Reverse Engineering the Mind with Brain Corp. CEO Eugene Izhikevich
The brain initiative that President Obama unfurled last week calls for spending over $100 million a year on neuroscience research over the next decade, including the development of innovative neurotechnologies to gain new insights into the way the brain works.
As the Salk Institute neuroscientist Terrence Sejnowski put it, “This is the start of the million neuron march”—an ambitious quest to accomplish for the central nervous system what the $3 billion Human Genome Project has done for our understanding of genetics.
Part of what’s known as the Brain Research Through Advancing Innovative Neurotechnologies (BRAIN) Initiative already is underway in San Diego, in what might be described as the ultimate hack: Reverse engineering the human brain to create a duplicate system that can be implemented using semiconductors and software.
Eugene Izhikevich has been at the forefront of this emerging field for more than a decade. A computational neuroscientist, Izhikevich moved to San Diego in 2000 when he began a fellowship in theoretical neurobiology at The Neurosciences Institute in La Jolla. In 2009, he founded Brain Corp., a startup backed by Qualcomm Ventures. Izhikevich serves as the CEO and chairman of Brain Corp., which has been incubating at the San Diego headquarters of Quaclomm (NASDAQ: QCOM), the world’s largest wireless chipmaker.
The first research breakthrough came in 2003, when Izhikevich published an algorithm to describe the pulsed signal of “spiking” neurons in the brain. With this, he created a computer model in 2005 that simulates the signaling activity of 100 billion neurons and a quadrillion synapses, which he says is roughly equivalent to the size of the entire human brain.
Izhikevich says the Obama administration’s BRAIN initiative has three thrusts: 1) recording the brain’s activity to a level of detail that was not previously possible; 2) stimulating the brain in controllable ways to identify the structure and function of different regions; and 3) developing computer systems that model the biological processes of the brain.
In an e-mail, he writes, “During my years at the Neurosciences Institute, I was mostly doing 1, but a bit of 1 and 2. I did 3 on the largest possible scale.”
In laying out the BRAIN initiative last week, the White House said it plans to provide $110 million in federal funding for the initiative in fiscal year 2014, which begins in October. According to a White House fact sheet, $40 million would come from National Institutes of Health, $20 million from the National Science Foundation, and $50 million from the Defense Advanced Research Projects Agency (DARPA), which provided some funding to Brain Corp. in 2010.
(Private research institutes also have promised to provide funding for the BRAIN Initiative, including $60 million from the Allen Institute for Brain Science, $30 million from the Howard Hughes Medical Institute, $28 million from the Salk Institute, and $4 million from the Kavli Foundation. The total from both government and private sources is $232 million.)
Izhikevich agreed to answer a few questions about his work by e-mail. He would not discuss aspects of the technology under development at Brain Corp., or how it would be applied commercially.
Xconomy: How is the technology you’re developing conceptually different from advanced artificial intelligence and neural network systems in use today?
Eugene Izhikevich: We base our models on neuroscience, rather than on computer science. We try to mimic the neuro-computational processes that take place in the brain, as opposed to capturing logical or statistical computations as AI researchers do.
What we do could be classified as a “biological neural network” system. However, our models are much closer to biology (in dynamics and scale) compared with traditional neural nets. In particular, we use spiking neurons and spike-timing-dependent synaptic plasticity [a biological process that adjusts the strength of connections between neurons in the brain].
X: What is a spiking neuron? I’m visualizing an electric signal that pulses up and back, like the strongman hitting the bell in a carnival high striker game. Or is it more accurate to say that a neuron is either “on” or “off?”
EI: The strongman analogy is good. Neurons are typically quiet for long periods of time, then fire a brief pulse (called an action potential or a spike) or a burst of spikes, in response to signals arriving from other neurons. The exact details depend on where the neuron is.
Imagine 100 such spiking neurons, each firing just one spike per second. If we consider the timing of these neuron spikes, such a system has more than 10^160 (ten with 160 zeros) of possible combinations, each representing a pattern of spikes.
Not only is this number greater than the number of particles in the known universe (which is 10^80), it also is greater than the number of pair-wise combinations of all the particles. It is hard to comprehend how large this number is. It is infinite from any practical point of view, yet we can achieve such a combinatorially large capacity in a network of just 100 neurons—as long as we capture the timing of spikes. Now, imagine not 100, but 100 billion neurons! Any researcher or a company that figures out how to use this will unlock the key to the neural computations in the brain, and enable a trillion-dollar technology. Even a partial success would enable smart consumer devices that behave less like robots and more like animals.
X: Is it accurate to say your first breakthrough was developing an algorithm to describe the biological process of spiking neurons? How, or why, was it important?
EI: Correct. Thousands of researchers use my model (and refer to it by my name) as a computationally efficient way to simulate spiking and bursting activity in neurons. The model, published in 2003, paved the way to simulate millions or billions of neurons with firing patterns similar to those observed in the brain. The model captures the essence of neural computation taking place inside each neurons to the degree that if I stimulate a real neuron and the model neuron with the same stimulus, show the results to an expert neurobiologist, the expert would not be able to tell the difference.
X: Did this algorithm make it possible to develop the large-scale computer model of a normal human brain?
EI: Yes, all the way to the 100 billion neurons.
X: Was this the key innovation that made you think this was technology that could be commercialized? If not, what was that key innovation?
EI: The key technology breakthroughs are (a) the development of the efficient model of spiking neurons, (b) the development of various forms of spike-timing-dependent synaptic plasticity resulting in emergence of neuronal computations, and (c) the availability of high-performance processors and a path to develop a new generation of specialized processors that take our simulations to the next level.