If you want a computer that works like a living brain you need to build it with the processes of living organisms. That’s the notion underlying neuromorphic engineering, a field of computer science that draws inspiration from biological nervous systems. Conceived in the late 1980s, it’s a fundamental departure from conventional computer designs and it’s shown some big advances lately. A research team announced the creation of a “neurosynaptic” chip that works much like the human brain, and it could be the necessary enabler for the next generation of smart technologies.
This time it’s different
The chip is called “TrueNorth” by its developers, who described its working concepts in the August 8 issue of Science. TrueNorth is roughly the size of a postage stamp, but with the processing potential of a supercomputer. It’s built from conventional silicon transistors, but the resemblance to conventional chips pretty much ends there.
TrueNorth uses 5.4 billion transistors to simulate a huge network of 1 million programmable neurons and 256 million programmable synapses. Like a living nervous system, stimulating the input synapses with data initiates a stream of electrical spikes that are propagated through the neurons. Each neuron simply responds to its input, sending a spike of its own whenever the number of arriving spikes reaches a critical threshold. There is no linear sequence of operations.
“We have not built a brain,” cautioned study leader Dharmendra Modha. “What we have done is learn from the brain’s anatomy and physiology.”
Conventional computer engineering is grounded in concepts defined by John von Neumann after World War II. Most computers include a memory unit and a central processing unit with parts that carry out program logic and program control. They’re great for crunching numbers in sequential chains of calculations, but they have inherent drawbacks. Their sequential calculation processes, for example, can’t perform operations and retrieve instructions at the same time, which limits their speed. Their linear processes must also be coordinated by an internal clock, which means electrons are flowing all the time, generating heat.
TrueNorth processes data asynchronously. A neuron sends a single only when incoming signals sum to a threshold. This means low energy consumption between processing steps. Furthermore, both memory and operations are contained within a single network, saving the time and energy normally invested in accessing the external memory of a von Neumann design.
The TrueNorth chip consumes only about 70 milliwatts – the equivalent of a hearing aid battery, or about 1/10,000 the power of most modern microprocessors. And it’s fast. There are deep differences between neurosynaptic and conventional operating concepts, but at least one metric – calculations per second per watt – has been proposed to compare their performance. According to the research team, an energy-efficient supercomputer scores 4.5 billion on this scale, while TrueNorth scores between 46 and 400 billion.
A potent package
Neuromorphic processing like this could make a variety of existing applications better and make some emerging applications practical.
The chip is small enough to install in almost anything, its low energy needs can free it from external power or from extra batteries, and its processing capabilities can free it from the need for outside data sources. Think of a more powerful Siri that didn’t have to communicate with other computers to generate its answers. Or sensors that could interpret and respond appropriately to images, sounds, and smells in real time. It’s not a leap to imagine distributed networks of devices containing neurosynaptic chips, communicating and coordinating with each other.
These are exactly the kings of capabilities that will be needed for roads filled with driverless cars, for remote medical care, or for communities built on the internet of things – capabilities that required real time sensing, interpreting, and decision-making.
Progress in neuromorphic engineering is picking up speed. Thanks to large sponsorship enterprises like the European Union Human Brain Project and the DARPA SyNAPSE program, brain-based chip designs are emerging from labs around the world. Right now, however, TruNorth appears to have a clear edge among alternatives in its size, computing potential, and extremely lower power consumption.
Success is already pointing out a few gaps that will have to be filled before such chips find wide use. TrueNorth, for example, is configured by choosing which neurons are connected and how strongly they influence one another. It doesn’t really have a programming language like conventional computers, although its research team is working fast to develop one. Whatever comes out of that effort, though, will be as different as the chip itself. A generation of programmers will have to be trained to think in entirely new ways.
Soon, a society surrounded by really, really smart devices may have to think in entirely new ways, too.