Computational perspectives and probability models help decipher how neurons communicate with each other

The brain is a vast network of interconnected neurons. There are more neurons in every person’s skull than people in the world. Each neuron speaks to a small town’s worth of correspondents by broadcasting a message consisting of a few short discrete electrical impulses called spikes. The select group of correspondents of any given neuron—some of which are next-door neighbors while others live overseas—is partly the result of random wiring and partly the outcome of a life-long process of training.

How are messages encoded by these brief bursts of neural activity? For simple sensory events, like the turning on of a light, the brightness correlates well with the rate of spiking in early neural structures, like the retina. Should we assume that all neural representations follow this single format? A number of theoreticians in Brown’s Brain Science Program (BSP), such as Professors Elie Bienenstock, Stuart Geman, David Mumford, and Michael Black, think that this extrapolation—which says that the brain works exclusively with rate codes—is not warranted. Other neural codes can be envisioned, which, while not contradicting the existence and importance of rate codes provide the basis for explaining how the brain can cope in a highly precise and efficient way with an infinite variety of never-experienced stimuli: objects, scenes and problems. Such versatility calls for codes that have more of a hierarchical nature, reproducing the structure of symbol systems like language.

If not restricted to rate codes, what would the language of the brain consist of? An intriguing hypothesis actively investigated at Brown both theoretically and experimentally, is that the brain uses a temporal code. Messages are passed back and forth as fine and complex spatio-temporal patterns between the 10¹⁰ neurons that make up the highly interconnected neocortex. In this view, firing rates, measured on the time scale of a second, are an impoverished part of the global picture. To get a more complete view we should take into account the precise timing, on the millisecond time scale, of spikes with respect to each other.

Data for testing these theories comes from several laboratories affiliated with the BSP. For example, in the laboratory of professor David Sheinberg, neurons in the infero-temporal cortex of monkeys are recorded while the animals analyze visual scenes that contain complex objects (see "How do we see?"). In the laboratory of professor James Simmons, neurons are recorded from the echolocation system of the bat, and based on these observations, models of high-resolution biosonars are elaborated in collaboration with the theoretical group of Professor Leon Cooper. In the laboratory of Professor John Donoghue, extracellular recordings from the motor cortex of awake, behaving monkeys are used for the development of mathematical algorithms for the prediction of arm movements from brain activity. While the monkey performs a simple motor task, an experimenter records the activity of about a hundred individual neurons in the animal’s motor cortex. Bayesian statistical algorithms are then used to construct models of the firing of the cells at specific times and as a function of the kinematic parameters, e.g. the speed and direction of motion of the arm.

Our Bayesian algorithms then allow one to predict the future position of the monkey's arm from the activity of the monkey's neurons. These results open up a number of important practical applications such as controlling a prosthetic robotic arm. Further, these theoretical and empirical studies provide the foundation for understanding neural activity in general and eventually, cracking the neural code.




Posted 11/03