory Actual objects in image y ? episodic memory Noisy Memory Content x 10 Data Model 1 1 Time = 10secs Time = 2secs 0.95 Scene Cue 10 Secs. Cumulative Accuracy Precision Cumulative Accuracy 0.95 0.9 0.85 0.8 2 Secs 0.75 0.7 0.65 0.6 0.55 2 4 6 8 10 12 Output Position 14 16 0.9 0.85 0.8 0.75 0.7 0.65 =0.200 0.6 =0.150 =0.000 0.55 2 4 prior only 6 8 10 12 14 16 Output Position 11 Limitations of Model Independence assumption between objects Recall often clusters correlated objects “computer” “mouse” “keyboard” 12 Reconstructing the Style of Written Digits Study Digit space from “Earth Mover” distances personal prior Reconstructed reconstructed studied digit 13 Reconstructing Drawings from Memory Studied Drawing Reconstruction 14 Reconstructing Lists of Words Study this list: PEAS, CARROTS, BEANS, SPINACH, LETTUCE, HAMMER, TOMATOES, CORN, CABBAGE, SQUASH HAMMER, PEAS, CARROTS, ... 15 Conclusion Prior knowledge has strong effect on memory Guessing based on prior knowledge leads to quite good performance Machine learning methods help formalize “schemas” Useful to understand human memory when designing information retrieval systems 16 Thanks! 17