Yann LeCun on Self-Supervised Learning | ICLR 2020

Welcome to AIP.
- The main focus of this channel is to publicize and promote existing SoTA AI research works presented in top conferences, removing barriers for people to access the cutting-edge AI research works.
- All videos are either taken from the public internet, or the Creative Common licensed, which can be accessed via the link provided in the description.
- To avoid conflict of interest with the ongoing conferences, all videos are published at least 1 week after the main events. A takedown can be requested if it infringes your right via email.
- If you would like your presentation to be published on AIP, feel free to drop us an email.
- AI conferences covered include: NeurIPS (NIPS), AAAI, ICLR, ICML, ACL, NAACL, EMNLP, IJCAI

If you would like to support the channel, please join the membership:
https://www.youtube.com/c/AIPursuit/join

Subscribe to the channel:
https://www.youtube.com/c/AIPursuit?sub_confirmation=1

Donation:
w/ BEP20 (BTC, ETH, USDT, SOL, BNB, Doge, Shiba) ⇢ 0x0712795299bf00eee99f13b4cda0e19dc656bf2c

USDT (TRN20) ⇢ THV9dCnGfWtGeAiZEBZVWHw8JGdGCWC4Sh

Play smarter and safer on Stake while staying anonymous.
Use my affiliate link now: stake.com/?c=fefa962a46

The video is reposted for educational purposes and encourages involvement in the field of AI research.

A good GitHub repo on self supervised learning: https://github.com/jason718/awesome-self-supervised-learning#machine-learning

0:00 Introduction
1:32 Three challenges for Deep Learning Deep Supervised Learning works well for perception
2:26 Self-Supervised Learning = Filling in the Blanks
4:38 Energy-Based Model: unconditional version
5:23 Architecture for Multimodal Output: latent variable EBM
6:21 Latent-Variable EBM: inference
8:25 Contrastive Methods vs Regularized Architectural Methods
9:09 Problem with Max Likelihood / Probabilistic Methods
11:44 Contrastive Embedding
13:25 GANS: training a network to generate contrastive samples
14:03 Architectural Methods & Regularized Methods
19:18 Learning a Forward Model of the World
20:45 Conclusions / Conjectures