1. Why were subjects with these particular ages chosen for this experiment?

The question of whether it is possible to perform cross-modal object identification, or to develop this ability over time, needs to be addressed for all subject ages. In principle, it is conceivable that young subjects may perform better than older subjects for the obvious reasons, such as shorter periods of sight deprivation and greater levels of plasticity in the young. Thus, we tried to maximize the age range of our subjects. Young children are often rather uncooperative or unable to understand our instructions, but we were lucky enough to work with an intelligent and well-behaved 8 year old as our youngest subject. Finding subjects that fit the very stringent requirements of our test was very difficult, so we did not have the opportunity to test subjects older than 17, but the results do not indicate that we would get substantially different results in an older age group.

It is possible, of course, that a much younger subject may exhibit "cross-modal transfer" (that is, be able to perform the object identification task across modalities) immediately after sight onset, indicating innateness. However, it would be surprising if an ability which is so quickly learned in the adult without resorting to innate abilities is also available through innate channels to the infant. Moreover, an adaptable and learned mapping between touch and vision is even more critical to the infant's development than to the adult, since an infant's limbs and digits grow and its muscles gain strength quite rapidly in ways that the brain would not be able to anticipate. Thus, any mapping between the sense of touch and vision would be rapidly decalibrated during normal growth. Upon reflection, it seems necessary to have a cross-modal mapping system that can be trained and retrained relatively quickly, and our study suggests that indeed this is the case.

2. Can the Molyneux Problem be resolved with brain imaging technologies, rather than through behavioral experiments?

With most experiments utilizing brain imaging, many subjects are required in order to find an "average" result which is somehow different for the various experimental conditions. It is very rare to find such a result in just a few individuals. In the case of the Molyneux Question, one would have to know how to interpret the "brain signal", which is quite difficult in normal individuals. In our subjects, the periods of long-term visual deprivation would alter their brain responses even more from the normal patterns, so it would be doubly difficult to interpret them. Thus, even if we found differences in brain activation patterns between our subjects and normal subjects, or between our subjects immediately after sight recovery and our subjects after some time, we would not know the significance or reasons for this difference. There is no "cross-modal area" or "object identification area" that we could monitor for changes, as far as we know.

At this point in Neuroscience's understanding of brain imaging techniques, they are not sufficient to answer the Molyneux Question.

3. What problem in Neuroscience did this experiment resolve, and what new questions have arisen from it?

This experiment speaks largely to the notion of "representation", that is, in what form does the brain "store" an object? Given that the subjects were able to perform well in the touch-to-touch and the vision-to-vision tasks, we can surmise that they did have visual representations and tactual representations available to them. However, were these representations the same? The answer seems to be "no", because if they were, the touch-to-vision task would have been trivial. Furthermore, if these representations are not the same, is there an innate connection between the two types of representations? The answer again seems to be "no", because such a connection (or "mapping") would have allowed the subjects to perform the touch-to-vision task. The most likely explanation is that the representations formed in the two different modalities (touch and vision) are connected to each other via learning gained through experience.

Following the result in this paper, the next big question is: HOW does this mapping occur? That is, how is it that in just a week or so of undirected real-world experience, a cross-modal map is formed (or, at least, enough of a mapping is formed to complete the task we tested)? The particular mechanisms of this learning are still to be explored, though we have our theories ...

4. What impact do these findings have on the field of Artificial Intelligence?

The implications to AI (artificial intelligence) are quite interesting. The crux of the Molyneux Question is whether cross-modal mappings are innate or learned. If innate, then an algorithm for creating cross-modal mappings would have been derived through hundres of millions of years of evolution, involving countless mini-experiments in the form of exponential breeding and genetic mutations. Such a result would have suggested that researchers working in the artificial domain would need to develop an equally sophisticated algorithm for this specific problem. The other result--namely, that cross-modal mappings are learned--suggests that artificial algorithms NEED be good at learning, but NEED NOT self-contain all the relevant information. The rapidity of learning that we observed also suggests that the system does not necessarily need a very large amount of input to accomplish this learning.

If these inferences are correct, then it seems quite likely that it is possible to build, say, a robot that can learn properties of the physical world that engineers do not explicitly anticipate, and the robots may learn and adapt robustly even as its physical components may change due to gravity (in the case of space travel), humidity, or physical damage. How this learning actually occurs, however, is still a very open question.

5. Tell us a bit about your academic career ... How did you end up at MIT? If you had studied in the Russian Federation, how would that have impacted your academic progression? What is the current state of Russian Neuroscience?

I grew up in the U.S., attending public schools in my small town. I studied Computer Science at Harvard, where I gained an appreciation both for computer algorithms and for the surprises that the brain yields, having also studied linguistics and vision science during my years as an undergraduate. Upon completing my undergraduate studies, I briefly worked in IT (Information Technology) until I decided to go get a PhD in order to apply my computer science knowledge to the study of the brain and vice versa. MIT seemed like a very good place to do that, and in fact it was the only university to which I applied. I do feel that a computational mindset while studying the brain's function gives a certain type of insight that is sometimes missing from research projects in neuroscience.

I don't know very much about the state of neuroscience in Russia, but I have noticed that there is a curious lack of participation currently among Russian scientists in my own subfield of visual neuroscience, as evidenced by the papers I read and the presenters I see at conferences. I would be curious to know what is the reason for that.