Generalization via analogy in young children and Large Models.
summary
Empirical tests show toddlers use relational rules to generalize: 24‑month‑olds (n=100 preregistered) pick the right machine to make a hat fit after seeing a change machine demonstration, and preschoolers predict size, number, rotation and reflection outcomes. A Kid‑inspired Visual Analogies dataset, the KiVA benchmark (1,400 visual transformations of Toys4K objects), compares humans and models: adults and children identify what changed, quantify how it changes, and extrapolate; GPT‑4V, LLaVA‑1.5 and MANTIS reliably detect what but fail to quantify the change and to extrapolate to new objects.
AI-written summary · about
chapters (10)
- 0:00Introduction to analogy
- 1:38Science and analogy
- 4:29AI challenges with analogies
- 6:53Developmental findings
- 10:18Causal reasoning in toddlers
- 15:35The KIVA benchmark
- 20:38Results and performance
- 25:30Testing AI capabilities
- 28:29The path forward
- 33:43Discussion and Q&A
related talks
20mKiVA: Kid-inspired Visual Analogies for Large Multimodal Models
33mLearning Abstract Causal Structure From Causal Interventions In Children And Artificial Agents
01h 16mGeneralization in natural language processing
01h 05mAdaptive Generalization: The Role of Dynamic Evaluation in Low-Resource Settings
01h 10mModern paradigms of generalization, the heliocentric model of Aristarchus,...
50mKnowledge is embedded in language neural networks but can they reason?
27mBuilding Mini Minds Our Adventures in Teaching Small LLMs to Reason
01h 18mComputer Vision after the Victory of Data