simulations at the scale of small mammalian brains. Though we have only humble achievements to report, our aspirations are lofty. We seek nothing less than to discover, demonstrate, and deliver the core algorithms of the brain and gain a deep scientific understanding of how the mind perceives, thinks, and acts. This will lead to novel cognitive systems, computing architectures, programming paradigms, practical applications, and intelligent business machines.

Rationale

Our rationale was aptly and eloquently
captured by Churchland and Sejnows-
ki, writing, “It would be convenient
if we could understand the nature of
cognition without understanding the
nature of the brain itself. Unfortu-
nately, it is difficult if not impossible
to theorize effectively on these mat-
ters in the absence of neurobiologi-
cal constraints. The primary reason
is that computational space is con-
summately vast, and there are many
conceivable solutions to the problem
of how a cognitive operation could be
accomplished. Neurobiological data
provide essential constraints on com-
putational theories, and they conse-
quently provide an efficient means for
narrowing the search space. Equally
important, the data are richly sugges-
tive in hints concerning what might
really be going on and what computa-
tional strategies evolution might have
chanced upon.”
Neuroscience today is rich in de-
tailed biological observations, as
reflected in the sheer size— 1,414
pages—of Principles of Neural Science,
a modern introductory textbook by
Kandel et al. 18 As neuroscientists, we
view these observations as a web of
clues to the biological mechanisms of
cognition. As engineers, we view them
as something else entirely. The brain
is an example solution to the problem
of cognitive computing, and the ob-
servations of neuroscience are a par-
tial set of constraints on the form of
that solution. The trick to leveraging
neuroscience in the name of cognitive
computing is to separate the wheat
from the chaff.

Here, we explore the fundamental
neuroscientific constraints on building
a functional simulation of the brain,
first describing structural constraints
learned from the wiring diagram of
the brain. The central message is the
brain’s neuronal network is a sparse,
directed graph organized at multiple
scales. In particular, local, short-range
connections can be described through
statistical variations on a repeating
canonical subcircuit, whereas global,
long-range connections can be de-
scribed through a specific, low-com-
plexity blueprint. We highlight what
neurophysiology has taught us about
the dynamics of computation and
communication within this network.
Our thesis is that the computational
building blocks of the brain (neurons
and synapses) can be described by rela-
tively compact, functional, phenom-
enological mathematical models, and
that their communication can be sum-
marized in binary, asynchronous mes-
sages (spikes).

neuroanatomy

A central tenet of neuroscience, sometimes called the “neuron doctrine,” posits that specialized cells in the brain, the neurons, are the biological substrate of brain computation. The function of individual neurons is covered later in the section on neurophysiology, but for now, neuronal function can be abstracted to receiving, integrating, and sending binary messages. These messages are communicated at points of contact, dubbed synapses by Sir Charles Sherrington in England in 1897. Through messaging, neurons collaborate to form networks that engender powerful capabilities, vastly more sophisticated than the processing capacity of individual neurons. To understand brain function, it is crucial to understand the organization of neural circuitry.

Connectivity in the brain is sparse. Adult humans have about 100 trillion synapses, six orders of magnitude less than would be required to completely and directly connect the tens of billions of neurons that make up the brain. Moreover, there is strong evidence that biology has a relatively compact algo-