The End Of Knowledge - Vault 4/5 - AI For Good - 2017 - Peter Lee

graph LR classDef industry fill:#d3f4ff, font-weight:bold, font-size:12px classDef learning fill:#ffd3e0, font-weight:bold, font-size:12px classDef transfer fill:#d3ffd4, font-weight:bold, font-size:12px classDef applications fill:#fff3d3, font-weight:bold, font-size:12px classDef specialization fill:#f3d3ff, font-weight:bold, font-size:12px classDef democratization fill:#f9f9d4, font-weight:bold, font-size:12px classDef growth fill:#d4f9d4, font-weight:bold, font-size:12px classDef responsibility fill:#f9d4d4, font-weight:bold, font-size:12px A[Peter Lee
Microsoft AI and
Research] A --> B[AI transitioning from
research to industry. 1] B --> C[Machine learning depends
on labeled data. 2] C --> D[Training data is
labor-intensive, expensive. 3] D --> E[Machine learning needs
skilled operators. 4] A --> F[Transfer learning improves
performance across languages. 5] F --> G[Transfer learning mirrors
human learning. 6] A --> H[Industry racing to
acquire AI talent. 7] H --> I[AI surprises, like
Skype for teachers. 8] A --> J[Computer vision: deep nets
caption images. 9] J --> K[AI in healthcare: speeds
medical imaging. 10] A --> L[Machine learning models
specialized, not generalized. 11] L --> M[Democratizing AI: tools
for innovators. 12] A --> N[Exponential growth marks
historical inflections. 13] N --> O[AIs emergence transformative
like printing press. 14] A --> P[Consider AI disruptions,
like literacy needs. 15] P --> Q[Oversight needed beyond
industry alone. 16] A --> R[Cybersecurity tech misused
in Iran, 2009. 17] R --> S[Powerful tech used
for good, ill. 18] A --> T[Researchers now recognize
techs dual nature. 19] T --> U[All tech community
responsible for AI. 20] class B,C,D,E industry class F,G transfer class H,I applications class J,K applications class L,M specialization class N,O growth class P,Q responsibility class R,S responsibility class T,U responsibility

Resume:

1.- AI is transitioning from a research pursuit to an industrial one, but it's not fully there yet - in an "in-between" artisanal stage.

2.- Machine learning, the core of AI practice today, is highly dependent on large amounts of human-labeled training data.

3.- Obtaining training data is labor-intensive and expensive. Companies try to hoard or monetize this valuable data.

4.- The machine learning process requires highly skilled people to operate the systems, set parameters, and integrate the resulting models into applications.

5.- Transfer learning, where a model trained on one language can improve performance on another language, is an important phenomenon.

6.- While not truly biological, transfer learning is alluring as it mirrors how humans learn, leading to hype about AI.

7.- Industry is racing to acquire skilled labor to build AI models and products, like Skype's real-time translation of 9 languages.

8.- Surprising applications emerge once AI is deployed at scale, like teachers using Skype Translator to accommodate students with hearing loss.

9.- Computer vision is rapidly advancing, with applications like using deep neural nets to caption images taken with smartphones.

10.- AI is augmenting healthcare, such as using computer vision to speed up analysis of medical imaging for radiotherapy planning.

11.- Most valuable machine learning models are highly specialized for single applications and don't generalize well, requiring new models for each use case.

12.- Democratizing AI means building tools to allow more innovators to create machine learning models, which companies like Microsoft are working on.

13.- Exponential technology growth marks inflection points in human history, such as the rapid growth of printed books in the 15th century.

14.- The emergence of practical AI may be a similar transformative period to the impact of the printing press, which helped spur the Renaissance.

15.- We must be thoughtful about the disruptions AI will cause, much like how the printing press made literacy a necessary skill.

16.- Peter Lee doesn't think oversight of AI development should fall only on industry, but is an issue for the whole tech community.

17.- In 2009, U.S. developed cybersecurity tech was used by Iran to crack down on citizens using social media to protest election results.

18.- This was an early lesson that powerful technologies can be used for good or ill, something many researchers hadn't considered before.

19.- Since then, the research community has made progress in recognizing the dual-edged nature of the technologies they develop.

20.- It's up to not just industry, but researchers and tech innovators to continue advancing this understanding of responsible AI development.

Knowledge Vault built byDavid Vivancos 2024