No video :(

2019-05-03 Pieter Abbeel ODSC Keynote: Recent Advances and Trends in Artificial Intelligence

Подписаться 20 тыс.

Просмотров 4,5 тыс.

50% 1

Recent advances in Artificial Intelligence have enabled a wide range of breakthroughs in many domains, including image recognition, speech recognition, machine translation, learning to play video games, learning to control simulated and real robots, and mastering the classical game of Go. In this talk I will start from some of the eye-catching headlines in the popular press, spanning AI taking over the world, national strategies and priorities, major valuations, and claims that AI is the new electricity. From there I will highlight actual advances being made, provide background on ideas and trends that have powered these advances, and share examples on how AI is making impact in various organizations. Finally I'll discuss near and long term considerations, my thoughts on whether an AI winter is impending, and on the relationship between AI and humans.

Опубликовано:

21 авг 2024

Ссылка:

Скачать:

Готовим ссылку...

Добавить в:

Мой плейлист

Посмотреть позже

Комментарии : 4

@prodbydjsmalls 5 лет назад

Excellent talk Pieter.

@smugada 5 лет назад

Excellent insights into pros and cons of AI. Thanks

@ProfessionalTycoons 5 лет назад

very good talk

@eugenedsky3264 5 лет назад

29:00 I think, improvements in CAD, CAE software are even more important. There is an idea for you here: drive.google.com/file/d/1GSv89tiQmPDcnFEu4n4CqfaJcUJxVmL5KrSCJ047g4o/edit Instead of NN or ELGs you can use a learning Mealy machine to store already known observation-action-reward tuples, which are context dependent. By action I imply some logic transform. Reward depends on how beneficial that transform is. And then, while compressing the automaton, you transform its transition and value functions with hypothesis logic which assumes what unknown tuples ("Don't care" values) must be equal to. That Mealy machine based agent may simplify the simulation with low-detail VR engine in a brute-force algorithm easily. At least you may start with simple tasks and then as the agent learns how to simplify the extensive search for your robot's code you may add complexity. Of course, some tasks (like 3D reconstruction, maybe) are better programmed by humans and not by that extensive search... Hope that helps.