Abstract

Tomaso Poggio is Eugene McDermott Professor in the Department of Brain and Cognitive Sciences at MIT, where he is also member of the Computer Science and Artificial Intelligence Lab and of the new McGovern Institute for Brain Research. He received his doctorate in physics in Italy, then worked for ten years in the Max Planck Institute for Biological Cybernetics in Tübingen, Germany, before moving to MIT where he has been for the last 23 years. His papers span mathematics, computer science and neuroscience. His main contributions are in computer and biological vision (in both flies and humans), computational neuroscience, system theory and computer graphics. He currently leads a group of researchers investigating the problem of learning in terms of the underlying mathematics, its engineering applications and its role for understanding object recognition in visual cortex.

Why did you study physics? As a kid I was fascinated by the problem of intelligence. What made Einstein such a genius? I wanted to know what is intelligence, how to increase it and how to build intelligent machines. I thought about studying biology on the way to brain science but at the time in the University of Genoa in Italy, biology was little more than zoology. I opted for getting a mathematics background, and physics, which has always had a good tradition in Italy and excellent standards, won the day against engineering and mathematics.

Do you have any favourite papers? Yes, though they are not very original I am afraid. I am fascinated by Einstein's Gedankenexperiment in the section ‘Über die Relativität von Längen und Zeiten’ of his first paper on the theory of relativity, and I admire the elegance and brevity and historical importance of Crick and Watson's famous 1953 Nature paper ‘Molecular structure of nucleic acids; a structure for deoxyribose nucleic acid’, which marked the beginning of molecular biology.

Which of your papers are you most proud of? A recent one exploring the foundations of predictivity: what are the conditions that ensure that a learning algorithm will generalize? These are general and deep questions: what does it mean to predict the future? They are also closely related to the question of what makes a theory a scientific theory. Last but not least, for once I was able to work on many of its details myself with pencil and paper!

What is the best advice you've been given? Werner Reichardt, who founded the Max Planck Institute for Biological Cybernetics, where I spent the first ten years of my scientific life, taught me that, in neuroscience, models and theories should be developed in very close, daily contact with experiments. I do not believe that theories developed in a vacuum have any chance of success in brain science. I emphasize the same message to all the students who want to work in computational neuroscience.

Do you have a scientific hero? Yes, I have several heroes. I count two of them among my best friends: Francis Crick (who sadly died while I was preparing this Q&A) and Steve Smale. I admire both of them immensely not only because of their scientific achievements but also because of their great, refreshingly youthful attitude toward life.

What is your greatest research ambition? Of course, I would like to understand how the brain works, but right now I would be quite happy to find nontrivial consequences of the mathematical conditions of predictive learning with respect to the rules of synaptic plasticity in the brain. Alternatively, I would call it a real achievement to develop algorithms that can learn from the past and predict the future in really difficult problems, such as predicting financial markets!

Do you think theoretical approaches will play an important role in biology and especially in neuroscience? Bioinformatics has become an important set of tools within biology, and systems biology is just emerging at the forefront of the next revolution in biology. I believe that, in neuroscience, computational approaches will play a similar role, particularly with some of the harder problems of higher brain functions, such as visual recognition or motor control. Quantitative models will become powerful tools – rather like microelectrodes in neurophysiology – for summarizing and interpreting existing data and for planning and analyzing new experiments.

You have been working in computer science and neuroscience and even mathematics: how do you compare these areas of research? In the last 15 years, my coworkers and I have been working on the problem of learning. Learning is of course much more than memory. The problem of learning is the gate to understanding how to make intelligent machines and what is intelligence. It is not accidental that its complete solution will require solving it as a mathematical, engineering and natural science problem. This will certainly require the work of many different groups. It is probably a naïve renaissance dream to try to do everything at once, but right now I find it a lot of fun!

Why is the problem of learning the focus of your research? I believe that any modern definition of biological and artificial intelligence should extend the implicit definition given by Turing half a century ago, by explicitly including learning. I also believe that it is our ability to learn that preserves our individual freedom and our human dignity, and can effectively save each one of us from the imperialism of the genes. And even at the level of the evolution of our species, it is the process of learning, and now in particular of teaching and education, that allows ideas, culture and technology to spread, replicate and mutate. Ideas – Dawkins' memes – are now more important for the evolution and the survival of our species than new changes to our own genes. In short, my specific interest in the scientific aspects of learning is mirrored in my broader belief that research and education are the key engine of evolution for our culture and our society.

What do you think are the big questions to be answered next in neuroscience? It would really be good to see cortical physiology – including whole-brain imaging approaches, such as functional magnetic resonance imaging (fMRI) – go beyond the butterfly collection stage. We need to begin to address the real questions about how the brain solves the tremendously difficult problems of perception and thinking, the visual recognition of objects for example (something close to my heart). Of course, it is clearly important to know that neurons in a certain area of the brain are involved in, say, visual categorization, but physiologists should not be content with that. In molecular biology today it is not enough to know which genes are involved in a disease; you also need to know what those genes do and how they are controlled. Similarly, neurophysiologists should strive not only to measure what a neuron may compute but also to describe how it and the circuits feeding into it manage to do it.