Computers are great at performing functions in the same way over and over again. They're not so great at adapting to situations they haven’t seen before. Now a group of researchers from MIT is using the human brain as a model for the circuits in a machine -– and opening the way to a computer that can really learn.
The key is plasticity. Plasticity is the quality of human brains that makes them able to change in response to stimuli. It’s one reason many neuroscientists think we remember and learn. It also underlies the ability of the brain to recover from injuries.
The MIT researchers designed a computer chip that can simulate the activity of a single brain synapse. The synapse is the connection between two neurons, through which information flows. But it isn’t like a wire. Synapses are gaps and chemical signals, in the form of neurotransmitters, jump across them and bind to receptors. Those receptors activate ion channels. When an ion channel is opened and closed it changes the cell’s electrical potential and if the change is large enough, an electrical impulse is fired — called an action potential.
The ion channels control the flow of charged atoms of elements such as sodium, calcium and potassium, hence the name. When those charged atoms flow down the channels, they strengthen or weaken synapses.
The MIT team designed a computer chip whose transistors behave like ion channels. Instead of the digital on or off mode that is basic to any computer, these transistors have different levels of current, just as in an analog circuit. A gradient of electrical potential drives current to flow through the transistors just as ions flow through ion channels in a cell. This is similar to having a circuit with a variable resistor on it; analog volume dials function in just this way.
The chip has about 400 of these kinds of transistors, quite small but useful for modeling certain kinds of brain function. And that’s exactly what they plan to do. One could model the visual processing system much more realistically, and do so faster than with conventional computers. A high-speed computer might take days to model something like that, whereas this “analog chip” system can do it in seconds.
Chi-Sang Poon, a principal research scientist in the Harvard-MIT Division of Health Sciences and Technology, is the senior author of the paper describing the chip in the Proceedings of the National Academy of Sciences. He said in a press release that the technology will offer insights into how the brain works and could be used in prosthetics, a field that is rapidly expanding. It also sheds light on a longstanding debate about long term depression (LTD) and potentiation (LTP) –- weakening and strengthening of neural connections.
One theory holds that LTD and LTP depend on how often a neuron fires. More recently theorists have suggested they depend on the timing of the action potentials’ arrival at the synapse.
Both theories depend on the involvement NMDA receptors, which detect the activation of a neuron after the chemical signal arrives. One candidate for that second receptor is an endo-cannabinoid. Endo-cannabinoids are similar in structure to marijuana and are produced in the brain. They are involved in many functions, including appetite and memory. Some neuroscientists had theorized that endo-cannabinoids produced in the neuron are released into the synapse, where they activate endo-cannabinoid receptors in the neuron that sent the signal. If NMDA receptors are active at the same time, LTD occurs
When the researchers included on their chip transistors that model endo-cannabinoid receptors, they were able to accurately simulate both LTD and LTP. It was the first time that anyone had modeled them both and demonstrated that it could work.