GMU's Harry Wechsler

October 31, 2005
Technology Research News Editor Eric Smalley carried out an email conversation with Harry Wechsler, Professor of Computer Science and Director of the Distributed and Intelligent Computation Center at George Mason University.

Wechsler's research centers around making computers more intelligent by giving them the ability to recognize patterns. Pattern recognition technologies involve distinguishing patterns -- the distribution of pixels in images, words in documents or numbers in raw data. These technologies also involve recognizing specific instances or types of patterns -- faces, phrases or statistically significant variations. Training computers to recognize patterns often involves learning by example.

Wechsler's research touches on computer vision, automatic target recognition, signal and image processing, statistical learning theory and support vector machines, neural networks, machine learning, information retrieval, data mining and knowledge discovery, genetic algorithms and animats, face and gesture recognition, biometrics, and forensics.

He is an IEEE fellow and has directed NATO Advanced Study Institutes and Co-Chaired the International Conference on Pattern Recognition. He has authored more than 200 scientific papers, written a book titled Computational Vision, and edited books on neural networks and face recognition.

Wechsler has held visiting professorships at Institutes in France, Japan, Israel, Australia, Switzerland, Russia, and China, consulted for government agencies including the National Institutes of Health, Naval Research Laboratory, and NASA, and consulted for companies including Honeywell, Lockheed-Martin, Procter & Gamble, and 3M.

He received his Ph.D. in Computer Science from the University of California, Irvine, in 1975.

TRN: What got you interested in science and technology?

Wechsler: [I] was always interested about the brain and how it works. [I] started also to play chess at a young age -- 5 yrs. -- never made it to be a champion -- and loved to solve puzzles.

[I] still think that kids should be exposed to games. The earlier the better. They will learn reasoning... strategy, and responsibility.

TRN: What are the best games out there today, and why?

Wechsler: High on my list is chess for obvious reasons. Then comes Go, a game I never played but wished I did. It's much more complex than chess and combines reasoning and pattern recognition.

Video games are also good choice (within limits... due to violence) because [they] require hand eye coordination, speed, and (team) planning.

TRN: What are the important or significant trends you see in science and technology research?

Wechsler: On the positive side, pervasive computing and communication -- everywhere and everytime, 24/7, global village -- the drawback is about security, lack of privacy and 1984.

TRN: Tell me more about the positive significance of pervasive computing.

Wechsler: You will not have to remember many things -- people, appointments, medical condition (pervasive sensors and agents will travel inside and outside to monitor us), how to deal with unforeseen circumstances. But beware of flying by the wire (see Airbus). And make sure that somehow people have the last word.

TRN: Tell me more about the downside.

Wechsler: Privacy and security for obvious reasons. Delegate decisions that people might come to regret.

TRN: Do you see any practical ways these can be balanced?

Wechsler: Important decisions are explained to and require final approval from people.

TRN: Tell me about the trends in pattern recognition research. What are the pluses and minuses of these technologies as they exist today?

Wechsler: Not much different from 30 - 40 years back. Some of the big news, e.g., statistical learning theory and support vector machines (SVM) owe their existence to research done in the 60s.

TRN: What do you see as the most urgent needs in pattern recognition research?

Wechsler: Experimenting on real and large amounts of data, truth in reporting and competitive evaluations, take advantage of the temporal dimension and consider change and drift, and allow for noise, occlusion, and distortions. [I] see an important role for e-science.

TRN: These seem like major challenges. What needs to be done, and what will the practical effects be?

Wechsler: Such experimentation is more complex and more difficult. [It] will take time for people to get used to such scenarios and run large scale experiments (on hundred sequences rather than just a few) (and using real CCTV scenarios).

Look at what was done after 9/11 and recently in London. They went to find the perpetrators "after" the events and using mostly manual means. This is "state-of-the-art" and is not good enough. What needs to be done is to seek and accrue evidence in an intelligent fashion in order to reduce uncertainty, better learning methods, and taking advantage of spatiotemporal coherence.

TRN: One argument for performing experiments with canned data sets is that they provide benchmarks for comparisons. On the other hand, they often differ significantly from real world experience.

Is the pattern recognition research community relying too much on standard data sets, or are there practical issues that simply make it difficult to use real world data?

Wechsler: [It] is easier to pass tests on known sets. You are right. Testing should be on unseen data and should lack ground truth information.

Most people still do not get the difference between a priori and a posteriori threshold settings. Much of pattern recognition (PR) requires decision thresholds. Most people set them after completing the experiments based on ground truth knowledge and report best results achieved using the threshold that yields them.

Real operation requires learning both the classifier and the threshold ahead of time. Other tricks that bias performance are those using score normalization (even that NIST forbids them on speech competition).

TRN: What is "ground truth information"?

Wechsler: Ground truth means that I know what it is. Say you want to classify items. When you build your statistics and have to choose decision thresholds you know what label is associated with each class during testing. So you choose the threshold that yields best result.

In real life one does not have the labels and has to come prepared with a decision threshold based only on training and validation data.

TRN: Where does e-science [the use of grid computing and other distributed resources] come in?

Wechsler: One can try her latest "magic bullet" using the infrastructure already in place -- no need to build everything from scratch. [You] can play the game "what if" when you substitute one module for another one.

TRN: What are the issues here? Do you have concerns about the quality of pattern recognition research?

Wechsler: [Jean] Ponce, 2003: “Despite 40 years of research, however, today’s recognition systems are still largely unable to handle the extraordinary wide range of appearances assumed by common objects in typical images.”

TRN: Tell me about statistical learning theory and support vector machines. How have they changed pattern recognition technology?

Wechsler: [They] made the tradeoff between performance and complexity concrete using prediction risk. [It] defines "margin" as new criteria for optimal performance (now + future) and relates it to Popper's definition of what makes science. [It] can be refuted.

TRN: What is "prediction risk"? What do you mean by "margin" here?

Wechsler: Training and validation come first and determine what is known as empirical risk (R(emp)).

What performance should be expected during testing? Usually R(emp) + alpha, where alpha is kind of confidence interval.

The more complex "h" the classifier is, the more likely that there is overfit. The smaller number "n" of examples you trained with, the less likely you learned the correct decision boundary. So alpha [is explained by] f (h/n), where f is kind of penalty factor, and is expressed using a monotic increasing function. There is more to this story but those are the basics.

The margin is the area separating... two classes. [It's] better to think that the boundary is surrounded by no man's land where everything can happen. The larger the no man's land, where there is no commitment, the less likely the classifier will fail during testing. Margin can be also thought as margin of error.

Rather than having a sharp boundary you leave room for future data to fall on the sides of the no man's land and thus not be wrongly classified.

TRN: Research on giving machines the ability to accurately perceive their surroundings has advanced considerably in recent years but remains a major challenge. What will it take to build machines that can operate effectively in unfamiliar, dynamic environments?

Wechsler: MIT thought back in the 60s that will take [a masters student] to solve the problem over the summer. We are not closer today to the solution than we were then. Maybe the computer vision community should start from scratch and revisit their assumptions and operation mode. Maybe is too difficult and one should not expect that 40 years of research can achieve what nature and evolution did over millions of years.

TRN: Why were people unable to see how difficult a problem this is?

Wechsler: Maybe because much of this started with AI folks that never heard about signals and noise.

TRN: Machine perception and pattern recognition technologies are increasingly applied to problems of tracking and understanding human behavior. What are the social and economic implications of these technologies?

Wechsler: Improved HCI on one side. [It would be] better for medicine for sure and maybe good for education and dialog. [It would be] bad for privacy and security.

TRN: How will pattern recognition technologies be good for medicine and education?

Wechsler: [Look at] personalized medicine, sequencing, and microarray analysis. WSJ just reported a few days ago about some company [454 Life Sciences] in [San Francisco] that can speed up sequencing and makes personalization a reality.

TRN: How will they be bad for privacy and security?

Wechsler: [They] will know too much about us and what we do. [They] can follow us on roads and keep track of everything we do and with whom.

TRN: How should the downsides be addressed?

Wechsler: Information on people should be discarded or not even collected unless there is a court order to do that.

TRN: Where are we in the evolution of artificial neural networks? Are there technologies on the horizon that could give rise to artificial brains that are in any way comparable to biological brains?

Wechsler: I think that we are again at a dead end in neural networks. We don't know what the brain does and how it does it. So how can we duplicate something that we still don't understand?

TRN: I'm hoping you can elaborate on how the brain is a mystery. Surely we know something of how the brain works. And aren't there already practical applications of neural network technology?

Wechsler: Something, yes, but very little of the whole.

The practical applications mostly owe their existence to incorporating advance knowledge about sensors, signals, and systems. One can't say much about backprogation -- [it] is widely used but can't prove convergence and is not incremental. [It] needs to be [retrained] each time the repertoire of facts changes.

TRN: Can you describe for the layperson what "backpropagation" is?

Wechsler: [A] stick and carrot like mechanism [that] learns to map inputs to their correct outputs, [like] class label.

You have a multi-layer network that operates in a feed-forward mode from the input layer (where the unknown pattern is presented) through hidden layers (like in a black box) to the output layer.

Each layer is made up of node units that can compute. The nodes are connected using synaptic weights. Learning means changing the synaptic weights so most of the time inputs map correctly to outputs. This is what [a] backpropagation learning algorithm does.

TRN: What are the possibilities and limits of data mining, and what are the social and economic implications of using the techniques you and others are developing?

Wechsler: Major potential [in] individualized and predictive medicine (micro arrays). The social implications are grim. Machines will know about us, our health, and maybe our thoughts. They will try to sell us things -- see recent brouhaha over neuroeconomics, deny health care, and make life and death decisions on our behalf.

TRN: That's a rather bleak vision. How serious is this risk, what kind of timeframe are you considering, and are there technological solutions to the problem?

Wechsler: There is always somebody lurking in the background to misuse progress in science and technology. All what I am saying is that parallel to such developments there is great need for awareness and for means to combat misuse. See the Internet -- is great and will be very difficult to live without.

Still we have to develop means for protection against phising et al. Progress can be good or can be bad, depend[ing] on its use. It has always been like this and will always be the same.

TRN: How do the technologies you're working on relate to business, culture, and social life?

Wechsler: My face recognition research helps with biometrics and security, on one side, and pattern recognition, data mining, and machine learning, on the other side. [It] is also about human nature and culture because the human face occupies such a preeminent role in arts, law, and social life.

TRN: What are the important social questions related to today's cutting-edge technologies?

Wechsler: What is the price paid for change and outsourcing? Cut throat competition and the real possibility now to become obsolete so fast. Almost all is about technology and very little about society. Many technologies are yet brittle. Software is of poor quality.

TRN: What can be done -- government policy, education, socioeconomic initiatives -- to address these problems?

Wechsler: Real competition and less government intervention.

TRN: In terms of technology and anything affected by technology, what will be different about our world in five years? In 10? In 50? What will have surprised us in 10 years, in 50?

Wechsler: Everything will be connected in 10-20 years and at our fingerprints.

[It] will surprise me if computers will show same imagination as people and make breakthroughs, e.g., an Einstein, in science and technology, or find cures for diseases like cancer.

In social life [I] will be very surprised if faster and more intelligent computers will make people suddenly peaceful and make war a thing of the past.

Last but not least is always the possibility to encounter an alien civilization.

TRN: What will become more complicated 10 and 50 years down the road, and what will become simpler?

Wechsler: Life will be more complicated and regulated. Forget simplicity and privacy.

TRN: What do you imagine you or your successor will be working on in 10 years? In 20 years?

Wechsler: Similar problems. Maybe breakthrough in nanotechnology, energy (fusion), medicine (health), and brain sciences.

TRN: What's the most important piece of advice you can give to a child who shows interest in science and technology?

Wechsler: My children are 13 and 15 so this is a real problem for me. My advise is to study as much and as broad as you can. Learn to be flexible and keep in mind that learning never ends.

Learn math -- the universal language -- learn about human behavior, culture and languages, and social sciences. Maybe economics -- see recent Nobel prize awards -- rather than computer science alone.

TRN: Why not?

Wechsler: Too narrow. We now try to combat declining enrollments here with combined degrees, e.g., applied CS and biology or applied CS and history or geography (see GIS).

Sciences and liberal arts need computing. We can expand the pool of potential CS students and literate a larger population and eventually work force.

TRN: What are your thoughts on the state of the general public's conventional wisdom on science and technology?

Wechsler: [The] public is mostly uneducated. About 50% believe in intelligent design and its variations. [The] public enjoys being a greedy consumer.

TRN: What can we do about this?

Wechsler: Redesign those people using some intelligent design :)

TRN: What could be done to improve the pursuit of science and technology research in terms of business trends, politics, and/or social trends?

Wechsler: Improve education and start at an earlier age. Establish high SOL [Standards of Learning] standards. Learn from other countries.

TRN: Do you have some good examples in mind?

Wechsler: Asia seems to be ahead. [It] spends more time and has better teachers. There is much shortage here of qualified science teachers. Compensation and morale are low -- discipline, et al.

TRN: What books that have some connection to science or technology have impressed you in some way, and why?

Wechsler: Freakonomics, Blink. [They] show how science and technology can address some of our most important issues

TRN: What other readings do you recommend that would bring about more interest and/or a better understanding of science and technology?

Wechsler: Scientific American National Geographic, Discovery.

TRN: Is there a particular image (or images) related to science or technology that you find particularly compelling or instructive?

Wechsler: Civilization has a long history. People 2000-3000 years back were not less smart than we are today. Compelling and instructive? Go to Monticello and see Jefferson's house and his library. Compare with today's political class.

TRN: What are your interests outside of work, and how do they inform how you understand and think about of science and technology?

Wechsler: Arts, travel, and family -- raising my children in love and appreciation of learning.

TRN: What question would you like to be asked in an interview like this? What is the answer to that question?

Wechsler: How do we make our society a better one? Where most of us have access to good quality education, among other things.

Answer: educate people to appreciate education for its own rather than gadgets, money, and power.

Last            Next

Discuss this interview or anything else to do with science and technology on the TRN Forum.

TRN needs your help. Please click here for details.



Research News Roundup
Research Watch blog

View from the High Ground Q&A
How It Works

New: TRN's Jobs Center

RSS Feeds:
News  | Blog  | Books 

Ad links:
Information on RFID, VoIP, VPN, ZigBee and more

Buy an ad link


Ad links: Clear History

Buy an ad link

Home     Archive     Resources    Feeds     Offline Publications     Glossary
TRN Finder     Research Dir.    Events Dir.      Researchers     Bookshelf
   Contribute      Under Development     T-shirts etc.     Classifieds
Forum    Comments    Feedback     About TRN

© Copyright Technology Research News, LLC 2000-2006. All rights reserved.