Recently, at EPFL's machine learning summer camp, Turing Award winner Yann LeCun's latest report "From Machine Learning to Autonomous Intelligence" elaborated on the recent specific concepts of autonomous intelligence, which is very worthy of attention!
AI/ML/DL Applications Today
-
Relying heavily on supervised deep learning. There are some on Deep RL.
-
Increasing reliance on pre-training self-supervised learning
ML/DL sucks now compared to humans and animals
-
Humans and animals learn models of the world
Self-Supervised Learning
-
Main problems: representing uncertainty, learning abstractions.
energy-based model
-
Contrastive Learning Methods Samples
-
Non-contrastive learning methods
Deep SSL is a enabler for the next AI revolution
I will try to convince you:
-
Ditching Supervision and Reinforcement Learning
Well, not exactly, but as much as possible.
-
Abandonment probability model
Use an energy-based framework instead
Abandon generative models
Use a federated embedding architecture instead
Using a Hierarchical Latent Variable-Based Energy Model
Enable machines to reason and plan.
See manuscript: "The Path to Autonomous Machine Intelligence"
https://openreview.net/forum?id=BZ5a1r-kVsf
Expert knowledge and convenient viewing
Easy to download , please pay attention to the official account of Zhizhi (click the blue Zhizhi above to follow)
Reply " A192 " in the background to get the download link of " From Machine Learning to Autonomous Intelligence", the latest report of Turing Award winner Yann LeCun, with 192 pages of ppt and video .