Computing With Spikes

The brain is an excellent example of a hybrid of an analog computer and a digital computer. Neurons use "spikes" or pulses of voltage to communicate with each other, and these spikes have an inherently hybrid nature: The presence of the spike is a digital event, but the time between spikes is a continuous analog variable.

Hybrid computers like the brain can combine the analog advantages of efficient operation with the digital advantages of the ability to divide complex computational problems up into a sequence of simpler tasks, scalability, programmability, and robustness to noise. To execute a program stored in memory, digital computers operate as finite state machines. In each state, they accomplish a task, and then transition between states according to the value of an input or according to a prescribed sequence in the program. For example, your PC running a program that plays music might transition to a "loading state" and load a saved file from memory if you select "Load File" from a pull-down input menu. After loading, it may automatically transition to a "playing state" where it plays the file automatically from beginning to end. It may abort playing the music if you select "Stop Play" and transition to a "holding state" where it waits for further input.

When dealing with real-world signals like sound and images, it is traditional today to perform all complex computations by digitizing the signal with an analog-to-digital converter and then processing it with a finite state machine in the form of a digital microprocessor running a sequence of instructions. In a hybrid state machine (HSM), one that combines both analog and digital computation, we can effectively combine both these operations into one unit and do away with the analog-to-digital converter altogether.

The hybrid state machines [see illustration] we've built have both a reconfigurable analog computing unit and a digital finite state machine. The digital portion reconfigures the analog unit to accomplish different tasks in each of its digital states. The analog unit outputs voltage spikes when its task has been completed allowing the digital portion to proceed to the next state (and task) once the spike has been reported. The spikes output by the analog computing unit may in addition signal 'yes' or 'no' answers to questions posed in each digital state. For example, the HSM can be configured such that a spike arriving after the leading edge of a clock signal means 'yes' and one arriving after the edge means 'no'.

A sequence of spikes that yield 'yes' or 'no' bits convey significantly more meaningful information than bits that reflect just the raw numbers in an image or a sound signal. For example, by constantly reconfiguring analog circuits in each of its states, a hybrid state machine can ask a sequence of meaningful questions on an analog input image such as "Does it have eyes?", "Is it a face?", "Are the features of the face close to a stored set of faces?", "Are they close enough for me to recognize the face?" etc. The raw numbers that convey visual intensities of millions of pixels in the image are never digitized at all, rather only the meaningful information is digitized. Neuroscientists have studied how neurons in the brain's thalamus and cortex are connected, and it appears that a feedback architecture that span those two structures may work similarly: In each state of a successive sequence, the thalamo-cortical circuit classifies analog information into digital patterns, subtracts the information gleaned from these patterns, and then analyzes the residual information remaining after subtraction on the succeeding state; successively, finer and finer information in the whole analog image can then be obtained.

We've built hybrid state machines that can find patterns in an image or sound, recognize syllables in speech, learn a frequently present pattern in an analog input, control other devices, and that even convert an analog current to a digital number by analyzing it at successively finer levels of precision. The latter analog-to-digital converter is built with two spiking electronic neurons configured as an extremely simple hybrid state machine. This converter appears to be the most energy-efficient analog-to-digital converter reported thus far. It is energy efficient because several computations important in analog-to-digital conversion such as amplification, subtraction, and thresholding are extremely easy and natural to implement with spiking neurons. The converter is being applied to digitize biomedical signals that require moderate speed and precision but where energy efficiency is paramount because the electronics is implanted inside the body. Other researchers in the field are using spike-based processing to let chips in a computer talk to each other more efficiently, to model how parts of the brain compute, and to process information in bionic implants.

—R.S.