3 Computers That Mimic the Human Brain
For years, researchers have been hard at work on a concept that sounds like something out of an ambitious science fiction movie: computers that mimic the functions and structure of the human brain. While the computers that we use everyday have come a long way in their speed, memory, and capability, computers that function more like the human brain could complete a new array of tasks like operating robots, sensors, or drones, and handling complex analytical tasks that computers can’t currently undertake.
Read on to learn about some of the computers that have attempted to mimic the functions of the human brain and how our growing understanding of the human brain has inspired new types of machines that could eventually lead to more capable and more efficient computers.
DeepMind builds a Neural Turing Machine that mimics the brain’s working memory
Google’s DeepMind built a neural network that can access an external memory, like a Turing machine. As MIT’s Technology Review reports, the computer aims to mimic some of the properties of the brain’s short-term working memory. The computer is a new kind of neural network, adapted to work with an external memory. It learns as it stores memories and can retrieve them later to complete logical tasks — tasks beyond those that it’s been trained to do.
A cognitive psychologist named George Miller discovered in the 1950s that what defines the human brain’s short-term memory isn’t the amount of information that it contains. Instead, according to Miller’s theory, the working memory can hold approximately seven “chunks” of information.
The “chunks” that Miller’s research dealt with ranged from a single digit or letter to a small group of words. They could represent anything from a very small amount of information to a complex idea, equivalent to a much larger amount of information, effectively giving the brain a shortcut to remember large amounts of information.
Technology Review reports that in cognitive science, the ability to understand the components of a sentence and store them in the working memory is called “variable binding.” This ability enables the brain to receive pieces of information and assign it to a place in the working memory. The brain does this repeatedly.
In the 1990s and 2000s, computer scientists began to try designing algorithms, circuits, and neural networks that could mimic the human brain’s working memory. A computer with such brain-like ability would be able to parse out a simple sentence, dividing it into an actor, an action, and the receiver of the action. DeepMind’s new neural network takes on this task but also changes the fundamental nature of a neural network.
Traditionally, a neural network has been built of patterns of connected “neurons” that can change the strength of their connections based on external input. But they lack an external memory — a fundamental of the process of computing — that can be read from and written to during the computational process. So Alex Graves, Greg Wayne, and Ivo Danihelka at DeepMind added an external memory to the neural network, which they then named the Neural Turing Machine.
While the Neural Turing Machine learns from external input like a conventional neural network, it also learns to store and retrieve information. It can learn simple algorithms from example data, and then use those algorithms to generalize far outside its area of training. That ability represents a significant step toward making computers more like the human brain than ever before.
One of the next steps might be to tackle another ability of the brain: to recode the multiple chunks of information that Miller talked about into a single chunk, in a process that enables the brain to make sense of complex arguments. Miller considered this recoding ability to be key to artificial intelligence, and believed that until a computer could reproduce it, it would never match the performance of the human brain.
Stanford researchers develop the Neurogrid circuit, inspired by the human brain
In April, the Stanford University news service reported that bioengineers developed a new circuit modeled on the human brain. Kwabena Boahen and his team of researchers developed Neurogrid, a circuit board comprised of 16 “Neurocore” chips that can simulate 1 million neurons and billions of synaptic connections. The Neurogrid device, about the size of an iPad, can simulate more “orders of magnitude more neurons and synapses” than other brain-mimicking computers are able to, all on the power that it takes to run a tablet.
Boahen plans to lower the costs of building the Neurogrid and then create software that would enable engineers or computer scientists with no knowledge of neuroscience to solve problems like controlling a humanoid robot with Neurogrid.
In its current form, researchers need to know how the human brain works to program the $40,000 prototype. As the news release notes: “Its speed and low power characteristics make Neurogrid ideal for more than just modeling the human brain. Boahen is working with other Stanford scientists to develop prosthetic limbs for paralyzed people that would be controlled by a Neurocore-like chip.”
To make a system affordable enough to be used widely in research, Boahen would change the manufacturing process for the 16 Neurocores — which each supports 65,536 neurons — which relied on 15-year-old fabrication techniques. By shifting to more modern manufacturing processes and fabricating the chips in large volumes, he projects that he could a Neurocore’s cost 100-fold and theoretically build a million-neuron board for only $400 per copy.
IBM’s SyNAPSE project yields the neurosynaptic TrueNorth chip
In IBM’s SyNAPSE project — short for Systems of Neuromorphic Adaptive Plastic Scalable Electronics — researchers took on the task of redesigning computer chips to replicate the ability of neurons to make synaptic connections. As CNET reported at the time, IBM in August unveiled what it called the world’s first neurosynaptic computer chip, a processor that mimcs the human brain’s abilities and power efficiency.
The TrueNorth chip, about the size of a postage stamp, incorporates 5.4 billion transistors, 1 million programmable neurons, and 256 million programmable synapses. While those figures are lower than the 100 billion neurons and 100 trillion to 150 trillion in the human brain, the chip fits supercomputer-like abilities into a much smaller, more efficient microprocessor.
IBM principal investigator and senior manager Dharmendra Modha told CNET that TrueNorth has enough neurons and synapses to run devices that could proactively issue tsunami alerts, complete oil-spill monitoring, or enforce shipping lane rules, all while running on approximately the same amount of power used by a hearing aid.
CNET reports that rather than solving problems through brute-force mathematical calculations, the TrueNorth chip was designed to understand its environment, handle ambiguity, and take action in real time. Potential applications could include powering search-and-rescue robots, helping people with vision impairments to move around safely, or distinguishing between voices in a meeting and creating accurate transcripts for each speaker.
While the TrueNorth chip is still in its prototype phase, it could be just two to three years from its first commercial use. It’s possible that the TrueNorth chip or an innovation like it could help overcome the limitations of the von Neumann architecture, which has formed the core of almost every computer created since 1948.
In contrast to a Turing machine, a machine based on the von Neumann architecture has random-access memory (RAM), which enables each operation to to read or write any memory location. It also has a central processing unit (CPU), with one or more registers that hold data being operated on. Because the processor and memory are separate and data constantly moves between them, delays are inevitable. No matter how quickly a processor can work, the performance of the machine is limited by the rate of transfer between the processor and the memory.
As The New York Times reported when IBM unveiled TrueNorth, the idea that neural networks could be a useful tool for processing information has been around since the 1940s, before the invention of modern computers, but only recently — thanks to gains in memory capacity and processing speed — have neural networks become powerful computing tools. Google, Microsoft, and Apple have all used pattern recognition driven by neural networks to improve services like voice recognition and photo classification.
With TrueNorth, IBM wants to push computers beyond typical “left brain” mathematical tasks to complete “right brain” sensory processing functions with very little power. That would enable chips installed in cars or smartphones to perform calculations in real time, without a connection to the Internet.
Many other brain-inspired projects are currently underway
A variety of other projects are also in various stages of seeking to emulate the human brain’s functions with a computer. The European Union’s Human Brain Project, for instance, is a 10-year undertaking with objectives including the development of neuromorphic computing and neurorobotic systems, as well as the simulation of a human brain on a supercomputer. The U.S. BRAIN project — short for Brain Research through Advancing Innovative Neurotechnologies — challenges scientists to develop new kinds of tools to read out the activity of thousands or even millions of neurons in the brain, and write in complex patterns of activity.
ZDNet reports that researchers at Melbourne’s RMIT University have built a data storage nano structure that mimics the human brain, using a film of oxide material more than 10,000 times thinner than a human hair. The memory’s behavior is dependent on its past experiences, and the research aims to help open the door to the exploration of new materials as flash memory approaches scaling limits.
As part of Heidelberg University’s BrainScales project, researchers are developing analog chips that mimic the behavior of neurons and synapses. The HICANN chip — short for High Input Count Analog Neural Network — would accelerate brain simulations, which would enable researchers to simulate drug interactions that might otherwise take months to play out.
As ComputerWorld reported in May, researchers at Sandia National Laboratories are undertaking a long-term project to build neuro-inspired computers, which would unite processing and memory into a single architecture, so data would be processed and stored by the same components of the machine. Sandia says that researchers will be able to create that architecture in the next few years, but that commercial applications are likely still years away.
Gizmodo reported that researchers at the University of Zurich and ETH Zurich built 11,011 electrodes onto a 2-millimeter-by-2-millimeter piece of silicon, creating a microchip that mimics the human brain to create a microchip that can “feel” and complete complex sensorimotor tasks using the network’s cognitive abilities.
While it will be a monumental task to create a computer that can truly act like the human brain, scientists from a variety of disciplines and backgrounds have demonstrated that they are up to the challenge. The scope and ambition of the projects currently underway are broad, and the global effort to build computers that mimic the human brain will likely continue to yield fascinating inventions and insights into new architectures and materials to make computers more powerful and capable.