On This Page

Advanced Techniques
Three Approaches
Supercomputers

The Challenge:
Simulating millions of neurons

Advances in neural recording and imaging techniques have allowed researchers to acquire vast amounts of data about brain structure and function. Microelectrode penetrations measure voltage differences along fine neural processes, elucidating mechanisms underlying reception and transmission of electrical signals between neurons. Serial-scanning electron microscopes image fixed brain-tissue slice by slice, enabling the neurons' synaptic connectivity to be reconstructed. While two-photon microscopes image activity in the living brain layer by layer, mapping the behavior of thousands of neurons.

Cellular-Level Mapping of Cortical Function Physiologists mapped orientation preferences of almost every cell in a 0.3mm-cube of rat visual cortex by presenting grating stimuli. This was done by injecting a dye into the living brain that makes active cells glow—but only where illumination (red pencil) is intense enough to evoke (two-photon) fluorescence. This happens only at the microscope's focal plane (colors)—because the light comes through its optics—making it possible to resolve neural activity at different depths. [Reid et al. 2005]

The challenge is to relate observed cortical structure and function: the biological processes taking place at the molecular-level and the system-level functions they enable.

While dissecting the brain experimentally uncovers its hierarchical organization (molecules, synapses, neurons, columns, areas, systems, and organism), it fails to provide insight into how these organizational levels are integrated. Analyzing the brain mathematically explains behavior at one level using abstractions from the level beneath, but brushes aside details to simplify the brain's complexity (e.g., heterogeneity, nonlinearity, and stochasticity). Simulating the brain computationally permits inclusion of molecular-level details while replicating system-level functions, but requires enormous amounts of data to constrain parameters. Thus, these three approaches truly complement each other.

With the recent advances in neural recording and imaging techniques however, our ability to characterize the cortex's structure and function trumps our ability to simulate its behavior.

Imagine simulating the visual cortex’s neural activity during a visual task. You would need a supercomputer to evaluate the set of nonlinear differential equations that model electrical current flowing through each cell's repertoire of ion-channels. When researchers lucky enough to have access to a two-million-dollar Blue Gene rack (2048 processors) attempted this, it took one hour and twenty minutes to simulate one second of neural activity in a piece of cortex with 8 million neurons connected by 4 billion synapses. Why did this tiny piece—4% of primary visual cortex (macaque)—bring Blue Gene to its knees? It needed to evaluate those equations 40 trillion times!