Standard model for AI

Standard model for AI is human specifies the objective like here’s the discounted sum of rewards machine says okay I’m on it right now we know this model doesn’t work because we can’t specify the objective correctly and I should mention it’s not just our model this is the same model as control theory where you minimize the cost function in statistics you minimize risk in operations research in fact this is borrowed from operations research you maximize the discount of some rewards and economics you maximize a welfare function or GDP or profit or whatever objective it is all of these are the same basic standard model and it’s a bad model and King Midas could tell you that because he said okay here’s my objective I want everything I touch the turn to gold the machine said righty-ho in fact it was the gods in his case.

Deep Learning for System 2 Processing

Slides: https://photos.app.goo.gl/KfQjX5FqcGEsVEs49

1:09:37 [Talk: Deep Learning for System 2 Processing by Yoshua Bengio] 1:10:10 No-Free-Lunch Theorem, Inductive Biases Human-Level AI 1:15:03 Missing to Extend Deep Learning to Reach Human-Level AI 1:16:48 Hypotheses for Conscious Processing by Agents, Systematic Generalization 1:22:02 Dealing with Changes in Distribution 1:25:13 Contrast with the Symbolic AI Program 1:28:07 System 2 Basics: Attention and Conscious Processing 1:28:19 Core Ingredient for Conscious Processing: Attention 1:29:16 From Attention to Indirection 1:30:35 From Attention to Consciousness 1:31:59 Why a Consciousness Bottleneck? 1:33:07 Meta-Learning: End-to-End OOD Generalization, Sparse Change Prior 1:33:21 What Causes Changes in Distribution? 1:34:56 Meta-Learning Knowledge Representation for Good OOD Performance 1:35:14 Example: Discovering Cause and Effect 1:36:49 Operating on Sets of Pointable Objects with Dynamically Recombined 1:37:36 RIMS: Modularize Computation and Operate on Sets of Named and Typed Objects 1:39:42 Results with Recurrent Independent Mechanisms 1:40:17 Hypotheses for Conscious Processing by Agents, Systematic Generalization 1:40:46 Conclusions

Self-Supervised Learning

Slides: https://photos.app.goo.gl/9JwsDo8wMZatpj9w7

36:04 [Talk: Self-Supervised Learning by Yann LeCun] 36:25 What is Deep Learning? 38:37 Supervised Learning works but requires many labeled samples 39:25 Supervised DL works amazingly well, when you have data 40:05 Supervised Symbol Manipulation 41:50 Deep Learning Saves Lives 43:40 Reinforcement Learning: works great for games and simulations. 45:12 Three challenges for Deep Learning 47:39 How do humans and animals learn so quickly? 47:43 Babies learn how the world works by observation 48:43 Early Conceptual Acquisition in Infants [from Emmanuel Dupoux] 49:33 Prediction is the essence of Intelligence 50:28 Self-Supervised Learning = Filling in the Blanks 50:53 Natural Language Processing: works great! 51:55 Self-Supervised Learning for Video Prediction 52:09 The world is stochastic 52:43 Solution: latent variable energy-based models 53:55 Self-supervised Adversarial Learning for Video Prediction 54:12 Three Types of Learning 55:30 How Much Information is the Machine Given during Learning? 55:54 The Next Al Revolution 56:23 Energy-Based Models 56:32 Seven Strategies to Shape the Energy Function 57:02 Denoising AE: discrete 58:44 Contrastive Embedding 1:00:39 MoCo on ImageNet 1:00:52 Latent-Variable EBM for inference & multimodal prediction 1:02:07 Learning a (stochastic) Forward Model for Autonomous Driving 1:02:26 A Forward Model of the World 1:04:42 Overhead camera on highway. Vehicles are tracked 1:05:00 Video Prediction: inference 1:05:15 Video Prediction: training 1:05:30 Actual, Deterministic, VAE+Dropout Predictor/encoder 1:05:57 Adding an Uncertainty Cost (doesn’t work without it) 1:06:01 Driving an Invisible Car in “Real” Traffic 1:06:51 Conclusions