SCIENCE & ENGINEERING NEWS
San Diego, CALIF. — Bioengineers at the Johns Hopkins Medical Institutions, taking Plato’s concept of reality and illusion into the world of robots, have uncovered some of the algorithms of learning, the “primitives” the brain uses to comprehend the world. In particular, they have described the mathematical shapes used to control movements of the arms.
The primitives demonstrate why certain tasks are hard for us to learn, and that there may be fundamental limitations to what is learnable by the human brain.
In an article in the October 12 issue of Nature, Kurt Thoroughman, a graduate student in biomedical engineering, and Reza Shadmehr, Ph.D., assistant professor of biomedical engineering and neuroscience, report mathematically deconstructing the learning process.
Subjects in the experiment learn to control the motion of a robotic arm. Shadmehr and Thoroughman have mathematically described how they do it by constructing an internal model of reality in the brain. The report finds the specific shape and size of the primitives with which the brain builds internal models of physical dynamics.
“By programming the robot, we can produce the dynamics of things that either exist naturally or don’t make sense. We simulate the dynamics of systems and we look at how people learn to control those systems,” he says.
Behind the experiments is a concept called “primitives,” the abstract elements by which the brain deals with the physics of objects and how the arm interacts with them. One way to think of primitives is to think of a Lego set, which contains a finite number of blocks, in particular shapes and sizes. To perceive reality, the brain arranges the blocks in order and eventually builds an internal model of the physical dynamics being produced by the robot, just like one builds a castle with Lagos. The smaller the size of the Lagos, and the more movements the Lagos make, the more intricate the internal model the brain can build and the more precisely the brain can control movements. As reality changes, the brain must rearrange the blocks to reflect the new reality, which is the way we learn.
The experiment involves a subject guiding a spot across a transparent screen by moving a robot arm. The goal is to move from one place on the screen to another. In response to the subject’s movement, the computer produces a patter of forces that has to be learned. In about an hour’s time, subjects get much better at controlling the motion.
During the learning process, Shadmehr programs in occasional radical changes in the forces, totally upsetting reality.
The brain is forced to alter the primitives, readjusting to the reality rather than the perception of reality – enlightenment right out of Plato’s parable of the cave, where reality was depicted as shadows on the wall. By mathematically analyzing how these sudden changes affect the subsequent movements, the report infers the shape and size of the primitives used to construct the model of reality in the brain.
“We think the shapes are related to the neural firing behavior of cells in the brain, the cerebellum,” Shadmehr reports in the Nature article. It is possible that it is in this region of the brain that the internal model is constructed.
Then, Shadmehr programs in a radical change, totally upsetting reality.
For an un-Platonic analogy, think of a milk bottle, solid white and apparently full of milk, standing on a shelf. In order to pick it up, the brain sends a signal to the arm and hand to lift the bottle, making presumptions about how much the bottle weighs so sufficient force can be utilized. If the presumptions are correct, the procedure works flawlessly. But what if the bottle is only painted white, is actually empty? The hand, expecting it to weigh more, uses too much force and the bottle goes flying.
This shows that humans use internal models of reality to guide movements. The change in the program is the equivalent of secretly replacing the full milk bottle with an empty, lighter one. In this case, the robot arm scrambles where and what forces it applies so the previous learned path between spots on the screen no longer works.
The subject working the arm decides how to modify the much-altered model of reality. The paper describes mathematically how that happens.
First comes the process of unlearning the old paradigm before learning the complex new one. The subject has to think: “If I make my next movement, how will that affect my other movements? You changed the world for me; I see errors. If the errors are big enough, I have to reconsider the entire model.” The next move the subject makes likely is worse than the one before until the subject starts realigning the primitives correctly.
Because the changes in the reality were done deliberately, not randomly, Shadmehr has been able to predict with considerable accuracy what the subsequent moves will be, how the primitives are changed.
Experiments show that the theoretical limits are closely observed in people’s learning abilities. In other words, humans have a limited number of primitives to deploy when learning skills.