Development strategy and philosophy
##Research Goals
There are many tasks humans perform well that robots perform poorly. Visual recognition, bipedal locomotion, manipulation, and verbal communication are a few examples. Robots and computers are able to do these tasks in simplified and structured environments, but typically not with the noise, pace, and variability present in the natural world. The primary goal of my research is to enable machines to solve the problems of natural world interaction more robustly.
There are a lot of hard problems that need to be solved to do this, including perception, natural language communication, concept learning, procedural learning, and reasoning. In order to focus my efforts on these problems, I've chosen a specific challenge that contains all these elements: the search and retrieve task. In it, a human coach gives verbal instructions to a robot to help it find an object and put it in a bin. Neither the coach nor the robot knows anything about the environment or the target object beforehand. The only feedback available to the coach is audio from the robot, as when a parent talks a child through a task over the telephone.
Although this task doesn't encapsulate all human capabilities, the skills required to succeed on it would cover a broad swath of intelligent behavior. In its complete form, the search and retrieve task is far too difficult for current systems to tackle, so I'm beginning with a highly simplified version and gradually adding complexity as my robots get smarter.
##Biases
These are the assumptions I bring to my work.
-
Since biological organisms have solved the problems of natural-world interaction already, studying and mimicking them may provide insights into the solutions.
-
While the notions of cognitive psychology (e.g. working memory, semantic memory, attention) are useful in cataloging the results of experiments, they may not necessarily be useful in describing the neural processes taking place in the brain.
-
While neurophysiological data is being generated and aggregated at a remarkable rate, it is still far too sparse to deduce neurological function.
-
While functional neurological imaging is producing a prodigious amount of tantalizing data, the new phrenology of associating very specific functions to a highly localized regions of the brain is unsupportable.
-
Although not sufficient for deducing brain function, cognitive psychology, neuroanatomy, and brain imaging studies provide a large and rich body of empirical observations from which hints may be gleaned.
-
Phylogenetically older organisms contain nature's earliest solutions to the problem of natural-world interaction. They also contain the rudiments upon which the human nervous system was built. Studying and mimicking the simplest aspects of animal behavior will be more fruitful in the long run than directly trying to mimic the complete human intellect.
-
If I can't build a machine that solves a problem, I can't pretend that I understand how biology has solved that problem.
-
If I can build a machine that solves a problem, I still can't claim that I know how biology has solved that problem. But at least then I can plausibly propose a mechanism and use it to help guide the biological investigation.
-
Nature rewards versatility rather than optimality. Optimality is just another way to say that if conditions change even slightly, performance will degrade. Conditions in nature change constantly.
-
If I can build a machine that solves all the problems of natural-world interaction as well or better than an animal or human, I don't really care whether it does it in the same way.
##Approach
The confluence of these biases leads me to use a particular approach. It is a cycle consisting of three steps.
A. Study how biology solves a particular problem. Include the perspectives of as many researchers as possible, spanning neuroscience and psychology of children, adults, and animals.
B. Formulate a potential solution for a highly simplified version of the problem that is consistent with as many of the observations as possible. Heavily emphasize versatility of the solution over optimality.
C. Implement the solution in physical hardware and demonstrate it successfully. While the problem addressed may be simple, don't fall into the trap of adding structure (like a model of the environment) in order to boost performance. I find that this step is a strong selection mechanism. The challenges of getting an idea to work in hardware highlight deficiencies that might have remained hidden in a simulated environment.
After completing a pass through the cycle, identify a slightly more ambitious problem and repeat.
