How Can Machines Learn Human Values? - With Brian Christian

Explore the complexities and ethical challenges of machine learning in this thought-provoking lecture by Brian Christian. Delve into the alignment problem, examining how AI systems can fail to meet human expectations and the potential risks that arise. Learn about issues in training data, representation in datasets, and the open category problem. Investigate fairness in machine learning, particularly in criminal justice applications, and understand the intricacies of reinforcement learning. Discover innovative approaches like TCAV, behavior cloning, and inverse reinforcement learning that aim to address these challenges. Gain insights into the ongoing efforts to create AI systems that better align with human values and expectations.

Intro
walter Pitts
Imagenet
The alignment problem
Machine learning systems
Training data
Labels in the wild
Representation and data sets
What can we do about this
TCAV
Fire Truck
Robustness
Open Category Problem
Examples
What makes a system fair
Predicting reoffending
Ground truth
Potential ethical issue
Criminal justice
Risk vs Needs
Reinforcement Learning
Facebook Reinforcement Learning
Coast Runners 3
Sparsity
Designing reward functions
Behavior cloning
Shadow mode
Inverse reinforcement learning