the newsworthy milestone of cat-scale cortical simulations, roughly equivalent to 4.5% of human scale, fully utilizing the memory capacity of the system. 2 The networks demonstrated self-organization of neurons into reproducible, time-locked, though not synchronous, groups. 3 The simulations also reproduced oscillations in activity levels often seen across large areas of the mammalian cortex at alpha (8Hz–12Hz) and gamma (> 30Hz) frequencies. In a visual stimulation-like paradigm, the simulated network exhibited population-specific response latencies matching those observed in mammalian cortex. 2 A critical advantage of the simulator is that it allows us to analyze hundreds of thousands of neural groups, while animal recordings are limited to simultaneous recordings of a few tens of neural populations. Taking advantage of this capability, we were able to construct a detailed picture of the propagation of stimulus-evoked activity through the network; Figure 5 outlines this activity, traveling from the thalamus to cortical layers four and six, then to layers two, three, and five, while simultaneously traveling laterally within each layer.

The C2 simulator provides a key
integrative workbench for discover-
ing algorithms of the brain. While our
simulations thus far include many key
features of neural architecture and dy-
namics, they only scratch the surface
of available neuroscientific data; for
example, we are now incorporating the
long-distance white-matter projections
(see the first two sidebars and Figures
1 and 2), other important sub-cortical
structures (such as the basal ganglia),
and mechanisms for structural plastic-
ity. We remain open to new measure-
ments of detailed cortical circuitry of-
fered by emerging technologies.

Prospective

The quest for intelligent machines ulti-

figure 5. simulated response of thalamocortical circuitry to a triangle-shape stimulus.

Included are 1. 6 billion
neurons and 8. 61 trillion
synapses. The plot
is the propagation of
the evoked response
through the circuit,
depicting the latency of
the first spike at every
topographic location in
each cortical layer and
the thalamus following
stimulus delivery.

l2/3

l4

l5

l6

thalamus

100

50

0

first spike time (ms)

 

Included are 1. 6 billion neurons and 8. 61 trillion synapses. The plot is the propagation of the evoked response through the circuit, depicting the latency of the first spike at every topographic location in each cortical layer and the thalamus following stimulus delivery.

mately requires new breakthroughs in
philosophy, neuroanatomy, neurophys-
iology, computational neuroscience,
supercomputing, and computer archi-
tecture orchestrated in a coherent, uni-
fied assault on a challenge of unprece-
dented magnitude. The state of today’s
effort in cognitive computing was best
captured by Winston Churchill: “Now
this is not the end. It is not even the be-
ginning of the end. But it is, perhaps,
the end of the beginning.”
On the heels of the unprecedented
simulation scale and the trends in de-
velopment of supercomputer technol-
ogy, the good news is that human-scale
cortical simulations are not only with-
in reach but appear inevitable within a
decade.

The bad news is that the power and space requirements of such simulations may be many orders of magnitude greater than those of the biological brain. This disparity owes its genesis to the salient differences between the von Neumann architecture and the brain itself. 39 Modern computing posits a stored program model, traditionally implemented in digital, synchronous, serial, centralized, fast, hardwired, general-purpose, brittle circuits, with explicit memory-addressing imposing a dichotomy between computation and data. In stark contrast, the brain uses replicated computational units, neurons and synapses, implemented in mixed-mode analog-digital, asynchronous, parallel, distributed, slow, reconfigurable, specialized, fault-tolerant biological substrates, with implicit memory addressing blurring the boundary between computation and data.

The elegance and efficiency of biology entices us to explore entirely new computing architectures, system designs, and programming paradigms. Under the umbrella of the U.S. Defense Advanced Research Projects Agency (DARPA) Systems of Neuromorphic Adaptive Plastic Scalable Electronics initiative, beginning in 2008, we have embarked on an ambitious program to engender a revolutionarily compact, low-power neuromorphic chip comprising one million neurons and 10 billion synapses per square centimeter by exploiting breakthroughs in nanotechnology and neuromorphic very large-scale integration. 26