The End Of Knowledge - Vault 5/81 - CVPR - 2023 - Revisiting Old Ideas With Modern Hardware

graph LR classDef ai fill:#f9d4d4, font-weight:bold, font-size:14px classDef computation fill:#d4f9d4, font-weight:bold, font-size:14px classDef expectations fill:#d4d4f9, font-weight:bold, font-size:14px classDef challenges fill:#f9f9d4, font-weight:bold, font-size:14px classDef insights fill:#f9d4f9, font-weight:bold, font-size:14px A[Revisiting Old Ideas
With Modern Hardware] --> B[Brooks: influential, challenged status quo. 1] A --> C[Computational power increase enables AI. 2] C --> D[Neural networks long, key history. 3] A --> E[Short-term overestimation, long-term underestimation common. 4] E --> F[Advanced tech appears magical, overestimated. 5] E --> G[Task-specific AI misinterpreted as general. 6] E --> H[Real-world AI deployment surprisingly slow. 7] E --> I[Hollywood depicts unrealistic, sudden changes. 8] A --> J[True exponential growth rare, plateaus. 9] J --> K[Exponential computation ≠ human intelligence. 10] A --> L[AI hype leads to overgeneralization. 11] A --> M[Language-vision integration requires embodied grounding. 12] M --> N[Human intelligence connects words to
models, objects, geometry. 13] M --> O[Deep learning may not solve
symbol grounding problem. 14] A --> P[Perception: fuzzy boundaries, functional understanding. 15] P --> Q[Smith: registration deeper than labeling. 16] A --> R[Old papers origins, early generations
yield new ideas. 17] A --> S[Large language models surprising, novel. 18] A --> T[Unique research directions more impactful. 19] A --> U[Women made foundational AI contributions. 20] C --> V[Computation enables Asimovs laws implementation. 21] A --> W[Deep learning temporary, new paradigms
will emerge. 22] A --> X[Geometry important, often overlooked in
AI, vision. 23] C --> Y[Cheap edge computation enables intelligent
robots. 24] A --> Z[Backing old ideas with modern
computation yields breakthroughs. 25] class A,B,U ai class C,D,V,Y computation class E,F,G,H,I,J,K expectations class L,M,N,O,P,Q,W,X challenges class R,S,T,Z insights

Resume:

1.- Rod Brooks is an influential AI researcher who challenged the status quo and shifted focus to real-world systems and nature-inspired intelligence.

2.- Computational power has increased dramatically, enabling new possibilities for AI research.

3.- Neural networks have a long history, with key ideas like shift invariance, backpropagation, and convolutional neural networks developed over decades.

4.- Overestimating short-term impact and underestimating long-term effects of technology is common.

5.- Sufficiently advanced technology appears like magic, leading to overestimation of AI capabilities.

6.- AI performance on specific tasks is often misinterpreted as general competence.

7.- Real-world deployment of AI systems takes much longer than expected due to practical challenges and inertia.

8.- Hollywood scenarios depict unrealistic, sudden technological changes without considering co-evolution of society and technology.

9.- True exponential growth in technology is rare; most improvements eventually plateau.

10.- Exponential computational growth does not guarantee functional equivalence to human intelligence.

11.- Imprecise language and hype around AI leads to overgeneralization of narrow advances.

12.- Combining language models and computer vision requires more than simple integration; grounding symbols in embodied experiences is crucial.

13.- Human intelligence relies on connecting words to models, embodied objects, and geometry, not just linguistic descriptions.

14.- Deep learning may not fully solve the symbol grounding problem, which involves stability and ongoing perceptual integration.

15.- Perception involves fuzzy boundaries and functional understanding, not just labeling based on image characteristics.

16.- Brian Cantwell Smith introduces the concept of registration, a deeper understanding than labeling, as a key challenge for AI.

17.- New ideas can be found by studying the origins and early generations of old papers in depth.

18.- Large language models were a surprising and novel development in AI over the past decade.

19.- Pursuing unique research directions, rather than following crowded fields, can lead to more impactful contributions.

20.- Women have made significant foundational contributions to AI research.

21.- Increased computational power enables new capabilities, such as implementing Asimov's laws in robotics.

22.- Deep learning is likely a temporary focus in AI; new paradigms will emerge in the future.

23.- Geometry is an important but often overlooked aspect of AI and computer vision.

24.- Cheap, massive computation is now available at the edge, near sensors like cameras, enabling intelligent robots.

25.- Backing and reading old ideas with modern computational power may lead to new breakthroughs in AI.

Knowledge Vault built byDavid Vivancos 2024