The Challenge:
Simulating a million neurons
Brain simulations are difficult because computers operate sequentially, with one or a few cores executing a preprogrammed set of instructions step-by-step, while the brain operates in parallel, with a multitude of neurons processing information distributed throughout a highly interconnected network. The computer can compensate for its lack of parallelism by executing instructions blazingly fast, but it pays a steep cost in energy and time to shuttle far-flung data through its central processing unit—putting cortex-scale simulations out of reach. To address this shortfall, our lab is building an affordable supercomputer: Neurogrid.
Spike activity of one Neurocore's 65,536 silicon neurons, simulating cortical neurons that inhibit each other. Nearby neurons spike in synchrony (top, color-coded) at about 14 Hz (bottom, sharp peaks). These groups drift in and out of synchrony at about 2 Hz (bottom, broad peaks). With each of its sixteen Neurocores programmed to model a distinct cell type, Neurogrid can simulate up to sixteen cortical cell layers at once [John Arthur 2009].
Neurogrid's speciality is simulations large enough to include interactions between cortical areas yet detailed enough to account for distinct cellular properties.
Neurogrid is part of a profound shift in computing, away from the sequential, step-by-step Von Neumann architecture towards a parallel, interconnected architecture more like the brain. Its fundamental component is not a logic gate, like in a digital computer, but a silicon neuron, whose behavior and connectivity are programmable. This neuromorphic approach, developed over the past two decades, yields hitherto unimagined levels of efficiency that make Blue-Gene performance affordable on a Dell-cluster budget.
A promising alternative to GPUs and FPGAs, Neurogrid will make the computational power required to explore various hypotheses about how the cortex works affordable. Neurogrid's speciality is modeling interactions between cortical areas, of which there are over three dozen in the visual system alone, connected by lateral, feedforward, and feedback projections. Feedback projections constitute about half the total—virtually every area projects back to the areas that feed it—yet their role remains mysterious. One hypothesis is that they integrate these areas' myriad representations into a coherent percept. Another hypothesis is that they mediate attention, zooming in on the most informative area and excluding the others.In contrast, Blue Brain has the complementary goal of modeling interactions within a cortical column at subcellular resolution.
