Wenjie Li @CogSci
@wenjieli.bsky.social
PhD student at @CarnegieMellon with @ybisk.me | prev: Research Scientist at @NYUDataScience with @brendenlake.bsky.social. Concept, Symbols, Abstraction, Cognitive Science, AI
https://wliwenjieli.github.io
https://wliwenjieli.github.io
(3/3) This suggests more abstract concepts need more data and capacity to process it, potentially revealing that uniquely human abstraction emerged through expanded information capacity. Swing by to chat about the poster or anything: representation, relations, abstraction, symbols, compositionality!
August 2, 2025 at 6:18 PM
(3/3) This suggests more abstract concepts need more data and capacity to process it, potentially revealing that uniquely human abstraction emerged through expanded information capacity. Swing by to chat about the poster or anything: representation, relations, abstraction, symbols, compositionality!
(2/3) We compared 1,691 neural networks with humans, children, and monkeys on abstraction tasks. As abstraction demands grow, networks diverge from adults and mirror children/primates. Model scale and training data are key predictors, esp for relational tasks where only 0.5% of models succeeded.
August 2, 2025 at 6:18 PM
(2/3) We compared 1,691 neural networks with humans, children, and monkeys on abstraction tasks. As abstraction demands grow, networks diverge from adults and mirror children/primates. Model scale and training data are key predictors, esp for relational tasks where only 0.5% of models succeeded.
(2/3) We compared 1,691 neural networks with humans, children, and monkeys on abstraction tasks. As abstraction demands grow, networks diverge from adults and mirror children/primates. Model scale and training data are key predictors, esp for relational tasks where only 0.5% of models succeeded.
August 2, 2025 at 6:15 PM
(2/3) We compared 1,691 neural networks with humans, children, and monkeys on abstraction tasks. As abstraction demands grow, networks diverge from adults and mirror children/primates. Model scale and training data are key predictors, esp for relational tasks where only 0.5% of models succeeded.