Our Innovation:
Neurogrid
To address the computational challenge of cortex-scale models, our lab is building Neurogrid, an affordable supercomputer. In contrast with Blue Brain, which has the complementary goal of modeling interactions within a cortical column, Neurogrid is for neuroscientists interesred in interactions between cortical areas. There are over three dozen in the visual system alone. Virtually every one projects back to the areas that feed it, making these feedback projections about half the total, yet their role remains mysterious. One hypothesis is that they play a role in integrating these areas' myriad representations into a single coherent percept. Another hypothesis is that they mediate attention, zeroeing in on the most informative area and excluding the others.
Neurogrid can handle simulations large enough to include interactions between multiple cortical areas yet detailed enough to account for distinct cellular properties.
While progress has been made linking cellular-level mechanisms to network-level rhythms—which are thought to modulate cortico-cortical interactions—scaling these models up to the area- and system-levels has proved difficult. To this end, rather than Blue Brain's 10,000 thousand-compartment cells, Neurogrid has 1 million two-compartment cells—one compartment for a pyramidal cell's apical dendrites, the other for its basal dendrites. This minimal cell model proved sufficient to capture interactions among feedforward, feedback, and lateral projections, which terminate in distinct cortical layers. It also replicates firing patterns of various pyramidal cell types, simply by varying how strongly the compartments are coupled.
When it is completed in 2008, Neurogrid will emulate a million neurons in the cortex in real-time, rivaling the performance of two-hundred Blue Gene racks—at under a thousandth the cost.
Neurogrid is a multichip system that consists of a 4-by-4 array of Neurocores, each of which contains a 256-by-256 array of silicon neurons, each of which can have up to 6,000 synaptic connections. Based on its similarity to GRAPE-6, we estimate that you could have one for $60K. You would assign each Neurocore to a different cortical layer (or cell-type), tailor its silicon models to match that cell-population's ion-channel repertoire, and route its softwires to match their synaptic connectivity. All this is done with a few mouse clicks using a GUI, which also presents simulation results in real-time. You can select a single cell, an entire layer, or any level of complexity in between. While you could only dream of simulating a million cortical neurons in real-time before, you really wish you had tens of millions, don't you? We're working on that: Neurogrids with larger grids (16 by 16) and larger arrays (1,024 by 1,024) should be available in a few years.
Students
Paul Merolla and John Arthur have built a four-chip 1D grid.
Collaborators
Bill Newsome
Tirin Moore
Shaul Hestrin
Yiannis Aloimonos
