Did you think you have the
most powerful computer in front of you?
You are wrong; it is
inside your body!
Results from neuroscience
confirm the astonishing computational power of our brain. Complex tasks like
face recognition is carried out in real time using minimal power beating any
available digital computer. Two decades ago Carver Mead at California Institute
of Technology took up the challenge of making microelectronics based on models
from biology and coined the term; neuromorphic. Since then significant
advances have been made and a number of neuromorphic systems like artificial
eyes and ears have been reported.
While trying to catch the computational paradigms of
biology, the strong connection between the representation of the neural
state and the neural computation itself unveils. Although there are discrete
states in biology, the dominating state variable is continuous, making our
nervous system predominantly an analog computational system with computations
taking place in real and continuous time. In contrast to our completely
discrete or digital systems, biology seems to carry out computational tasks in
a fundamentally different way. Based on badly conditioned data, complex
matching operations are carried out in real time with extremely limited
computational elements with slow and noisy interconnections. Unlike artificial
neural networks used as models for computer programs, neuromorphic systems are
closer to the real neurology carefully modeling both the analog computations and
the neurological representation.
The achievements of microelectronics are evident to everybody, as we are all affected in our everyday life. The famous Moore’s law of doubled performance every second year seems to be unbreakable. The success of digital microelectronic systems, utilizing millions of transistors, is indisputable, while the old art of analog circuit design is still stuck with less than a hundred transistors. While the physical sizes of digital chips have almost stayed constant, the complexity is increasing due to reduction of feature sizes. We still chop up a dinner-plate sized silicon wafer in small thumbnail sized dies. Why not make larger dies? For two reasons:
· Defects – The digital computational paradigm demands perfection. The way digital systems work they do not tolerate one, single transistor not working out of a billion.
· Power – Increasing die-size would demand even more power to be alleviated. Already we need fans to prevent overheating.
Unlike digital systems, our nervous system seems to handle defects nicely. Colorblind men are missing approximately 1/3 of the color sensors in the eye, but they are still doing pretty well. In neuromorphic electronics we are trying to maintain these fault-tolerant and robust properties. It is feasible to use micropower (weak inversion) circuits for neuromorphic electronics due to reasonable precision demands. Since neuromorphic electronics can handle defects and systems may be realized in micropower CMOS, we may exceed the fundamental limits of digital microelectronic systems. In the future we will see neuromorphic systems taking us beyond the reach of digital systems overcoming some of the basic obstacles of technology scaling.
Be practical! How do we make real neuromorphic silicon?
Good old analog circuits like amps and filters are essential, but in neuromorphic systems all the classic tricks-of-the-trade does not always pay off. In larger systems individual tuning of circuits is impossible, so the number of knobs must be reduced to a minimum. High performance must be traded for robustness. In neuromorphic systems current-mode circuit design is often preferred in addition to non-linear circuits and compression techniques like log-domain filters.
An interesting coding found in biology is the spiking behavior of neurons were discrete spikes are fired along nerves (axons). In more engineering terms, the number of spikes pr. second is proportional to the analog input value (pulse-density coding). These spikes are discrete or digital in value but time is still continuous (no clock). This is not the whole story. Important temporal information is also coded with a stochastic behavior. Building spiking neuron circuits in microelectronics is easy. More important it is possible to take advantage of the discrete nature of each spike and make an analog communication channel across a digital bus. Every spiking neuron circuit is given a unique identification. Whenever there is a spike, the neurons ID is encoded as a digital word and sent out on the shared bus. We carefully avoid clocks, but have to manage possible collisions. This coding scheme was proposed by Misha Mahowald a decade ago and was named AER (Address Event Representation). In the receiver end of the bus a simple decoder is converting back to spikes, which may be integrated for reconstruction of the original analog value. With this neuromorphic communication scheme we have established virtual analog wires over a digital bus.
We may take this further and connect these asynchronous buses to computers through a parallel interface. Through that interface it is possible both to read analog values from neuromorphic systems and generate spiking input to neuromrophic systems. In fact we have a simple form of analog-to-digital and digital-to-analog conversion without using data-converters. Involving a computer in the loop is of cause violating the continuous-time nature of our bus and these virtual analog wires are limited to low bandwidth connections.
Be serious! How
may neuromorphic electronics be put to real use?
Neuromorphic electronics has still not matured into an established engineering discipline with products available in the market place (although there are some available). Neuromorphic systems is getting interesting as human computer interaction is shifted from keyboard and screen to more natural means like speech and sound. Much of human information flow is low bandwidth. Looking for yourself in a picture is a simple single bit binary information, but may involve significant computing if the picture is scanned into a computer with high resolution and some face-recognition software is applied for the same task.
Implantable medical electronics is however a niche where neuromorphic electronics comes natural. No doubt we will see an increasing amount of electronics implanted inside our bodies and many of these devices are nerve simulators. The spiky output generated from neuromorphic systems is perfectly suited for nerve stimulation, because the complete system is based on neurological models. One of the more successful implantable microelectronic systems are cochlea implants stimulation the auditory nerve with sound picked up in an external microphone. A neuromorphic cochlea implant is under investigation and preliminary results are promising.
The merits of neuromorphic electronics are still not impressive, but fundamental properties of this way of exploring technology may well bring us beyond the reach of digital systems, both technologically and functionally.