Tom M. Mitchell founded and chairs the Machine Learning Department at Carnegie Mellon University, where he is the E. Fredkin University Professor. His research uses machine learning to develop computers that are learning to read the web, and uses brain imaging to study how the human brain understands what it reads. Mitchell is a member of the U.S. National Academy of Engineering, a Fellow of the American Association for the Advancement of Science (AAAS), and a Fellow of the Association for the Advancement of Artificial Intelligence (AAAI).
See Carnegie Mellon University www.cs.cmu.edu/~tom
To listen to the interview, click on this MP3 file link
:00:22: How can computers learn to read the web? "....We have this research project going which we call our Never-Ending Language Learner (or NELL for short), and the goal here is to build a program that runs 24 hours a day 7 days a week and everyday it has two things it has to do. One is that it has to extract some more facts from the web by reading (facts like 'Obama is president' or 'blue jeans are often worn with t-shirts'), and the second task it has to do each day is to learn to read better than it could the day before so that tomorrow it can collect some more facts more accurately....We started this program running 24 hours a day January 2010 so it's about 4 years old now and the results so far is that it is indeed learning...."
:02:28: So from a pure belief and language processing perspective would you say that it's better than a 3 year old? "....They have (in a funny way) a more complete factual set knowledge than a normal three year old, but they still don't have that kind of common sense that a 3 year old has...."
:03:15: What’s the underlying hardware platform for this Never-Ending Language Learner (NELL) project? "....The hardware is a collection of about a hundred computers that Garth Gibson (another faculty in our department here), runs as a research project and he lets us use that collection to run NELL....Maybe more interesting than the hardware is how the learning itself works and I could take a few minutes to explain that...."
:09:50: How does the human brain represent word meanings? "....Let's say I give you the word 'mom' or 'computer'....it's really the same brain and the only difference is that when I say the word 'mom' it was one pattern of neurons firing and when I say 'computer' it is a different set of neurons firing in your brain. We've become very interested in how the brain processes natural language and for the past 12 years with my colleague Marcel Just in our psychology department, we've been studying brain imaging with fMRI and MEG brain imaging, the patterns of neural activity that occur in people's brains when they read. We very much took a machine learning approach to this problem....For me the fun thing is that it starts to tie together the research that we’re doing on NELL of trying to understand how computers can learn to read, and it ties it together with the studies that we're doing of representations in the brain in terms of neural activity when people read. It's a direction we are pushing on very hard now to put these two research projects even more closely together...."
:18:21: Long term there are some practical applications of this, but I guess the problem is that you are averaging based on typical neural patterns and if you have dysfunction, like a brain injury, then those neural patterns would get distributed. Have you done any work on that to see if it would still be applicable for people with head injuries? Also can you take this work and apply it to the animal world? Can your work be extrapolated to other types of brain structures, for example crows (corvids), where the intelligence is located elsewhere (nidopallium caudolaterale)? "....I think that is a really fascinating question. We haven't done that work yet....You also asked about longer term applications. If we can get to an understanding of how the brain reads then I think that we will in the future provide a lot of opportunities; for example understanding dyslexia and different aphasias, different language disorders in a much more accurate way....We're not there yet, but I would characterize the work we're doing as the underlying basic science that will hopefully provide the scaffolding for that...."
:21:05: You are doing some very fascinating work and there will be some philosophical extensions to this work. Do you get into some of these philosophical discussions at all? "....I'm fascinated by these discussions, but I'm frustrated by the fact that things like consciousness are not directly observable to the physical instruments that we have available. I read a paper that changed my thinking a little bit about this and I'll briefly describe some work by a researcher named Stan Dehaene in France; he was interested in using brain imaging to study awareness....At least in the case of reported self-awareness or awareness of the external stimuli, Stan has shown us how to devise experiments that can distinguish those so maybe there’s a future of much more interesting lines of work studying things like awareness and what are the neural mechanisms that underlie them...."
:24:40: More generally, where is Cognitive neuroscience heading? "....I think the future for Cognitive neuroscience in the coming decade is more of that kind of work where we don't just look at the neural data, but we simultaneously build computer programs that can do those same things that we're asking the brain to do, and then we look at the computer process itself (the sequence of processing steps that was required in order for the computer to solve the algebra problem etc.) and that becomes a model for what the brain is doing. For me that's the exciting opportunity in the coming decade...."
:32:54: When will we understand the brain and what does it mean to understand the brain? "....I think the best answer I can come up with for the question of 'when we understand the brain what will the answer look like', is that the answer will look like an artificial intelligence program that functionally performs very similarly to the brain and is organized by some design principles and those design principles will also be part of our theory of the brain....For me that would be a large part to have a theory of understanding how the brain really works. So it's very closely tied to computer science (not a distinct field) and it's very closely tied to artificial intelligence. AI and Cognitive neuroscience will become increasingly close over the next decade...."
:35:04: What is the current state of brain research and the challenges to this research? "....To me what's probably going to be the two main drivers for Cognitive neuroscience will be our ability to come up with even better instruments and to also take advantage of the growing volume of experimental data that exists across labs in the world. The second challenge is really to come up with better computer simulations of the cognitive processes themselves. That's the task that will be shared between Cognitive neuroscience and AI...."
:43:33: There's all of this work going on both at a micro and a macro level that you talked about, but ultimately it comes down to individuals. What are the implications of all of this happening and for example, what are the future coming advances in mining personal data? "....I think the most obvious upcoming change will be in our mobile devices, because personalization in mining personal data is partly determined by what algorithms you use to mine that data, but it is even more determined by what data is available. And your cellphone has the opportunity to collect more data about you than any other thing on earth....I think learning your personal habits is one part of it but another equally important part is just the growing competence of programs in terms of perception, speech, text, vision and others, and planning and cognitive competence...."
:48:13 There's a lot of controversy in terms of agencies mining data of governments and of people and even this ability to turn on webcams without you even realizing that it has been turned on. What does that mean to personal privacy? "....It's really a social question, a political question and a society question that we all need to get engaged in....The question of whether they should or should not collect any type of data is one question, but a distinct type of question is should they or should they not allow society to make the decision about whether they collect that data...."
:50:25: I guess we could extrapolate on what you talked about earlier. Let's say that we have 10⁵ higher resolution on fMRIs and the same kind of increase in resolution but with a temporal aspect that you talked about with MEGs and a Cloud-based AI system that can sort of reason what's happening, so now we have walking entities with embedded devices and we can read each other's thoughts and anticipate what they're thinking. What does that mean? "....To be honest that's pretty far off unless there's some miracle instrument devised. Right now the fMRI scanner that we use weighs about 10 tons so you're not going to walk around with this. Similarly the MEG scanner is a super cool magnet and you have to sit in a chair with your head stuck inside (it's not a portable device), so we won't have people walking around with that in the near term....I think one of the most underestimated upcoming disruptive technologies is really that computers will learn to read...."
:52:01: Tom talks about machines learning to read. "....When computers learn to read like that they won’t be constrained the way you and I are. They are very fast and they would be able to read the entire web, hundreds of billions of webpages out there. Once they learn to read, computers will be more well-read than we ever could be. That's a very disruptive technology because right now the web is the world's largest knowledge base for human consumption. We write it in text for each other and right now it is inaccessible to computers. Once computers learn to read, the web will become the world's biggest knowledge base accessible to computers. What will we want to do then?...."
:56:47: There is a predictive analytics portion of this. For example there are various models of the stock market and you can pick the leading one and say 'examine all the world trends in terms of these 60 variables and give me some kind of conclusion so what's emergent'? Or it's like with Google, Bing or some of the search engines, you can predict activity that's going to occur in disease just by emergent types of news occurring in a particular region. So what does that mean to the economy? "....Like you say maybe we'll be able to predict for the world economy. I guess the implication for the world economy is that the economists would probably say it may lead for a more productive economy because if you can see what's coming next you can plan for it more efficiently and produce what needs to be available for that upcoming future instead of being in the dark about it....I think the question that you raise leads immediately to this question of access to that data and whether that will be private or public and that will be a very important question for us all...."
:01:01:16: Throughout this interview you talked about machine learning, but there's people in the audience who may not fully realize what machine learning is. Can you talk more generally about what is machine learning, where it's headed and how it will impact us all? "....Machine learning tries to answer the following question, how can we build computer programs that automatically improve through experience? A spam filter is a good example of that....Machine learning doesn't just mean using historical pre-collected data, it can also mean 'as you go' through the world or over time using what you now understand so you can go back and train yourself so that in the future you will be able to anticipate better. Roughly machine learning is about getting the computer to improve from experience...."
:01:04:58: How do you feel as a leading proponent of machine learning and somebody who is recognized worldwide as a thought leader and researcher in this area? You must feel like a kid in a candy shop? "....Machine learning is just the current answer to the next natural question about computation which is 'what about those things that we can't write down the recipe'?....I think machine learning has a very bright future and I don't think machine learning is a solved problem by any means. I think we are still in the very early young days of the field of machine learning and when we look back in 25 years at the algorithms that we're using today we will smile in a bemused way at how primitive some of them are..."
:01:10:52: Taking a broader view, where is Artificial Intelligence (AI) heading and why should the public care?? "....I find it hard to think about Artificial Intelligence at a technical level independent of thinking about machine learning. It's become an important component of artificial intelligence....Some would say the artificial intelligence defining question is how do we build computers that are as intelligent as we are....I think one way to think about the future of artificial intelligence is to give some thought to what kind of assistance would we like and what are the feasible types of assistance that we might be able to get computers to provide...."
:01:13:52: Where is computer science heading? "....Maybe one way to see where it's heading is to see how our perception of computer science is different to what it was in the early days. In the early days we thought of computer science as a field that was about getting machines to calculate things. Now when we think about computer technology we think as much about communication as we think of calculation...."
:01:21:43: What are your views on Judea Pearl's work, especially his most recent work and the implications to what you are doing? "....I think it is great work and very interesting because one of the key difficulties in machine learning is that machine learning programs are quite good at finding correlations between variables, but quite bad at determining which of those correlations are actually causal....One of the challenges for machine learning is really to move on beyond correlation to be able to infer causation too. I think the theoretical work and formalisms that Pearl and others are developing are really great in that direction. Sadly it turns out I think that while those formalisms give us a better handle on what it means to infer causality, they are not finding easier ways to infer causality they are just so far telling us more precisely what it means to infer causality and just how difficult it is to do it in terms of the data that you have to have available. But it's great research and it's an important fundamental direction for the whole field..."
:01:27:05: Do you have a view of Kurzweil's singularity? Isn't there a tie-in with the work you are doing? "....I take this singularity hypothesis as something like there will come a time when computers are more intelligent than people and what's going to happen then?....The whole idea that there is a single question to ask (are they smarter than us or dumber than us?) is I think misguided and we should replace the singularity notion by multilarity and just say it's not an all or nothing or one or zero kind of question. There's a whole vector of different competences....I don't think computers will get stupider over time and I don't think people will either, but I do think that computers will get better than us at more of those things and I think that's the more interesting way to think of it. I think it's a disservice to the discussion to pretend that there is a single notion of computers being more intelligent than we are...."
:01:32:02: When will robots have free will and what does free will mean (in the sense of a human being)? "....I have a problem with that. Computers make decisions and we can give them the ability and the command to make certain types of decisions....I'm not sure how to precisely define free will, but I am sure how to think about giving computers the responsibility to learn to achieve goals, and then the difference between an explicit and fixed goal for it to learn to achieve versus the other decision that it will have to make and learn to make well in order to achieve those goals..."
:01:34:34: Can you profile your top research successes and what value they provide to research, and the broader implications/applications to the broader business and computing industry? "....At a technical level within the field of machine learning one of the results that we've come up with that I'm most happy about is the idea of how to get computers to learn from unlabeled data. That's a very useful idea both theoretically and in practice for how to use unlabeled data to achieve better accuracy in learning....I think that the work that we've been able to do introducing machine learning as a method for analyzing brain image data is a result that I'm also happy with both in terms of what it's been able to tell us about fundamental questions about how the brain represents meanings of words, but also just the idea that machine learning could be a useful way to practice Cognitive neuroscience...."
:01:44:09: Predict what you consider to be your future successes and how they will be realized to the general audience? "....It's hard to predict success but I can tell you what my 5 year goal is. My 5 year goal is to put together what are now these 2 distinct research projects that my teams are working on. One is in the area of brain science in human language processing, the other is in the area of computer science in machine learning studying how get a computer to process language. These 2 projects grew kind of independently, but my 5 year goal is to have them integrate and the result should be a computer program that both understands the text that you show it and as a side effect of that predicts the neural activity in the brain when the person reads that very same text...."
:01:48:19: Can you profile some of your past and current leadership roles and what value they provide to research, and the implications/applications to the broader marketplace, business and industry? "....One of the things that I enjoy very much about chairing the Machine Learning department here is that it gives me an opportunity to interact with people....I've been able to learn a lot in this position, not only about the different technical and scientific problems that people in our department are working on, but just as importantly what are the things that we are doing and not doing that are of most interest to people outside of the department who are intrigued in what we’re doing and interested in our students...."
:01:53:37: You have this leadership role as chair of this very important department and you are one of the leading experts in the world, noted researcher and authority. Do you ever think about commercialization of that work? "....I have been and I am involved in small ways in a number of companies and at the same time I decided some time ago that I'm an academic at heart. For me I really enjoy the life of the university in part because even though you can do wonderful things with this technology in companies (and I admire people who are doing that), there are benefits of doing work at a university....It's just a very different thing. I love the research part and the idea of brainstorming with people and communicating the way you do around the university...."
:01:55:48: You are a leading authority in the world in your field and there are implications to your field by economics, genomics, robotics, sociology, etc. so from your perspective, what are the best online resources to use? "....There is a wonderful course Machine Learning offered by Coursera taught by Andrew Eng....The Carnegie Mellon Machine Learning Department website....My own webpage (a video course that I've put online, videoed from a course that I teach here at Carnegie Mellon)....videolectures.net (has literally thousands of more technical videos of computer scientists presenting work but a large fraction of that is Machine Learning)...."
:01:58:18: What are your views on policy changes that are needed by the government? "....I think there is a huge opportunity for society to benefit from this growing volume of data that's available online and that isn't just in terms of a single application, but across many aspects of society....To put it in a phrase, 'Data will become wealth'; and we need discussion and new thinking, and we need to begin the discussion about how to deal with this future where (unlike in the past), a large part of wealth creation will surround who collects and who owns and who has access to the data and how we deal with the privacy issues....I'm not sure it's a policy change at this point, but more of an awareness and discussion at the society level...."
:02:05:14: We talked about a lot during this interview, are there any disruptive changes that are we missing? "....We are getting more and more reliant on computation and our communication infrastructure and it is rather amazing to think about. If for example, the power was out for a week or we had a meltdown of computational resources at some point...."
:02:06:32: Do you feel computing should be a recognized profession on par with accounting, medicine and law with demonstrated professional development, adherence to a code of ethics, personal responsibility, public accountability, quality assurance and recognized credentials? [See www.ipthree.organd the Global Industry Council, http://www.ipthree.org/about-ip3/global-advisory-council] "....I can see benefits of having that kind of professional code especially getting to some of the ethical questions. As a rule, without that code many of us have the opportunity to just ignore the ethical questions (and they are with us), so one advantage of having such a professional code is that it focuses our attention on those ethical questions and gets us to come to some consensus...."
:02:07:36: Tom shares a story from his extensive speaking, travels, and work, (perhaps something amusing, surprising, unexpected, amazing). "....This was to one of the more memorable humorous things and Saul and I laughed about that many, many times afterwards. He had forgotten that story. For me that is more about my fondness for Saul and the importance that his mentorship was to me...."
:02:10:40: If you were conducting this interview, what question would you ask, and then what would be your answer? "....I think I would ask one more question....What's your advice to new graduate students and researchers in computer science?...."
:02:14:45: Tom, with your demanding schedule, we are indeed fortunate to have you come in to do this interview. Thank you for sharing your deep experiences with our audience.