Knowledge Vault 2/70 - ICLR 2014-2023
Aida Nematzadeh Jessica Hamrick Kaylee Burns Joshua B Tenenbaum Alison Gopnik Emmanuel Dupoux ICLR 2020 - Workshop - Bridging AI and Cognitive Science (BAICS)
<Resume Image >

Concept Graph & Resume using Claude 3 Opus | Chat GPT4 | Gemini Adv | Llama 3:

graph LR classDef cognition fill:#f9d4d4, font-weight:bold, font-size:14px; classDef play fill:#d4f9d4, font-weight:bold, font-size:14px; classDef reasoning fill:#d4d4f9, font-weight:bold, font-size:14px; classDef memory fill:#f9f9d4, font-weight:bold, font-size:14px; classDef communication fill:#f9d4f9, font-weight:bold, font-size:14px; A[Aida Nematzadeh et al
ICLR 2020] --> B[AI-cognitive science collaboration explored. 1] A --> C[Play supports learning world models. 2] C --> D[Children's play integrates reasoning, cues. 3] C --> E[Play tracks discriminability of contrasts. 4] C --> F[Play invents problems, bootstraps ideas. 5] A --> G[Cognition's challenge: generating new hypotheses. 6] G --> H[Problems constrain solution search space. 7] G --> I[Play hacks utility functions, goals. 8] A --> J[AI: understand development, diversity, caregiving. 9] A --> K[Cognitive science: understand AI tools. 10] K --> L[Experiments inform AI, innate knowledge. 11] K --> M[Kaelbling: shared AI-intelligence principles, differences. 12] A --> N[Children, RL compared in exploration. 13] A --> O[Autism-inspired AI models atypical reasoning. 14] A --> P[Visual reasoning uses hierarchical imagery. 15] A --> Q[Self-consciousness model improves dialogue consistency. 16] A --> R[Brain-inspired replay reduces catastrophic forgetting. 17] A --> S[Global-local networks learn 3D perception. 18] A --> T[Kalman filter enables transition revaluation. 19] T --> U[SR eigenvectors support hierarchical exploration. 20] A --> V[Hippocampus: episodic memory, planning, task-switching. 21] A --> W[Resource constraints shape decision making. 22] A --> X[Sketch-object mapping relies on reasoning. 23] X --> Y[Learning analogical sketch mappings challenging. 24] A --> Z[Signaling games lead to efficient communication. 25] A --> AA[Product-Kanerva memories enable relational learning. 26] A --> AB[Graph matching enables analogical transfer. 27] A --> AC[Marr's levels disentangle AI debates. 28] A --> AD[Edge/segment models predict mouse vision. 29] A --> AE[Acoustic embeddings resemble human perception. 30] class A,B,J cognition; class C,D,E,F,N play; class G,H,I,O,P,W,X,Y,AB,AC reasoning; class R,S,T,U,V,AA memory; class Q,Z,AE communication;

Resume:

1.-The workshop bridges AI and cognitive science, exploring their shared history and potential for collaboration.

2.-Play and exploration in children supports learning by helping build better models of the world.

3.-Children's play is sophisticated from the start, integrating social cues and reasoning counterfactually.

4.-Experiments show children's exploratory play quantitatively tracks the discriminability of contrasts.

5.-Children's play involves making up arbitrary problems and inventing solutions to bootstrap new ideas and plans.

6.-The hard problem in cognition is thinking and generating new hypotheses, not just learning.

7.-Problems contain information that constrains the search space for solutions even before having the answers.

8.-Children violate rational action principles in play to create novel goals by hacking their utility functions.

9.-AI researchers should understand development, caregiving, diversity of cognition and purposes beyond task performance from cognitive science.

10.-Cognitive scientists should understand the variety of tools and approaches in AI beyond just neural networks.

11.-Natural experiments are proposed to understand innate knowledge, approximations, modularity, spatial reasoning in humans for informing AI.

12.-Leslie Kaelbling discusses pursuing shared principles between AI and natural intelligence while acknowledging differences.

13.-Young children and RL agents are compared exploring the same virtual environment to study efficient exploration.

14.-Autism-inspired AI looks at how atypical visual and social reasoning abilities can inform cognitive modeling.

15.-Visual spatial reasoning is modeled using visual imagery operations in a hierarchical process inspired by cognitive theories.

16.-Public self-consciousness modeled using an imaginary listener improves consistency in dialogue agents without additional annotation.

17.-Brain-inspired replay combining generative feedback and context-dependent gating reduces catastrophic forgetting in neural networks.

18.-Global-local networks using sparse depth cues learn 3D scene perception in a baby-like way through vision and interaction.

19.-A Kalman filter applied to the successor representation enables flexible value computation and transition revaluation.

20.-Eigenvectors of the SR matrix support hierarchical representation and temporally abstract exploration in a grid cell model.

21.-The hippocampus supports key cognitive functions difficult for AI - episodic memory, model-based planning, task-switching, imagination.

22.-Standard RL optimality criteria give way to natural decision making under resource constraints in a cognitive model.

23.-Sketch-object correspondence relies on high-level visual features and social reasoning to communicate relevant information efficiently.

24.-Learning object-sketch mappings that expose analogical structure is a challenge for current visual reasoning systems.

25.-Entropy regularization emerges in signaling games, leading to efficient communication that transmits only necessary information.

26.-Product-Kanerva memories enable compression, binding, relational learning and selective reconstruction of objects in a neural network.

27.-Continuous relaxation of graph matching enables knowledge transfer in analogical reasoning, qualitatively matching human asymmetries.

28.-Marr's levels of analysis help disentangle assumptions and focus debates at the right level in AI research.

29.-Mouse visual cortex responses are best predicted by models capturing edges/segments rather than supervised object recognition features.

30.-Acoustic word embeddings exhibit properties like onset bias and encode phonetic features similar to human percepts.

Knowledge Vault built byDavid Vivancos 2024