Center for AI Safety at Stanford University
A Quick Look at Whitepaper, Membership and Flagship Projects
Yesterday I discovered AI Safety at Stanford and it seemed like an interesting project. After visiting their website I observed they had a whitepaper online relating to their mission; they were doing corporate membership schemes; and have a series of flagship projects. Due to my 50 day focus on AI safety within my #500daysofAI it made sense to dive into this information.
What is Center for AI Safety at Stanford University ?
First to state the obvious: it is a centre focused on AI Safety based at Stanford University. Stanford has other initiatives such as the recently established Institute for Human-Centered Artificial Intelligence. Further back it has the Stanford Artificial Intelligence Laboratory (SAIL) which has been a center of excellence for Artificial Intelligence research, teaching, theory, and practice since its founding in 1962. Therefore we can assume some possible spillover and collaboration. Besides this I could not immediately find more relevant information (write me a response and I will update this article).
Whitepaper Overview
Immediately on the website there is a mission statement with a link to a whitepaper simply named Stanford Center for AI Safety. The paper is six pages long, so you can easily go to the website yourself and spend some time going through the text.
They outline three research directions in the paper: (1) Formal Techniques for AI Safety; (2) Learning and Control for AI Safety; (3) Transparency for AI Safety.
The first formal techniques relates to specification and verification of systems with AI components. As well as analysis of adversarial robustness and automatic test-case generation.
The second learning and control relates to AI agents. Safe exploration and learning for perception by AI systems. In addition to safe control of AI agents.
The third is transparency. Within this lies the focus on explainable, accountable and fair AI. However it is diagnosis and repair for systems with AI components is important, particularly detecting and diagnosing in an online setting.
Here is one quote from the whitepaper that resounded with me:
Because such code is written by humans, good software
engineering practices coupled with formal methods can ensure that it is also guaranteed to perform as expected. In machine-learned systems, however, the
program amounts to a highly complex mathematical formula for transform-
ing inputs into outputs. Humans can barely parse the formulas defining these
systems, let alone reason about them.
Corporate Membership
It is said on their website that their corporate members are a ‘vital and integral part’ of the Center for AI Safety. Since their core members are General Electric, NVIDIA and Uber the promise of providing insight on real-world use cases, valuable financial support for research, and a path to large-scale impact does not seem unfounded. Core sponsors are able to influence flagship research and sit at the advisory board.
Their federal sponsors are also DARPA (The Defense Advanced Research Projects Agency) and NSF (National Science Foundation).
Flagship Projects
I will list excerpts of the flagship projects underneath. They are of course not ranked in order of importance simply appearence. See if any catches your interest:
- Leader Follower. Leading and following can emerge naturally in human teams. However, such roles are usually predefined in human-robot teams due to the difficulty of scalably learning and adapting to each agent’s roles. Our goal is to enable a robot to learn how leader-follower dynamics emerge and evolve in human teams and to leverage this understanding to influence the team to achieve a goal.
- Robustness Verification. ACAS sXu is a protocol for collision avoidance for small drones. In this project we aim to do a formal robustness analysis of a specific implementation of sXu being used by GE. This implementation uses a deep neural network (DNN).
- Adaptive Stress Testing — Avionics. The control of unmanned aircraft systems must be rigorously tested and verified to ensure their correct functioning and airworthiness. Incorporating components that use novel techniques such as deep learning can pose a significant challenge because traditional approaches for detecting errors require deriving a model of a correctly performing controller, which can be intractable.
- Adaptive Stress Testing — Automotive. Adaptive stress testing (AST) is an approach for validating safety-critical autonomous control systems such as self driving cars and unmanned aircraft. The approach involves searching for the most likely failure scenarios according to some measure of likelihood of occurrence.
On their website they have link to a series of courses taught at Stanford related to this topic, as such it may be possible to explore the curriculum of the different courses to understand AI Safety better. Additionally there is a list of researchers and link to publications.
I hope this was short, to the point and informative. I will do my best to follow the developments at the Center for AI Safety at Stanford University going forward.
This is day 63 of #500daysofAI. My current focus for day 50-100 is on AI Safety.
