A recent controversy erupted over OpenAI’s new version of
their language model for generating well-written next words of text based on
unsupervised analysis of large samples of writing. Their announcement and
decision not to follow open-source practices raises interesting policy issues
about regulation and self-regulation of AI products. OpenAI, a non-profit AI
research company founded by Elon Musk and others, announced on
February 14, 2019, that “We’ve trained a large-scale unsupervised language
model which generates coherent paragraphs of text, achieves state-of-the-art
performance on many language modeling benchmarks, and performs rudimentary
reading comprehension, machine translation, question answering, and
summarization—all without task-specific training.”
The reactions to the announcement followed from the decision behind the following statement in the release: “Due to our concerns about malicious applications of the technology, we are not releasing the trained model. As an experiment in responsible disclosure, we are instead releasing a much smaller model for researchers to experiment with, as well as a technical paper.”
Examples of the many reactions are TechCrunch.com and Wired. The Electronic Frontier Foundation has an analysis of the manner of the release (letting journalists know first) and concludes, “when an otherwise respected research entity like OpenAI makes a unilateral decision to go against the trend of full release, it endangers the open publication norms that currently prevail in language understanding research.”
This issue is an example of previous ideas in our Public Policy
blog about who, if anyone, should regulate AI developments and products that
have potential negative impacts on society. Do we rely on self-regulation or
require governmental regulations? What if the U.S. has regulations and other
countries do not? Would a clearinghouse approach put profit-based pressure on
developers and corporations? Can the open source movement be successful without
Welcome to the eighth interview in our se- ries profiling senior AI researchers. This month we are especially happy to interview our SIGAI advisory board member, Thomas Dietterich, Director of Intelligent Systems at the Institute for Collaborative Robotics and In- telligence Systems (CoRIS) at Oregon State University.
Dr. Dietterich (AB Oberlin College 1977; MS University of Illinois 1979; PhD Stan- ford University 1984) is Professor Emeritus in the School of Electrical Engineering and Computer Science at Oregon State University, where he joined the faculty in 1985. Dietterich is one of the pioneers of the field of Machine Learning and has authored more than 200 refereed publications and two books. His research is motivated by challenging real world problems with a special focus on ecological science, ecosystem management, and sustainable development. He is best known for his work on ensemble methods in machine learning including the development of error- correcting output coding. Dietterich has also invented important reinforcement learning algorithms including the MAXQ method for hierarchical reinforcement learning. Dietterich has devoted many years of service to the research community. He served as President of the Association for the Advancement of Artificial Intelligence (2014-2016) and as the founding president of the International Machine Learning Society (2001-2008). Other major roles include Executive Editor of the journal Machine Learning, co-founder of the Journal for Machine Learning Research, and Program Chair of AAAI 1990 and NIPS 2000. Dietterich is a Fellow of the ACM, AAAI, and AAAS.
Getting to Know Tom Dietterich
When and how did you become interested in CS and AI?
I learned to program in Basic in my early teens; I had an uncle who worked for GE on their time-sharing system. I learned Fortran in high school. I tried to build my own adding machine out of TTL chips around that time too. However, despite this interest, I didn’t really know what CS was until I reached graduate school at the University of Illinois. I first engaged with AI when I took a graduate assistant position with Ryszard Michalski on what became machine learning, and I took an AI class from Dave Waltz. I had also studied phi- losophy of science in college, so I had already thought a bit about how we acquire knowledge from data and experiment.
What would you have chosen as your career if you hadn’t gone into CS?
I had considered going into foreign service, and I have always been interested in policy issues. I might also have gone into technical management. Both of my brothers have been successful technical managers.
What do you wish you had known as a Ph.D. student or early researcher?
I wish I had understood the importance of strong math skills for CS research. I was a software engineer before I was a computer science researcher, and it took me a while to understand the difference. I still struggle with the difference between making an incremental advance within an existing paradigm versus asking fundamental questions that lead to new research paradigms.
What professional achievement are you most proud of?
Developing the MAXQ formalism for hierarchical reinforcement learning.
What is the most interesting project you are currently involved with?
I’m fascinated by the question of how machine learning predictors can have models of their own competence. This is important for mak- ing safe and robust AI systems. Today, we have ML methods that give accurate predictions in aggregate, but we struggle to provide point-wise quantification of uncertainty. Related to these questions are algorithms for anomaly detection and open category detection. In general, we need AI systems that can work well even in the presence of “unknown unknowns”.
Recent advances in AI led to many success stories of AI technology undertaking real-world problems. What are the challenges of deploying AI systems?
AI systems are software systems, so the main challenges are the same as with any soft- ware system. First, are we building the right system? Do we correctly understand the users’ needs? Have we correctly expressed user preferences in our reward functions, constraints, and loss functions? Have we done so in a way that respects ethical standards? Second, have we built the system we intended to build? How can we test software com- ponents created using machine learning? If the system is adapting online, how can we achieve continuous testing and quality assurance? Third, when ML is employed, the re- sulting software components (classifiers and similar predictive models) will fail if the input data distribution changes. So we must mon- itor the data distribution and model the pro- cess by which the data are being generated. This is sometimes known as the problem of “model management”. Fourth, how is the deployed system affecting the surrounding social and technical system? Are there unintended side-effects? Is user or institutional behavior changing as a result of the deployment?
One promising approach is combining humans and AI into a collaborative team. How can we design such a system to successfully tackle challenging high-risk applications? Who should be in charge, the human or the AI?
I have addressed this in a recent short paper (Robust Artificial Intelligence and Robust Human Organizations. Frontiers of Computer Science, 13(1): 1-3). To work well in high- risk applications, human teams must function as so-called “High reliability organizations” or HROs. When we add AI technology to such teams, we must ensure that it contributes to their high reliability rather than disrupting and degrading it. According to organizational researchers, HROs share five main practices: (a) continuous attention to anomalous and near-miss events, (b) seeking diverse explanations for such events, (c) maintaining continuous situational awareness, (d) practicing improvisational problem solving, and (e) delegating decision making authority to the team member who has the most expertise about the specific decision regardless of rank. AI systems in HROs must implement these five practices as well. They must be constantly watch- ing for anomalies and near misses. They must seek multiple explanations for such events (e.g., via ensemble methods). They must maintain situational awareness. They must support joint human-machine improvisational problem solving, such as mixed-initiative plan- ning. And they must build models of the expertise of each team member (including them- selves) to know which team member should make the final decision in any situation.
You ask “Who is in charge?” I’m not sure that is the right question. Our goal is to create human-machine teams that are highly reliable as a team. In an important sense, this means every member of the team has responsibil- ity for robust team performance. However, from an ethical standpoint, I think the human team leader should have ultimate responsibil- ity. That task of taking action in a specific situation could be delegated to the AI system, but the team leader has the moral responsibility for that action.
Moving towards transforming AI systems into high-reliable organizations, how can diversity help to achieve this goal?
Diversity is important for generating multiple hypotheses to explain anomalies and near misses. Experience in hospital operating rooms is that often it is the nurses who first detect a problem or have the right solution. The same has been noted in nuclear power plant operations. Conversely, teams often fail when the engage in “group think” and fixate on an incorrect explanation for a problem.
How do you balance being involved in so many different aspects of the AI community?
I try to stay very organized and manage my time carefully. I use a machine learning system called TAPE (Tagging Assistant for Productive Email) developed by my collaborator and student Michael Slater to automatically tag and organize my email. I also take copi- ous notes in OneNote. Oh, and I work long hours…
What was your most difficult professional decision and why?
The most difficult decision is to tell a PhD student that they are not going to succeed in completing their degree. All teachers and mentors are optimistic people. When we meet a new student, we hope they will be very successful. But when it is clear that a student isn’t going to succeed, that is a deep disappointment for the student (of course) but also for the professor.
What is your favorite AI-related movie or book and why?
I really don’t know much of the science fiction literature (in books or films). My favorite is 2001: A Space Odyssey because I think it depicts most accurately how AI could lead to bad outcomes. Unlike in many other stories, HAL doesn’t “go rogue”. Instead, HAL creatively achieves the objective programmed by its creators, unfortunately as a side effect, it kills the crew.
Nominations, including self nominations, are invited for a three-year term
as TALLIP EiC, beginning on June 1, 2019. The EiC appointment may be renewed at most one
time. This is an entirely voluntary position, but ACM will provide appropriate
Appointed by the ACM Publications Board,
Editors-in-Chief (EiCs) of ACM journals are delegated full responsibility for
the editorial management of the journal consistent with the journal’s charter
and general ACM policies. The Board relies on EiCs to ensure that the content
of the journal is of high quality and that the editorial review process is both
timely and fair. He/she has final say on acceptance of papers, size of
the Editorial Board, and appointment of Associate Editors. A complete list of
responsibilities is found in the ACM
Volunteer Editors Position Descriptions. Additional information can be
found in the following documents:
Nominations should include a vita along with a brief statement of why the
nominee should be considered. Self-nominations are encouraged, and should
include a statement of the candidate’s vision for the future development of TALLIP.
The deadline for submitting nominations is April 15, 2019, although nominations
will continue to be accepted until the position is filled.
A recent item in Science|Business “Artificial intelligence nowhere near the real thing, says German AI chief”, by Éanna Kelly, gives policy-worthy warnings and ideas. “In his 20 years as head of Germany’s biggest AI research lab Wolfgang Wahlster has seen the tech hype machine splutter three times. As he hands over to a new CEO, he warns colleagues: ‘Don’t over-promise’. … the computer scientist who has just ended a 20 year stint as CEO of the German Research Centre for Artificial Intelligence says that [warning] greatly underestimates the distance between AI and its human counterpart: ‘We’re years away from a game changer in the field. I always warn people, one should be a bit careful with what they claim. Every day you work on AI, you see the big gap between human intelligence and AI’, Wahlster told Science|Business.”
For AI policy, we should remember to look out for over promising, but we also need to be mindful of the time frame for making effective policy and be fully engaged now. Our effort importantly informs policymakers about the real opportunities to make AI successful. A recent article in The Conversation by Ben Shneiderman “What alchemy and astrology can teach artificial intelligence researchers,” gives insightful information and advice on how to avoid being distracted away “… from where the real progress is already happening: in systems that enhance – rather than replace – human capabilities.” Shneiderman recommends that technology designers shift “from trying to replace or simulate human behavior in machines to building wildly successful applications that people love to use.”