Thomas Gilbert submits "The Passions and the Reward Functions - Rival Views of AI Safety?" to FAT*2020

28 Aug 2019

Thomas Krendl Gilbert submitted “The Passions and the Reward Functions: Rival Views of AI Safety?” to the upcoming Fairness, Accountability, and Transparency (FAT*) 2020 Conference.

Rohin Shah Publishes "Clarifying Some Key Hypotheses in AI Alignment" on the Alignment Forum

27 Aug 2019

CHAI PhD student Rohin Shah, along with Ben Cottier, pubished the blog post “Clarifying Some Key Hypotheses in AI Alignment” on the AI Alignment Forum. The post maps out different key and controversial hypotheses of the AI Alignemnt problem and how they relate to each other.

Siddharth Srivastava Publishes "Why Can't You Do That, HAL? Explaining Unsolvability of Planning Tasks"

17 Aug 2019

CHAI PI Siddharth Srivastava, along with his co-authors Sarath Sreedharan, Rao Kambhampati, David Smith, published “Why Can’t You Do That, HAL? Explaining Unsolvability of Planning Tasks” in the 2019 International Joint Conference on Artificial Intelligence (IJCAI) proceedings. The paper discusses how, as anyone who has talked to a 3-year-old knows, explaining why something can’t be done can be harder than explaining a solution to a problem. The paper then goes into new work in having AI explain unsolvability.

Michael Wellman gave talk "Trend-Following Trading Strategies and Financial Market Stability" at ICML 2019

16 Aug 2019

CHAI PI Michael Wellman gave a talk at the ICML’s Workshop on AI and Finance on how one form of algorithmic (AI) trading strategy can affect financial market stability. The video for the talk can be found here.

Mark Nitzburg Publishes WIRED Article Advocating for an FDA for Algorithms

15 Aug 2019

CHAI’s Executive Director Mark Nitzburg, along with Olaf Groth, published an article in WIRED Magazine that advocates for the creation of an “FDA for algorithms.”

Rohin Publishes "Learning Biases and Rewards Simultaneously"

05 Jul 2019

Rohin Shah published a short summary of the CHAI paper “On the Feasibility of Learning, Rather than Assuming, Human Biases for Reward Inference”, along with some discussion of its implications on the Alignment Forum.

CHAI Releases Imitation Learning Library

05 Jul 2019

Steven Wang, Adam Gleave, and Sam Toyer put together an extensible and benchmarked implementation of imitation learning algorithms commonly used at CHAI (Notably GAIL and AIRL) for public use. You can visit the Github here.

CHAI Paper Featured in New Scientist Article

01 Jul 2019

A recent New Scientist article features a paper that Tom Griffiths and Stuart Russell wrote along with David D. Bourgin, Joshua C. Peterson, and Daniel Reichman. The article discusses how the researchers were able to make a machine learning model that took into account human biases, like risk adversion, that are usually hard for computer systems to model.

CHAI Presentations at ICML

15 Jun 2019

CHAI faculty and graduate students presented their papers at the latest International Conference on Machine Learning.

CHAI Presents Paper on Adversarial Learning at ICML

14 Jun 2019

CHAI researchers Michael Dennis, Adam Gleave, Cody Wild, Neel Kant, and Stuart Russell, along with Sergey Levine, gave a talk on their paper Adversarial Policies: Attacking Deep Reinforcement Learning at the International Conference on Machine Learning 2019. There is a video of the talk on the ICML Github (starts at 1h:35m) and the slides can be here