Explanation & learning slides (talk @ pittsburgh science of learning center)
Download as PPTX, PDF1 like333 views
- Explaining "why" helps learning by engaging learners and promoting the discovery of generalizations that connect facts and observations. - A "subsumptive constraints" account holds that explaining favors broad generalizations suggested by prior knowledge to apply beyond specific cases. - Explaining can impair learning when it leads learners to rely on unreliable generalizations but can also promote belief revision in response to anomalous evidence that contradicts beliefs.
Explanation & learning slides (talk @ pittsburgh science of learning center)
- 1. Why does explaining “why?” help learning? A subsumptive constraints account Joseph Jay Williams Joseph_williams@berkeley.edu Josephjaywilliams.com
- 2. Explanation
- 3. Explanation & Learning • Education: Biology, physics, math. (Chi et al, 1994; 1989; Nokes et al, 2011; Siegler, 2002; Renkl, 1997) • Cognitive Development: Conceptual change in theory of mind and number conservation. (Siegler, 1985; Amsterlaw & Wellman, 2006; Wellman & Liu, 2006) • Cognitive Psychology: Category learning, causal reasoning, property induction. (Ahn & Kalish, 2000; Murphy, 2002; Rehder, 2006; Williams & Lombrozo, 2010) • Artificial Intelligence (Mitchell & Cedar-Kabelli, 1986; DeJong, 2008) • Philosophy of Science (Salmon, 1990; Woodward, 2010)
- 4. Explanation • “Why?” Explanation: Why a fact is true • Explaining the meaning of a text passage • Explaining one’s reasoning • Explaining how a conclusion was reached • Explaining how something works • Explaining what is anticipated • Step-focused, Gap-filling, Mental-model revising (Nokes et al, 2011) 4
- 5. Explaining explanation • General effects: • Learning Engagement (e.g. Siegler, 2002) • Metacognitive monitoring (e.g. Chi, 2000) • Selective effects: • Subsumptive Constraints account (Williams & Lombrozo, 2010, Cognitive Science) Explaining engages search for underlying generalizations
- 6. Subsumptive Constraints • Subsuming: An explanation of why a fact or observation is true shows how it is subsumed as an instance of a generalization • Unifying: Better explanations have broader scope Why?
- 7. Subsumptive Constraints • Subsumption: An explanation of why a fact or observation is true shows that it might be expected as an instance of a generalization • Ex. 1: Observation: John is a teacher. – Generalization: Caring people tend to become teachers. – Subsumed: John is an instance of a caring person being a teacher. • Ex. 2: Giraffes have long necks. – Generalization: Wanting a goal can make animals grow. 7
- 8. Predictions Subsumptive constraint on explanation: 1. Promotes discovery of generalizations 2. Favors broad generalizations: which prior knowledge suggests apply beyond specific cases 3. Impairs learning when generalizations are unreliable 4. Promotes belief revision from anomalies 8
- 9. I. Discovery in Category Learning • Learn to categorize robots on the planet ZARN. 9
- 10. Body Shape Generalization 10
- 11. Broad Foot Shape Generalization 11
- 12. Design Overview EXPLAIN: Explain why this Explain might be a Glorp. vs Control DESCRIBE: THINK ALOUD: FREE STUDY Describe this glorp. Say out loud what you are thinking. Categorization & MemoryTest Differences between categories? 12
- 13. Discovery of foot shape generalization (Williams & Lombrozo, 2010) Proportion discovering 0.5 Explain generalization 0.4 Control 0.3 0.2 0.1 0 Describe Think Aloud Free Study 13
- 14. II. Subsumptive scope & prior knowledge Antenna Generalization Foot Generalization Blank Informative Labels Labels Glorp Outdoor Drent Indoor 14
- 15. Discovery of knowledge-relevant generalization 1.0 Proportion discovering Explain knowledge-relevant 0.8 generalization Free Study 0.6 0.4 0.2 0.0 Blank labels Low PK Informative labels High PK 15 N = 407
- 16. Children’s sensitivity to subsumptive scope Drives Conceptual Frequent & Sophisticated Change (Chouinard, 2008; (Siegler, 2005; Hickling & Wellman & Liu, 2007) Wellman, 2001) • Less developed than adults • Explaining a guide to favor subsumptive scope? – In breadth of observations – In using prior knowledge
- 17. Learning about a novel cause Blicket Detector
- 18. Explanation Prompt GO Explain: “Why did 100% Pattern: Green = GO this one make my 75% Pattern: machine play RED = GO music?” Control: “What happened to my machine when I 100% Pattern: put this one on?” Yellow = NO GO 18
- 19. 100% vs. 75% Pattern 1 favoring 100% Pattern 0.8 * Proportion Choices 0.6 Verbalize 0.4 Explain 0.2 0 PK matched PK favors 75% 19
- 20. PK favors 75% pattern PK favors 75% pattern: GO Big blocks make it go 100% Pattern: Green = GO 75% Pattern: Explain: “Why did = GO this one make my machine play music?” Control: “What happened to my 100% Pattern: Yellow = NO GO machine when I put this one on?” 100% Pattern: = NO GO 20
- 21. Effect of prior knowledge 1 favoring 100% Pattern 0.8 * Proportion Choices 0.6 0.4 Verbalize Explain 0.2 0 PK matched PK favors 75% 21
- 22. Hazards of explanation Malle, 2011
- 23. Explanation’s benefits and hazards Explain Reliable patterns Spurious or misleading patterns Learning Engagement: Motivation, Attention, Enhancement Enhancement Extended Processing Subsumptive Constraints: Enhancement Impairment Driven to find generalizations Williams, Lombrozo & Rehder, CogSci 2011; Kuhn & Katz, 2009
- 24. Predicting people’s behavior 10 people: 5 rarely donated, 5 frequently Anna is living on East Coast Person dominating 28 A science major Predict Rarely or frequently donates? Feedback Anna rarely donates to Explain why charities. Anna rarely donates Explain Anna is living on East Coast to charities. dominating Vs. 28 A science major Control (Study) N = 182
- 25. Instances vs. Generalizations Unique features Pattern related Irrelevant features Unique features: features 10/10 predictions Picture Name Age Personality "living on the" "a graduate of a" Rarely donates to charities Anna 28 dominating East coast science major Reliable Patterns: Joseph 32 friendly West coast humanities major 10/10 predictions Sarah 24 boastful West coast science major Jessica 26 26 self-assured East coast science major Misleading Patterns: Mistakes in predictions Kevin 30 energetic energetic West coast humanities major Frequently donates to charities Steven 42 cautious East coast science major Josh 38 discreet West coast humanities major Effect of Laura 37 studious West coast science major explaining? Janet 45 45 self-conscious West coast humanities major Karen 39 quiet quiet East coast science major
- 26. Explanation Impairment Effect Anna donates to charities. All 10 descriptions shown 28 five times. On East Coast Dominating science major 0.5 Explain Learning Error Free Study 0.25 0 Reliable Unreliable
- 27. Category Learning * blocks to learn Number of 10 Explain Control Think Aloud 5 Reliable Misleading N = 240
- 28. Anomalous Evidence • Anomalies contradict current or prior beliefs • Often ignored (Chinn & Brewer, 1993 ) • When does explaining lead to their use? • Manipulated number of anomalies • Learning about importance of deviation in ranking scores
- 30. Strength of Anomalous Evidence
- 31. Interaction with strength of anomalous evidence
- 32. Conclusions • Explaining “Why?” guided by a subsumptive constraint: 1. Promotes discovery of generalizations 2. Favors generalizations that prior knowledge suggests apply broadly beyond specific cases 3. Impairs learning when generalizations are unreliable 4. Promotes belief revision from anomalies • Implications for predicting explanation’s effects • Future directions – More complex contexts – How does explaining recruit other cognitive processes? 32
- 33. Acknowledgements • Bob Rehder • Norielle Adricula, Dhruba Banerjee, Adam Krause, Sam Maldonaldo, Kelly Whiteford • Joe Austerweil, Randi Engle, Nick Gwynne, Luke Rinne, and Karen Schloss. • Concepts and Cognition Lab. 33
Editor's Notes
- #3: Not traditional, but ubiquitous
- #4: Mention: simply explaining, with NO FEEDBACK about whether the explanations were correct.Experience of explaining to someone else, not just showing a solution to a problem, but explaining WHY. Prompting children to provide explanations- for people’s behavior- drives major conceptual change- like in theory of mind.
- #5: Explaining as elaborationExplaining as communication
- #6: Despite demonstrations of its effects, less research probing underlying mechanisms, so it’s challenging to predict what kind of knowledge gained through explaining, contexts in which it will be effective, how can support learners in constructing good explanations.
- #7: Not traditional, but ubiquitous
- #9: Philosophical theories.For example. Explain why something is an instance of a concept in terms of a rule. Explain why an event occurred or person behaved certain way can uncover underlying causal regularity. Explain why you solve a problem a certain way helps gain insight into the underlying principle.
- #10: Conceptual analyses of subsumption in philosophy have a broad scope of application: instances of a concept, causal reasoning, mathematical principles.Investigate here because of the importance of understanding how people learn concepts and categories, because many other learning tasks share similar formal structure to learning a category concept, and because its extensive investigation in cognitive psychology allows experimental control, over people’s prior knowledge and in more precise operationalization of explanation and measures of generalization.
- #13: Good control conditions.
- #15: Minimal manipulation of PK. TO suggest one pattern was more likely to subsume observations beyond study context.
- #18: There were therefore 2 rules that were associated with items that made the machine go. One rule, the 100% rule, accounted for all of the data – 100% of the time, if the object had a green piece on it, it would make the machine play music. The other rule, the 75% rule, accounted for most of the data – 75% of the time, if the object had a red piece on it, it would make the machine play music.
- #24: *Like visual illusions.*Almost exclusively positive effects of explanation, impairment of explanation hasn’t been predicted.SET OURSELF the goal of identifying a context where manipulate how reliable pattern is, manipulate whether people explain, to test this interaction.*Many contexts where could be an issue.
- #25: CONTROL LIKELY EXPLAINED AS WELLNO SPECIAL REASON TO TEST SUBSUMPTIVE CONSTRAINTS HEREMore complex.Learn about individual instances or form generalizations.Explanations of behavior have been examiend (malle).Explanation plays a role in acquistion of stereotypes?Choice of this domain.
- #28: Fewer blocks reflects quicker learning.Impairment EVEN though constant feedback.
- #33: Selective vs General benefit.Implications: interpret previous findings: why explaining uniquely beneficial and key to understanding. educational settings. Children. Illusory correlation, Stereotyping.Constraint to find patterns spontaneously elicit & guide processes like comparison & analogy. How the target cases provided for explanation and prior knowledge about these cases guide what is learned.