Computer Architecture Branches Out
by Camille Sindhu
The next revolution in computer architecture may not come from research in computer science, but rather from the bench of a biology laboratory. Researchers working at the intersection of neurobiology and electrical engineering are already capable of building artificial analogs of biological sensory systems in silicon, but difficulties arise in emulating the more complex neural networks of brain tissue.
Dr. Kwabena Boahen, Associate Professor in the Department of Bioengineering, believes these difficulties are rooted in the lack of extensive knowledge about brain function. Overcoming this hurdle has catalyzed a gradual shift in his research goals to design techniques that enable the study of brain function on a much deeper level than ever before.
Neural vs. Von Neumann Computation
To tease apart the intricate puzzle of brain function, experimentation is required at the molecular and biochemical levels, a nearly impossible task to carry out in vivo. Instead, experiments are conducted and analyzed using a computational model of the brain. Since all the computing done by the brain is via action potentials triggered by the flow of ions through selective channels, understanding higher brain function necessitates a thorough grasp of ion channel dynamics. An ideal experiment would therefore manipulate variables such as ion type, channel characteristics and cell type, then observe the changes in activity or capability of the neural network as a whole. Unfortunately, simulation of one second of neural activity at the level of molecular detail required by a biologist takes a modern supercomputer one hour and 20 minutes, making practical and informative simulations nearly impossible.
The startling difficulty in simulation of neural activity is due to fundamental differences between the von Neumann model of computer architecture, the design upon which most computers today are built, and the neural architecture of an animal brain. Von Neumann machines operate in a sequential manner, executing step-by-step a programmed set of instructions, in marked contrast to the brainÕs highly parallel and interconnected architecture. These differences are reconciled by the Boahen labÕs Neurogrid computer: a neurobiology lab on a chip born from combining knowledge of neuroscience, computer architecture and electrical engineering. This device provides previously unattainable insights into brain function by modeling the underlying architecture of an animal brain.
Mimicking the Brain
The fundamental component of the Neurogrid computer is not a logic gate like in most computing devices, but a Òsilicon neuronÓ -- an electrical circuit of transistors and capacitors arranged in a pattern that mimics the voltage pulse of a real neuronÕs action potential. These voltage spikes are read and processed by an external computer running a software application that assembles and analyzes the generated data. This arrangement enables an incredible versatility in experimental analysis, from selecting a single cell in the network and plotting its firing pattern to examining the activity of the whole array or any level of complexity in between. Any change in connectivity patterns or cell types can easily be programmed and the experiment can be conducted again. In essence, Neurogrid models the layered organization of the cortex, with each cortical layer corresponding to another chip. This type of parallel architecture facilitates studying large systems of neural cells while taking into account the characteristics of individual neurons, thereby elucidating the connection between molecular events at the ion channel level to macroscopic changes in brain function.
The Need for a New Computational Paradigm
While the applications of neuromorphic processor technologies like Neurogrid may appear limited to providing neurobiologists with a brain on a chip, they are actually part of a much more profound shift in computing paradigms. Today we know how to build machines that compute precisely and with blazing speed. The challenge is to do it efficiently in terms of power and space. Fortunately, biological evolution has already created a computing paradigm whose design is dictated by energy efficiency: the brain.
When comparing the power used per computation by a brain versus that of a modern computer, one finds that computers use up to a staggering billion times more energy. Boahen is convinced that the key to reducing this massive energy consumption is to approach computer design in a way that emulates the brainÕs organization. ÒThis question is really a scientific one, not just an engineering one," Boahen explains. "Say youÕre a biologist and you come and tell me youÕve figured out how the brain works Ð well IÕm going to ask you, how does it do it with just 10 watts?Ó Any system seeking to mediate human-machine, real world, real time interactions must possess the same computing power as the brain, and in a similarly compact package running on comparable amounts of energy.
NeurogridÕs Future
Boahen insists that computers like the Neurogrid wonÕt replace our current von Neumann machines, which are already very efficient for current purposes. Neuromorphic processors would be complementary, specialized modules called upon in situations where adaptability, not precision, is required. The concept of a computer that can react to its surroundings rather than execute a sequence of programmed steps, as well as last longer on a small, portable energy supply, is dictating evolution of computation. Surprising as it may be, the solution to our computational challenges today is looking more and more like what biology has already designed.