For more than a decade, a grand challenge posed to computer researchers has been to understand, and eventually replicate, the way the brain computes – “reverse engineer the brain”, so to speak. Despite its universally recognized importance, computer researchers have made little forward progress. In fact, theoretical neuroscientists have assumed leadership in architecting plausible computing models and consequently have taken significant first steps toward solving the problem.
The challenge needn’t be cast in strictly neuroscience terms. It can also be addressed in computer architecture terms. The research reported here supports the computer architecture perspective by proposing a computation model that includes an important class of neuroscientific models as a subset and which possesses at least some of the look-and-feel of conventional computer design methods. Rather than being based on logical principles, however, it is a radically different model based on temporal principles.
This talk first describes biologically-based neuron models commonly used by the neuroscience community. Biological neurons communicate and compute using information encoded as voltage pulses, or spikes. The focus here is on an important class of spiking neuron models in which information is conveyed and processed via precise spike timing relationships measured across multiple communication paths.
Then, a “space-time” algebra is proposed as a way of capturing the essential features of the spiking neural networks that we are targeting. The algebra models the passage of time among inter-operating spatial computing elements (e.g., neurons ). Spiking neuron models, as envisioned by neuroscience researchers, can be implemented using the primitives of the proposed space-time algebra. This construction and modeling approach is aligned with conventional computer design methods and is very different from the current neuroscience approach of discretizing real-valued biologically-based models.
Finally, potential applications of space-time algebra may be much broader than spiking neural networks. It is shown that space-time algebra also supports a generalization of “race logic”. A key feature of race logic is that it can be directly implemented with off-the-shelf CMOS digital circuits with times of signal transitions (edges) representing temporally coded values. An important implication is that we may be able to design cognitive systems in the spiking neural network domain, and then, by representing temporal events as voltage edges rather than spikes, implement them directly using off-the-shelf CMOS.