Over the Christmas break, I had the opportunity to sit back, relax and unwind. Among other things, I found myself flipping through some Calvin and Hobbes cartoons and one of the comic strips (in which Calvin explains how much it takes for his bones and muscles to coordinate and form his childish drawings) got me thinking about a question I find myself repeatedly coming back to – how and where do we store (in our “wet” brains) our knowledge of complex dynamical behaviours?
We acquire a variety of dynamically dexterous skills during a lifetime, ranging from repeated activities such as folding laundry to rarer events such as driving to a hole-in-one (or some other spectacular feat in your favorite sport). Even lesser animals (such as some four-legged grazing animals in the Sahara) perform some incredible feats such as walking – reliably enough to not be eaten by hyenas – within minutes of being born, and subsequently – rather quickly – acquiring the ability to outrun and outmaneuver an impressive predator like the cheetah. Clearly, some of this is the result of adapted morphology – but much of what I am talking about here is “strategy”. Where is this strategy stored and how is it improved over time?
I am yet to hear a sufficiently good biological answer to this question. I am aware of the standard developmental argument (i.e., evolutionary adaptation) that shapes the morphology and improves it over time. But, genes don’t store the sort of strategies I am talking about. At best, they can specify a learning architecture that is tuned to some representation that is highly efficient for such dynamical behaviors. If so, I am very curious to know the details. Here again, I am aware of the neuroscience work on motor control but, on that path, we seem to be surprisingly far from understanding sophisticated and dynamically dexterous behaviors. And, of course, machine learning provides a few answers – but none that can do so much with such little and noisy data. Moreover, machine learning at this level of dynamic complexity seems elusive without “human intuition” – which needs to be seeded somehow. So, I want to know how nature does this?
Many people in the developmental robotics community are interested in similar questions but I rarely see anyone take on a task with nontrivial dynamics, of the sort that I mention above. And the representations that are used in the simpler scenarios seem unlikely to scale easily.
From my own work, I have at least one hypothesis in response to this question, in the context of robots. Meanwhile, I would love to learn more about how this is achieved in natural animals!