190 likes | 222 Views
Authors: S. Friedman (SIFT, USA), M. McLure & K. Forbus (Northwestern Uni., USA) Presentation of Georgios Samaras, for master course “ADVANCED ARTIFICIAL INTELLIGENCE”, prof. Panagiotis Stamatopoulos. Extending Analogical Generalization with Near-Misses (ALIGN). Intro.
E N D
Authors: S. Friedman (SIFT, USA), M. McLure & K. Forbus (Northwestern Uni., USA) Presentation of Georgios Samaras, for master course “ADVANCED ARTIFICIAL INTELLIGENCE”, prof. Panagiotis Stamatopoulos. Extending Analogical Generalization with Near-Misses (ALIGN)
Intro • Learning concepts from examples, for cognitive systems • Similarity-based supervised learning, labeled examples • Near-miss by Winston (1970). His system used analogical matching to compare the structured representations • Winston-Limits: domain-specific analog. Matcher, teacher must know internal representation & label near-misses and one representation per concept => no disjunctive concepts • ALIGN built on structure-mapping (based on theory: analogy & similarity) • Pros: 1)learns characteristic+discriminative properties, 2) able to learn disjunctive categories, 3)auto-identifies near-misses
Background (1/4) • Structure-Mapping(SM): Analogical comparison = extract commonalities between examples, to form a concept. • Structure-Mapping Engine (SME) is a simulation of SM. Local-to-global process: Build local match hypotheses=>global mappings with similarity scores, candidate inferences (base<=>target) and analogy skolems (entities that do not match)
Background (1/4) • MAC/FAC: model of similarity-based retrieval built on SME. Finds 3 match cases from its library with the input and then uses SME to actually compute the similarity and return all 3 or the best. • Sequential Analogical Generalization Engine (SAGE): model of analog. Generalization build on SME. Successor to SEQL(the general idea is the same). Merge example to generalization context if similar enough, or make it a new cluster. Prune low probability expressions.
Shock graphs, derived from skeleton, nodes relate to surrounding area->segment space
ALIGN: Detecting/exploiting Near-Misses with Analogy • Finds 1 Near-Miss from memory, with high similarity, but different label. • Use Near-Misses to learn which are the critical criteria for asserting category membership (inclusion hypothesis), derived from + → - and the sufficient criteria for blocking category membership (exclusion hypothesis), derived from - → +.
ALIGN: Revising Hypotheses with Analogical Generalization • Every label has a SAGE context. Update context with new example, prune generalized hypotheses that are not true for context's examples • The bigger the context, the more trustworthy • Multiply SAGE clusters for a context → disjunctive category structure. On new example, use structural similarity to determine which cluster(s) should be tested.
ALIGN: Classification via analogy • On new example, fetch from memory similar labeled examples. • Start by most similar: 1) map their correspondences, 2) test and match (watch out for exclusion/inclusion hypotheses). • If no match, repeat with next example (important for distinguishing similar categories)
Evalutaion • ALIGN is the full system described so far • Prototypes is ALIGN without near-miss analysis • Examples is ALIGN without near-misses and analogical generalization (so, Examples do similarity-based retrieval over the library of training examples)
Classification of sketched objects in GogSketch. A case contains 4.5 entities+31 facts
Classification in geospatial concepts by using GogSketch on Freeciv map. A case contains 8 entities+60 facts. Prototypes reduce to Examples.
Conclusion • ALIGN is better from Prototypes and Examples • Sweet spot for Similarity threshold, min creates too many near-misses, max creates small generalizations • ALIGN requires orders of magnitude fewer examples than other existing models