— Asks Neuroscientist and Postdoctoral Fellow
Only male songbirds sing. They sing to impress both females and other males — to attract the former, and to scare off the latter. But male birds are not born knowing how to sing. To learn, the baby has to grow up hearing an adult male sing. That tutor is necessary but not sufficient for birdsong learning.
How the young bird learns song appears to bear a striking resemblance not only to how a human baby learns speech, but also to how general goal-directed behaviors involving fine muscle control are learned. Or at least that is the working hypothesis for much of Ila Prasad Fiete’s research.
Fiete, a three-year postdoctoral fellow at the Kavli Institute for Theoretical Physics, is a theoretical neuroscientist.
KITP post-doc and theoretical neuroscientist, Ila Prasad Fiete. Photo by Peter Allen.
During her third year of graduate school in physics at Harvard, she attended a biophysics course given by Sebastian Seung. A theoretical neuroscientist at the Massachusetts Institute of Technology, Seung had been a student of Harvard condensed matter physicist David Nelson.
“I learned about protein kinetics in cells,” said Fiete, “and I learned about how neurons fire and communicate with each other.” She also learned that she, like Seung before her, wanted to switch fields from hard condensed matter to neuroscience. Having become Seung’s student, Fiete still collaborates with him, and with two other physicists-turned-neuroscientists, Michael Fee and Richard Hahnloser, on songbird learning.
Song by trial and error
If the baby bird grows up in isolation, Fiete explains, “he does not, in fact, learn how to sing anything recognizable as a successful song for mating.” On the other hand, if the juvenile bird is exposed to the tutor's song even briefly, and the tutor is removed before the baby bird has begun to practice its song, he will still eventually learn to produce an excellent match to the song of the adult (even if the two are unrelated). So, Fiete points out, the baby must rapidly acquire a template of the tutor song in his head.
Learning to generate an accurate vocal copy of this template is slow. At first, the baby produces sounds that are the bird equivalent of babbling. At this stage, auditory feedback is critical. If the bird is deafened after acquiring the template but before it has mastered song production, it will never be able to reproduce the tutor song. A normal bird, after much trial and error, eventually learns how to reproduce the song.
“So,” said Fiete, “even though the bird has the template in his mind, he can’t just play off the template in his vocal cords and produce the right song.” The bird has to be able to hear himself, and in effect compare the sounds he makes with the tutor template, and then gradually alter the vocal apparatus in order more closely to approximate the tutor song.
“He has to practice to get it right,” said Fiete. “It’s what we all do!”
“Let’s say I want to learn to serve a tennis ball,” she said. “My goal is to get the ball to the other side of the court. So I know roughly what I want done, but I don’t know how to do it. Specifically, my neurons have no idea how to do it. Even if I hire a tennis coach who tells me how to swing the racket, she can guide my hand and swing my racket for me, but she can’t tell me which neurons I should fire with greater or lesser intensity to make the muscles contract to make the ball do what I want it to do. In such an instance of goal-directed motor control, there are tens to hundreds of thousands of neurons involved in controlling the trajectory of the arm."
“My own research focuses on how individually ‘dumb’ neurons might each locally figure out how to change in order to learn to perform collectively a goal-directed task. Birdsong is just one example; the tennis serve is another. We know on a behavioral level we do it by trial and error, but don’t know what drives that behavior on the level of single neurons.”
Monte Carlo
“In my work with Sebastian, we hypothesize (like a few others before us) that to perform goal-directed learning, the brain itself performs Monte Carlo-like simulations on itself to decide how to change to improve performance! Our contribution is to provide a specific, biologically plausible algorithm through which realistic networks of neurons may do so.”
Monte Carlo simulation is a catchall name for algorithms that use random exploration to estimate answers to problems that cannot be solved analytically.
Said Fiete, “We hypothesize that a dedicated group of neurons in the brain acts as a source of randomness to drive variations in the activities of the motor control neurons, thereby producing behavioral variations that are necessary for trial-and-error learning.”
If, at a certain time during a motor action, a neuron receives a noise input that makes it momentarily more active than usual at that time, and if that event is followed by reinforcement, the neuron should strengthen all its incoming synapses (connections between neurons) that were active at the time. Conversely, if the noise temporarily suppresses activity, and this is followed by reinforcement, then the active synapse should be weakened.
“We can mathematically prove,” said Fiete, “that this learning rule will optimize performance on the task, and what’s more, will do so in very general models of neural networks. Neurons in the brain certainly have enough machinery to implement this rule, and we await experiments that can verify our strong predictions on how activity and reinforcement in the brain may drive synaptic change.”
Reductionist mind set
“One of the qualities that a physicist brings to the study of biology, for better or worse, is a reductionist mind set,” said Fiete. “You asked me what exactly do I use in neuroscience that I learned in physics. The answer is the valuable mathematical tools of our training and a physics outlook. I look at the world and think, 'Wow, is it complex!’ But, as a physicist, I believe that there are some essential features underlying the complexity, and that I can make ‘toy’ models that capture a few of the very important ingredients and that reproduce a lot of complexity."
“Biologists by training are very descriptive. They tend to emphasize the richness and complexity of the phenomena they study and don’t always believe in reductionist views of the underlying dynamics,” said Fiete.
“So, in some biologist circles, 'reductionism’ used to be a bad word. But these attitudes have been rapidly changing due to the sudden influx of huge amounts of genomic and neural data that call out for quantitative mining, and due to increasing numbers of fruitful collaborations between physicists and biologists."
“In fact, however,” she added, “a suspicion of reductionism may be partly justified, because biology could be a new frontier where the complexity is such that we cannot go to the same level of reductionism as we can in physics."
“But even then, if you don’t believe in reductionism at some level, it’s hard to understand anything,” said Fiete. “If you replace a complex phenomenon by an equally complex model, you haven’t gotten very far.”
On being a postdoc
Fiete has just completed her first year as a postdoctoral fellow. “There are a lot of advantages to being a postdoc at the KITP,” she said — “a lot of free time, and a lot of independence to develop an individualized research program in a way I probably would not be able to do in most places as a postdoc. Most postdocs elsewhere are tied to one particular faculty member, or a faculty member’s grant to do a specific project. Here, postdocs are just hired for their interests and their potential, broadly defined."
“These three years allow me to explore what I want to do,” said Fiete. “It’s the greatest opportunity because I have complete freedom to work on the problems I find exciting.”