We Could Build an Artificial Brain Right Now

Large-scale brainlike systems are possible with existing technology—if we’re willing to spend the money

Photo: Dan Saelinger

Brain-inspired computing is having a moment. Artificial neural network algorithms like deep learning, which are very loosely based on the way the human brain operates, now allow digital computers to perform such extraordinary feats as translating language, hunting for subtle patterns in huge amounts of data, and beating the best human players at Go.

But even as engineers continue to push this mighty computing strategy, the energy efficiency of digital computing is fast approaching its limits. Our data centers and supercomputers already draw megawatts—some 2 percent of the electricity consumed in the United States goes to data centers alone. The human brain, by contrast, runs quite well on about 20 watts, which represents the power produced by just a fraction of the food a person eats each day. If we want to keep improving computing, we will need our computers to become more like our brains.

Hence the recent focus on neuromorphic technology, which promises to move computing beyond simple neural networks and toward circuits that operate more like the brain’s neurons and synapses do. The development of such physical brainlike circuitry is actually pretty far along. Work at my lab and others around the world over the past 35 years has led to artificial neural components like synapses and dendrites that respond to and produce electrical signals much like the real thing.

So, what would it take to integrate these building blocks into a brain-scale computer? 

In 2013, Bo Marr, a former graduate student of mine at Georgia Tech, and I looked at the best engineering and neuroscience knowledge of the time and concluded that it should be possible to build a silicon version of the human cerebral cortex with the transistor technology then in production. What’s more, the resulting machine would take up less than a cubic meter of space and consume less than 100 watts, not too far from the human brain.

That is not to say creating such a computer would be easy. The system we envisioned would still require a few billion dollars to design and build, including some significant packaging innovations to make it compact. There is also the question of how we would program and train the computer. Neuromorphic researchers are still struggling to understand how to make thousands of artificial neurons work together and how to translate brainlike activity into useful engineering applications.

Still, the fact that we can envision such a system means that we may not be far off from smaller-scale chips that could be used in portable and wearable electronics. These gadgets demand low power consumption, and so a highly energy-efficient neuromorphic chip—even if it takes on only a subset of computational tasks, such as signal processing—could be revolutionary. Existing capabilities, like speech recognition, could be extended to handle noisy environments. We could even imagine future smartphones conducting real-time language translation between you and the person you’re talking to. Think of it this way: In the 40 years since the first signal-processing integrated circuits, Moore’s Law has improved energy efficiency by roughly a factor of 1,000. The most brainlike neuromorphic chips could dwarf such improvements, potentially driving down power consumption by another factor of 100 million. That would bring computations that would otherwise need a data center to the palm of your hand.

The ultimate brainlike machine will be one in which we build analogues for all the essential functional components of the brain: 

• the synapses, which connect neurons and allow them to receive and respond to signals; 
• the dendrites, which combine and perform local computations on those incoming signals; and 
• the core, or soma, region of each neuron, which integrates inputs from the dendrites and transmits its output on the axon.

Simple versions of all these basic components have already been implemented in silicon. The starting point for such work is the same metal-oxide-semiconductor field-effect transistor, or MOSFET, that is used by the billions to build the logic circuitry in modern digital processors.

These devices have a lot in common with neurons. Neurons operate using voltage-controlled barriers, and their electrical and chemical activity depends primarily on channels in which ions move between the interior and exterior of the cell—a smooth, analog process that involves a steady buildup or decline instead of a simple on-off operation.

MOSFETs are also voltage controlled and operate by the movement of individual units of charge. And when MOSFETs are operated in the “subthreshold” mode, below the voltage threshold used to digitally switch between on and off, the amount of current flowing through the device is very small—less than a thousandth of what is seen in the typical switching of digital logic gates.

The notion that subthreshold transistor physics could be used to build brainlike circuitry originated with Carver Mead of Caltech, who helped revolutionize the field of very-large-scale circuit design in the 1970s. Mead pointed out that chip designers fail to take advantage of a lot of interesting behavior—and thus information—when they use transistors only for digital logic. The process, he wrote in 1990 [PDF], essentially involves “taking all the beautiful physics that is built into...transistors, mashing it down to a 1 or 0, and then painfully building it back up with AND and OR gates to reinvent the multiply.” A more “physical” or “physics-based” computer could execute more computations per unit energy than its digital counterpart. Mead predicted such a computer would take up significantly less space as well.

In the intervening years, neuromorphic engineers have made all the basic building blocks of the brain out of silicon with a great deal of biological fidelity. The neuron’s dendrite, axon, and soma components can all be fabricated from standard transistors and other circuit elements. In 2005, for example, Ethan Farquhar, then a Ph.D. candidate, and I created a neuron circuit using a set of six MOSFETs and a handful of capacitors. Our model generated electrical pulses that very closely matched those in the soma part of a squid neuron, a long-standing experimental subject. What’s more, our circuit accomplished this feat with similar current levels and energy consumption to those in the squid’s brain. If we had instead used analog circuits to model the equations neuroscientists have developed to describe that behavior, we’d need on the order of 10 times as many transistors. Performing those calculations with a digital computer would require even more space.

Illustration: James Provost Synapses and Soma: The floating-gate transistor [top left], which can store differing amounts of charge, can be used to build a “crossbar” array of artificial synapses [bottom left]. Electronic versions of other neuron components, such as the soma region [right], can be made from standard transistors and other circuit components.

Emulating synapses is a little trickier. A device that behaves like a synapse must have the ability to remember what state it is in, respond in a particular way to an incoming signal, and adapt its response over time.

There are a number of potential approaches to building synapses. The most mature one by far is the single-transistor learning synapse (STLS), a device that my colleagues and I at Caltech worked on in the 1990s while I was a graduate student studying under Mead.

We first presented the STLS in 1994, and it became an important tool for engineers who were building modern analog circuitry, such as physical neural networks. In neural networks, each node in the network has a weight associated with it, and those weights determine how data from different nodes are combined. The STLS was the first device that could hold a variety of different weights and be reprogrammed on the fly. The device is also nonvolatile, which means that it remembers its state even when not in use—a capability that significantly reduces how much energy it needs.

The STLS is a type of floating-gate transistor, a device that is used to build memory cells in flash memory. In an ordinary MOSFET, a gate controls the flow of electricity through a current-carrying channel. A floating-gate transistor has a second gate that sits between this electrical gate and the channel. This floating gate is not directly connected to ground or any other component. Thanks to that electrical isolation, which is enhanced by high-quality silicon-insulator interfaces, charges remain in the floating gate for a long time. The floating gate can take on many different amounts of charge and so have many different levels of electrical response, an essential requisite for creating an artificial synapse capable of varying its response to stimuli.

My colleagues and I used the STLS to demonstrate the first crossbar network, a computational model currently popular with nanodevice researchers. In this two-dimensional array, devices sit at the intersection of input lines running north-south and output lines running east-west. This configuration is useful because it lets you program the connection strength of each “synapse” individually, without disturbing the other elements in the array.

Thanks in part to a recent Defense Advanced Research Projects Agency program called SyNAPSE, the neuromorphic engineering field has seen a surge of research into artificial synapses built from nanodevices such as

• memristors, 
• resistive RAM, and 
• phase-change memories (as well as floating-gate devices). 

But it will be hard for these new artificial synapses to improve on our two-decade-old floating-gate arrays. Memristors and other novel memories come with programming challenges; some have device architectures that make it difficult to target a single specific device in a crossbar array. Others need a dedicated transistor in order to be programmed, adding significantly to their footprint. Because floating-gate memory is programmable over a wide range of values, it can be more easily fine-tuned to compensate for manufacturing variation from device to device than can many nanodevices. A number of neuromorphic research groups that tried integrating nanodevices into their designs have recently come around to using floating-gate devices.

So how will we put all these brainlike components together? 
In the human brain, of course, neurons and synapses are intermingled. Neuromorphic chip designers must take a more integrated approach as well, with all neural components on the same chip, tightly mixed together. This is not the case in many neuromorphic labs today: To make research projects more manageable, different building blocks may be placed in different areas. Synapses, for example, may be relegated to an off-chip array. Connections may be routed through another chip called a field-programmable gate array, or FPGA.

But as we scale up neuromorphic systems, we’ll need to take care that we don’t replicate the arrangement in today’s computers, which lose a significant amount of energy driving bits back and forth between logic, memory, and storage. Today, a computer can easily consume 10 times the energy to move the data needed for a multiple-accumulate operation—a common signal-processing computation—as it does to perform the calculation.

The brain, by contrast, minimizes the energy cost of communication by keeping operations highly local. The memory elements of the brain, such as synaptic strengths, are mixed in with the neural components that integrate signals. And the brain’s “wires”—the dendrites and axons that extend from neurons to transmit, respectively, incoming signals and outgoing pulses—are generally fairly short relative to the size of the brain, so they don’t require large amounts of energy to sustain a signal. From anatomical data, we know that more than 90 percent of neurons connect with only their nearest 1,000 or so neighbors.

Another big question for the builders of brainlike chips and computers is the algorithms we will run on them. Even a loosely brain-inspired system can have a big advantage over digital systems. In 2004, for example, my group used floating-gate devices to perform multiplications for signal processing with just 1/1,000 the energy and 1/100 the area of a comparable digital system. In the years since, other researchers and my group have successfully demonstrated neuromorphic approaches to many other kinds of signal-processing calculations.

But the brain is still 100,000 times as efficient as the systems in these demonstrations. That’s because while our current neuromorphic technology takes advantage of the neuronlike physics of transistors, it doesn’t consider the algorithms the brain uses to perform its operations.

Today, we are just beginning to discover these physical algorithms—that is, the processes that will allow brainlike chips to operate with more brainlike efficiency. Four years ago, my research group used silicon somas, synapses, and dendrites to perform a word-spotting algorithm that identifies words in a speech waveform. This physical algorithm exhibited a thousandfold improvement in energy efficiency over predicted analog signal processing. Eventually, by lowering the amount of voltage supplied to the chips and using smaller transistors, researchers should be able to build chips that parallel the brain in efficiency for a range of computations.

When I started in neuromorphic research 30 years ago, everyone believed tremendous opportunities would arise from designing systems that are more like the brain. And indeed, entire industries are now being built around brain-inspired AI and deep learning, with applications that promise to transform—among other things—our mobile devices, our financial institutions, and how we interact in public spaces.

And yet, these applications rely only slightly on what we know about how the brain actually works. The next 30 years will undoubtedly see the incorporation of more such knowledge. We already have much of the basic hardware we need to accomplish this neuroscience-to-computing translation. But we must develop a better understanding of how that hardware should behave—and what computational schemes will yield the greatest real-world benefits.

Consider this a call to action. We have come pretty far with a very loose model of how the brain works. But neuroscience could lead to far more sophisticated brainlike computers. And what greater feat could there be than using our own brains to learn how to build new ones?

This article appears in the June 2017 print issue as “A Road Map for the Artificial Brain.”

About the Author

Jennifer Hasler is a professor of electrical and computer engineering at the Georgia Institute of Technology.

ORIGINAL: IEEE Spectrum

By JENNIFER HASLER 

Posted 1 Jun 2017 | 19:00 GMT

Google's Deep Mind Gives AI a Memory Boost That Lets It Navigate London's Underground

Photo: iStockphoto

Google’s DeepMind artificial intelligence lab does more than just develop computer programs capable of beating the world’s best human players in the ancient game of Go. The DeepMind unit has also been working on the next generation of deep learning software that combines the ability to recognize data patterns with the memory required to decipher more complex relationships within the data.

Deep learning is the latest buzz word for artificial intelligence algorithms called neural networks that can learn over time by filtering huge amounts of relevant data through many “deep” layers. The brain-inspired neural network layers consist of nodes (also known as neurons). Tech giants such as Google, Facebook, Amazon, and Microsoft have been training neural networks to learn how to better handle tasks such as recognizing images of dogs or making better Chinese-to-English translations. These AI capabilities have already benefited millions of people using Google Translate and other online services.

But neural networks face huge challenges when they try to rely solely on pattern recognition without having the external memory to store and retrieve information. To improve deep learning’s capabilities, Google DeepMind created a “differentiable neural computer” (DNC) that gives neural networks an external memory for storing information for later use.

“Neural networks are like the human brain; we humans cannot assimilate massive amounts of data and we must rely on external read-write memory all the time,” says Jay McClelland, director of the Center for Mind, Brain and Computation at Stanford University. “We once relied on our physical address books and Rolodexes; now of course we rely on the read-write storage capabilities of regular computers.”

McClelland is a cognitive scientist who served as one of several independent peer reviewers for the Google DeepMind paper that describes development of this improved deep learning system. The full paper is presented in the 12 Oct 2016 issue of the journal Nature.

The DeepMind team found that the DNC system’s combination of the neural network and external memory did much better than a neural network alone in tackling the complex relationships between data points in so-called “graph tasks.” For example, they asked their system to either simply take any path between points A and B or to find the shortest travel routes based on a symbolic map of the London Underground subway.

An unaided neural network could not even finish the first level of training, based on traveling between two subway stations without trying to find the shortest route. It achieved an average accuracy of just 37 percent after going through almost two million training examples. By comparison, the neural network with access to external memory in the DNC system successfully completed the entire training curriculum and reached an average of 98.8 percent accuracy on the final lesson.

The external memory of the DNC system also proved critical to success in performing logical planning tasks such as solving simple block puzzle challenges. Again, a neural network by itself could not even finish the first lesson of the training curriculum for the block puzzle challenge. The DNC system was able to use its memory to store information about the challenge’s goals and to effectively plan ahead by writing its decisions to memory before acting upon them.

In 2014, DeepMind’s researchers developed another system, called the neural Turing machine, that also combined neural networks with external memory. But the neural Turing machine was limited in the way it could access “memories” (information) because such memories were effectively stored and retrieved in fixed blocks or arrays. The latest DNC system can access memories in any arbitrary location, McClelland explains.

The DNC system’s memory architecture even bears a certain resemblance to how the hippocampus region of the brain supports new brain cell growth and new connections in order to store new memories. Just as the DNC system uses the equivalent of time stamps to organize the storage and retrieval of memories, human “free recall” experiments have shown that people are more likely to recall certain items in the same order as first presented.

Despite these similarities, the DNC’s design was driven by computational considerations rather than taking direct inspiration from biological brains, DeepMind’s researchers write in their paper. But McClelland says that he prefers not to think of the similarities as being purely coincidental.

“The design decisions that motivated the architects of the DNC were the same as those that structured the human memory system, although the latter (in my opinion) was designed by a gradual evolutionary process, rather than by a group of brilliant AI researchers,” McClelland says.

Human brains still have significant advantages over any brain-inspired deep learning software. For example, human memory seems much better at storing information so that it is accessible by both context or content, McClelland says. He expressed hope that future deep learning and AI research could better capture the memory advantages of biological brains.

DeepMind’s DNC system and similar neural learning systems may represent crucial steps for the ongoing development of AI. But the DNC system still falls well short of what McClelland considers the most important parts of human intelligence.

The DNC is a sophisticated form of external memory, but ultimately it is like the papyrus on which Euclid wrote the elements. The insights of mathematicians that Euclid codified relied (in my view) on a gradual learning process that structured the neural circuits in their brains so that they came to be able to see relationships that others had not seen, and that structured the neural circuits in Euclid’s brain so that he could formulate what to write. We have a long way to go before we understand fully the algorithms the human brain uses to support these processes. 

It’s unclear when or how Google might take advantage of the capabilities offered by the DNC system to boost its commercial products and services. The DeepMind team was “heads down in research” or too busy with travel to entertain media questions at this time, according to a Google spokesperson.

But Herbert Jaeger, professor for computational science at Jacobs University Bremen in Germany, sees the DeepMind team’s work as a “passing snapshot in a fast evolution sequence of novel neural learning architectures.” In fact, he’s confident that the DeepMind team already has something better than the DNC system described in the Nature paper. (Keep in mind that the paper was submitted back in January 2016.)

DeepMind’s work is also part of a bigger trend in deep learning, Jaeger says. The leading deep learning teams at Google and other companies are racing to build new AI architectures with many different functional modules—among them, attentional control or working memory; they then train the systems through deep learning. 

“The DNC is just one among dozens of novel, highly potent, and cleverly-thought-out neural learning systems that are popping up all over the place,” Jaeger says.

ORIGINAL: IEEE Spectrum

By Jeremy Hsu

12 Oct 2016

Former NASA chief unveils $100 million neural chip maker KnuEdge

Add caption

It’s not all that easy to call KnuEdge a startup. Created a decade ago by Daniel Goldin, the former head of the National Aeronautics and Space Administration, KnuEdge is only now coming out of stealth mode. It has already raised $100 million in funding to build a “neural chip” that Goldin says will make data centers more efficient in a hyperscale age.

Goldin, who founded the San Diego, California-based company with the former chief technology officer of NASA, said he believes the company’s brain-like chip will be far more cost and power efficient than current chips based on the computer design popularized by computer architect John von Neumann. In von Neumann machines, memory and processor are separated and linked via a data pathway known as a bus. Over the years, von Neumann machines have gotten faster by sending more and more data at higher speeds across the bus as processor and memory interact. But the speed of a computer is often limited by the capacity of that bus, leading to what some computer scientists to call the “von Neumann bottleneck.” IBM has seen the same problem, and it has a research team working on brain-like data center chips. Both efforts are part of an attempt to deal with the explosion of data driven by artificial intelligence and machine learning.

Goldin’s company is doing something similar to IBM, but only on the surface. Its approach is much different, and it has been secretly funded by unknown angel investors. And Goldin said in an interview with VentureBeat that the company has already generated $20 million in revenue and is actively engaged in hyperscale computing companies and Fortune 500 companies in the aerospace, banking, health care, hospitality, and insurance industries. The mission is a fundamental transformation of the computing world, Goldin said.

“It all started over a mission to Mars,” Goldin said.

Above: KnuEdge’s first chip has 256 cores.Image Credit: KnuEdge

Back in the year 2000, Goldin saw that the time delay for controlling a space vehicle would be too long, so the vehicle would have to operate itself. He calculated that a mission to Mars would take software that would push technology to the limit, with more than tens of millions of lines of code.

Above: Daniel Goldin, CEO of KnuEdge.

Image Credit: KnuEdge

“I thought, holy smokes,” he said. “It’s going to be too expensive. It’s not propulsion. It’s not environmental control. It’s not power. This software business is a very big problem, and that nation couldn’t afford it.”

So Goldin looked further into the brains of the robotics, and that’s when he started thinking about the computing it would take.

Asked if it was easier to run NASA or a startup, Goldin let out a guffaw.

“I love them both, but they’re both very different,” Goldin said. “At NASA, I spent a lot of time on non-technical issues. I had a project every quarter, and I didn’t want to become dull technically. I tried to always take on a technical job doing architecture, working with a design team, and always doing something leading edge. I grew up at a time when you graduated from a university and went to work for someone else. If I ever come back to this earth, I would graduate and become an entrepreneur. This is so wonderful.”

Back in 1992, Goldin was planning on starting a wireless company as an entrepreneur. But then he got the call to “go serve the country,” and he did that work for a decade. He started KnuEdge (previously called Intellisis) in 2005, and he got very patient capital.

“When I went out to find investors, I knew I couldn’t use the conventional Silicon Valley approach (impatient capital),” he said. “It is a fabulous approach that has generated incredible wealth. But I wanted to undertake revolutionary technology development. To build the future tools for next-generation machine learning, improving the natural interface between humans and machines. So I got patient capital that wanted to see lightning strike. Between all of us, we have a board of directors that can contact almost anyone in the world. They’re fabulous business people and technologists. We knew we had a ten-year run-up.”

But he’s not saying who those people are yet.

KnuEdge’s chips are part of a larger platform. KnuEdge is also unveiling KnuVerse, a military-grade voice recognition and authentication technology that unlocks the potential of voice interfaces to power next-generation computing, Goldin said.

While the voice technology market has exploded over the past five years due to the introductions of Siri, Cortana, Google Home, Echo, and ViV, the aspirations of most commercial voice technology teams are still on hold because of security and noise issues. KnuVerse solutions are based on patented authentication techniques using the human voice — even in extremely noisy environments — as one of the most secure forms of biometrics. Secure voice recognition has applications in industries such as banking, entertainment, and hospitality.

KnuEdge says it is now possible to authenticate to computers, web and mobile apps, and Internet of Things devices (or everyday objects that are smart and connected) with only a few words spoken into a microphone — in any language, no matter how loud the background environment or how many other people are talking nearby. In addition to KnuVerse, KnuEdge offers Knurld.io for application developers, a software development kit, and a cloud-based voice recognition and authentication service that can be integrated into an app typically within two hours.

And KnuEdge is announcing KnuPath with LambdaFabric computing. KnuEdge’s first chip, built with an older manufacturing technology, has 256 cores, or neuron-like brain cells, on a single chip. Each core is a tiny digital signal processor. The LambdaFabric makes it possible to instantly connect those cores to each other — a trick that helps overcome one of the major problems of multicore chips, Goldin said. The LambdaFabric is designed to connect up to 512,000 devices, enabling the system to be used in the most demanding computing environments. From rack to rack, the fabric has a latency (or interaction delay) of only 400 nanoseconds. And the whole system is designed to use a low amount of power.

All of the company’s designs are built on biological principles about how the brain gets a lot of computing work done with a small amount of power. The chip is based on what Goldin calls “sparse matrix heterogeneous machine learning algorithms.” And it will run C++ software, something that is already very popular. Programmers can program each one of the cores with a different algorithm to run simultaneously, for the “ultimate in heterogeneity.” It’s multiple input, multiple data, and “that gives us some of our power,” Goldin said.

Above: KnuEdge’s KnuPath chip.

Image Credit: KnuEdge

“KnuEdge is emerging out of stealth mode to aim its new Voice and Machine Learning technologies at key challenges in IoT, cloud based machine learning and pattern recognition,” said Paul Teich, principal analyst at Tirias Research, in a statement. “Dan Goldin used his experience in transforming technology to charter KnuEdge with a bold idea, with the patience of longer development timelines and away from typical startup hype and practices. The result is a new and cutting-edge path for neural computing acceleration. There is also a refreshing surprise element to KnuEdge announcing a relevant new architecture that is ready to ship… not just a concept or early prototype.”

Today, Goldin said the company is ready to show off its designs. The first chip was ready last December, and KnuEdge is sharing it with potential customers. That chip was built with a 32-nanometer manufacturing process, and even though that’s an older technology, it is a powerful chip, Goldin said. Even at 32 nanometers, the chip has something like a two-times to six-times performance advantage over similar chips, KnuEdge said.

“The human brain has a couple of hundred billion neurons, and each neuron is connected to at least 10,000 to 100,000 neurons,” Goldin said. “And the brain is the most energy efficient and powerful computer in the world. That is the metaphor we are using.”

KnuEdge has a new version of its chip under design. And the company has already generated revenue from sales of the prototype systems. Each board has about four chips.

As for the competition from IBM, Goldin said, “I believe we made the right decision and are going in the right direction. IBM’s approach is very different from what we have. We are not aiming at anyone. We are aiming at the future.”

In his NASA days, Goldin had a lot of successes. There, he redesigned and delivered the International Space Station, tripled the number of space flights, and put a record number of people into space, all while reducing the agency’s planned budget by 25 percent. He also spent 25 years at TRW, where he led the development of satellite television services.

KnuEdge has 100 employees, but Goldin said the company outsources almost everything. Goldin said he is planning to raised a round of funding late this year or early next year. The company collaborated with the University of California at San Diego and UCSD’s California Institute for Telecommunications and Information Technology.

With computers that can handle natural language systems, many people in the world who can’t read or write will be able to fend for themselves more easily, Goldin said.

“I want to be able to take machine learning and help people communicate and make a living,” he said. “This is just the beginning. This is the Wild West. We are talking to very large companies about this, and they are getting very excited.”

A sample application is a home that has much greater self-awareness. If there’s something wrong in the house, the KnuEdge system could analyze it and figure out if it needs to alert the homeowner.

Goldin said it was hard to keep the company secret.

“I’ve been biting my lip for ten years,” he said.

As for whether KnuEdge’s technology could be used to send people to Mars, Goldin said. “This is available to whoever is going to Mars. I tried twice. I would love it if they use it to get there.”

ORIGINAL: Venture Beat

A Scale-up Synaptic Supercomputer (NS16e): Four Perspectives

Today, Lawrence Livermore National Lab (LLNL) and IBM announce the development of a new Scale-up Synaptic Supercomputer (NS16e) that highly integrates 16 TrueNorth Chips in a 4×4 array to deliver 16 million neurons and 256 million synapses. LLNL will also receive an end-to-end software ecosystem that consists of a simulator; a programming language; an integrated programming environment; a library of algorithms as well as applications; firmware; tools for composing neural networks for deep learning; a teaching curriculum; and cloud enablement. Also, don't miss the story in The Wall Street Journal (sign-in required) and the perspective and a video by LLNL's Brian Van Essen.

To provide insights into what it took to achieve this significant milestone in the history of our project, following are four intertwined perspectives from my colleagues:

• Filipp Akopyan -- First Steps to an Efficient Scalable NeuroSynaptic Supercomputer.
• Bill Risk and Ben Shaw -- Creating an Iconic Enclosure for the NS16e.
• Jun Sawada -- NS16e System as a Neural Network Development Workstation.
• Brian Taba -- How to Program a Synaptic Supercomputer.

The following timeline provides context for today's milestone in terms of the continued evolution of our project.

Illustration Credit: William Risk

IBM’S ‘Rodent Brain’ Chip Could Make Our Phones Hyper-Smart

At a lab near San Jose, IBM has built the digital equivalent of a rodent brain---roughly speaking. It spans 48 of the company's experimental TrueNorth chips, a new breed of processor that mimics the brain's biological building blocks. IBM

DHARMENDRA MODHA WALKS me to the front of the room so I can see it up close. About the size of a bathroom medicine cabinet, it rests on a table against the wall, and thanks to the translucent plastic on the outside, I can see the computer chips and the circuit boards and the multi-colored lights on the inside. It looks like a prop from a ’70s sci-fi movie, but Modha describes it differently. “You’re looking at a small rodent,” he says.

He means the brain of a small rodent—or, at least, the digital equivalent. The chips on the inside are designed to behave like neurons—the basic building blocks of biological brains. Modha says the system in front of us spans 48 million of these artificial nerve cells, roughly the number of neurons packed into the head of a rodent.

Modha oversees the cognitive computing group at IBM, the company that created these “neuromorphic” chips. For the first time, he and his team are sharing their unusual creations with the outside world, running a three-week “boot camp” for academics and government researchers at an IBM R&D lab on the far side of Silicon Valley. Plugging their laptops into the digital rodent brain at the front of the room, this eclectic group of computer scientists is exploring the particulars of IBM’s architecture and beginning to build software for the chip dubbed TrueNorth.

'We want to get as close to the brain as possible while maintaining flexibility.'DHARMENDRA MODHA, IBM

Some researchers who got their hands on the chip at an engineering workshop in Colorado the previous month have already fashioned software that can identify images, recognize spoken words, and understand natural language. Basically, they’re using the chip to run “deep learning” algorithms, the same algorithms that drive the internet’s latest AI services, including the face recognition on Facebook and the instant language translation on Microsoft’s Skype. But the promise is that IBM’s chip can run these algorithms in smaller spaces with considerably less electrical power, letting us shoehorn more AI onto phones and other tiny devices, including hearing aids and, well, wristwatches.

“What does a neuro-synaptic architecture give us? It lets us do things like image classification at a very, very low power consumption,” says Brian Van Essen, a computer scientist at the Lawrence Livermore National Laboratory who’s exploring how deep learning could be applied to national security. “It lets us tackle new problems in new environments.”

The TrueNorth is part of a widespread movement to refine the hardware that drives deep learning and other AI services. Companies like Google and Facebook and Microsoft are now running their algorithms on machines backed with GPUs (chips originally built to render computer graphics), and they’re moving towards FPGAs (chips you can program for particular tasks). For Peter Diehl, a PhD student in the cortical computation group at ETH Zurich and University Zurich, TrueNorth outperforms GPUs and FPGAs in certain situations because it consumes so little power.

The main difference, says Jason Mars, a professor of a computer science at the University of Michigan, is that the TrueNorth dovetails so well with deep-learning algorithms. These algorithms mimic neural networks in much the same way IBM’s chips do, recreating the neurons and synapses in the brain. One maps well onto the other. “The chip gives you a highly efficient way of executing neural networks,” says Mars, who declined an invitation to this month’s boot camp but has closely followed the progress of the chip.

That said, the TrueNorth suits only part of the deep learning process—at least as the chip exists today—and some question how big an impact it will have. Though IBM is now sharing the chips with outside researchers, it’s years away from the market. For Modha, however, this is as it should be. As he puts it: “We’re trying to lay the foundation for significant change.”

The Brain on a Phone

Peter Diehl recently took a trip to China, where his smartphone didn’t have access to the `net, an experience that cast the limitations of today’s AI in sharp relief. Without the internet, he couldn’t use a service like Google Now, which applies deep learning to speech recognition and natural language processing, because most the computing takes place not on the phone but on Google’s distant servers. “The whole system breaks down,” he says.

Deep learning, you see, requires enormous amounts of processing power—processing power that’s typically provided by the massive data centers that your phone connects to over the `net rather than locally on an individual device. The idea behind TrueNorth is that it can help move at least some of this processing power onto the phone and other personal devices, something that can significantly expand the AI available to everyday people.

To understand this, you have to understand how deep learning works. It operates in two stages. 

• First, companies like Google and Facebook must train a neural network to perform a particular task. If they want to automatically identify cat photos, for instance, they must feed the neural net lots and lots of cat photos. 
• Then, once the model is trained, another neural network must actually execute the task. You provide a photo and the system tells you whether it includes a cat. The TrueNorth, as it exists today, aims to facilitate that second stage.

Once a model is trained in a massive computer data center, the chip helps you execute the model. And because it’s small and uses so little power, it can fit onto a handheld device. This lets you do more at a faster speed, since you don’t have to send data over a network. If it becomes widely used, it could take much of the burden off data centers. “This is the future,” Mars says. “We’re going to see more of the processing on the devices.”

Neurons, Axons, Synapses, Spikes

Google recently discussed its efforts to run neural networks on phones, but for Diehl, the TrueNorth could take this concept several steps further. The difference, he explains, is that the chip dovetails so well with deep learning algorithms. Each chip mimics about a million neurons, and these can communicate with each other via something similar to a synapse, the connections between neurons in the brain.

'Silicon operates in a very different way than the stuff our brains are made of.'

The setup is quite different than what you find in chips on the market today, including GPUs and FPGAs. Whereas these chips are wired to execute particular “instructions,” the TrueNorth juggles “spikes,” much simpler pieces of information analogous to the pulses of electricity in the brain. Spikes, for instance, can show the changes in someone’s voice as they speak—or changes in color from pixel to pixel in a photo. “You can think of it as a one-bit message sent from one neuron to another.” says Rodrigo Alvarez-Icaza, one of the chip’s chief designers.

The upshot is a much simpler architecture that consumes less power. Though the chip contains 5.4 billion transistors, it draws about 70 milliwatts of power. A standard Intel computer processor, by comparison, includes 1.4 billion transistors and consumes about 35 to 140 watts. Even the ARM chips that drive smartphones consume several times more power than the TrueNorth.

Of course, using such a chip also requires a new breed of software. That’s what researchers like Diehl are exploring at the TrueNorth boot camp, which began in early August and runs for another week at IBM’s research lab in San Jose, California. In some cases, researchers are translating existing code into the “spikes” that the chip can read (and back again). But they’re also working to build native code for the chip.

Parting Gift

Like these researchers, Modha discusses the TrueNorth mainly in biological terms. Neurons. Axons. Synapses. Spikes. And certainly, the chip mirrors such wetware in some ways. But the analogy has its limits. “That kind of talk always puts up warning flags,” says Chris Nicholson, the co-founder of deep learning startup Skymind. “Silicon operates in a very different way than the stuff our brains are made of.”

Modha admits as much. When he started the project in 2008, backed by $53.5M in funding from Darpa, the research arm for the Department of Defense, the aim was to mimic the brain in a more complete way using an entirely different breed of chip material. But at one point, he realized this wasn’t going to happen anytime soon. “Ambitions must be balanced with reality,” he says.

In 2010, while laid up in bed with the swine flu, he realized that the best way forward was a chip architecture that loosely mimicked the brain—an architecture that could eventually recreate the brain in more complete ways as new hardware materials were developed. “You don’t need to model the fundamental physics and chemistry and biology of the neurons to elicit useful computation,” he says. “We want to get as close to the brain as possible while maintaining flexibility.”

This is TrueNorth. It’s not a digital brain. But it is a step toward a digital brain. And with IBM’s boot camp, the project is accelerating. The machine at the front of the room is really 48 separate machines, each built around its own TrueNorth processors. Next week, as the boot camp comes to a close, Modha and his team will separate them and let all those academics and researchers carry them back to their own labs, which span over 30 institutions on five continents. “Humans use technology to transform society,” Modha says, pointing to the room of researchers. “These are the humans.”.

ORIGINAL: Wired

By CADE METZ

08.17.15

BrainCard, pattern recognition for ALL

BrainCard, pattern recognition for ALL Patern Recognition, Brain Card, Image Recognition, AI, Sound Recognition, Biosensors, speech recognition, Intel Edison, Neuromorphic Computing, General Vision, NeuroMem CM1K,

ORIGINAL: IndieGogo

Embedded recognition for images, speech, sound, biosensors or any signal with zero programming.

Petaluma, California, United States Technology

Text and Numbers

 

Pattern & image recognition module with neuromorphic learning for all your maker projects.

Robotics fans, drone pilots, hackers & data-miners - rejoice!

The BrainCard is an open source hardware platform featuring the worlds only fully functional and field-tested Neuromorphic Chip containing 1024 silicon neurons. It is able to learn and recognize patterns within any dataset generated by any source, from the physical (sensors), to the virtual (data).  

Offered here, for the first time, to makers in a format compatible with nearly all other popular electronics platforms — from Raspberry Pi to Arduino and Intel Edison —  we aim to help you add cognitive perception to any electronics project.

Add a brain to: Robots, toys or an old GoPro. Give them the ability to recognize and recall almost anything... You can also add a brain to any digital cameras including dash cams. Vision not your thing? The same technology can recognize patterns in data like that packet of code you're looking for in a sea of C++, a phrase in an eBook (regardless of the books length), even real time data: Build your own biosensors!  Make any appliance you like “smart”, like a coffee pot that recognizes you and starts making your coffee the way you like best.

Simply put; make it think.

The BrainCard is an open source hardware platform featuring the worlds only fully functional and field-tested Neuromorphic Chip containing 1024 silicon neurons. It is able to learn and recognize patterns within any dataset generated by any source, from the physical (sensors), to the virtual (data). Offered here, for the first time, to makers in a format compatible with nearly all other popular electronics platforms — from Raspberry Pi to Arduino and Intel Edison —  we aim to help you add cognitive perception to any electronics project.

Add a brain to: Robots, toys or an old GoPro. Give them the ability to recognize and recall almost anything... You can also add a brain to any digital cameras including dash cams. Vision not your thing? The same technology can recognize patterns in data like that packet of code you're looking for in a sea of C++, a phrase in an eBook (regardless of the books length), even real time data: Build your own biosensors!  Make any appliance you like “smart”, like a coffee pot that recognizes you and starts making your coffee the way you like best.

Simply put; make it think.

Cannot wait for technical details?
Before we carry on, for those of you that are quick studies and/or already know everything, we thought you might like to skip straight to the specs so here you go:

BrainCard Specifications (Hardware and API)

For everyone else - please read on...

Unfamiliar with Neural Networks or Neuromorphic Chips? Watch this:

(If you want some more background info, click here)

Now back to you project...

The BrainCard™ is a small electronics board with a NeuroMem® CM1K device plus a FPGA (Field Programmable Gate Array) chip to connect to platform buses and sensor inputs. There is even an optional image sensor featured on the BrainCard 1KIS (Image Sensor) version. It can be connected to almost any popular electronics platform including Arduino/Raspberry Pi/Intel Edison and enables users to massively boost any devices capability by creating a brain-like system architecture – hence the name.

The CM1K chip(s) on the BrainCard essentially acts as a right-brain hemisphere ready to learn, recognize and recall patterns/images/sounds/inputs from any incoming data stream. This allows the accompanying MPU device to concentrate on what it’s good at — left-brain functions such as logic, procedural computing and as a communications and I/O interface.

The BrainCard is an open source hardware platform featuring the world's only fully functional and field-tested Neuromorphic Chip containing 1024 silicon neurons. It is able to learn and recognize patterns within any dataset generated by any source, from the physical (sensors), to the virtual (data). Offered here, for the first time, to makers in a format compatible with nearly all other popular electronics platforms — from Raspberry Pi to Arduino and Intel Edison —  we aim to help you add cognitive perception to any electronics project.

Add a brain to: Robots, toys or an old GoPro. Give them the ability to recognize and recall almost anything... You can also add a brain to any digital cameras including dash cams. Vision not your thing? The same technology can recognize patterns in data like that packet of code you're looking for in a sea of C++, a phrase in an eBook (regardless of the books length), even real time data: Build your own biosensors!  Make any appliance you like “smart”, like a coffee pot that recognizes you and starts making your coffee the way you like best.
Simply put; make it think.


The key to success is teaching BrainCard as you would a child: Teach it too conservatively and it will not generalize enough; too moderately and it could get confused. It is not like traditional programming and we have found that part of the fun in building projects with the BrainCard is in this new learning parameter.

It’s really quite simple: Show the BrainCard what it must recognize and assign the example a category. So: This face is John, that voice is Emma, this vibration is made by your cat purring and so on.

Getting started:

The BrainCard is delivered with a default configuration which can communicate with either one of the proposed controllers (Arduino, Raspberry PI or Edison) through a same communication protocol over their SPI lines.  Access to generic pattern learning and recognition functions using the CM1K chip are made through a simple API delivered for the different IDE (Arduino and Eclipse). More specific function libraries will be released shortly after and we hope to start a repository of your libraries too! 

1. Install and connect the BrainCard to the MPU/Device of your choice. View the hardware datasheet
   
2. Install the API in the IDE of your choice (Arduino, Eclipse). View the BrainCard API preliminary datasheet
   
3. Now, you can program to teach the BrainCard using examples previously collected and saved to disk (waveforms, images, movies). Or you can program some GPIOs to trigger teaching (bush buttons, keyboard inputs and even voice control! As illustrated in the following video, teaching amounts to selecting examples and sending one of more signatures of this example to the neurons of the BrainCard. The neurons will decide if the example is worth learning based on what they already know. If applicable, some neurons will autonomosuly correct themselfves if they contradict the teacher and never repeat this mistake again.
4. Recognition is the same as learning except that this time, your program monitors the response of the neurons to the incoming signatures instead of sending them learning commands. Your program can then act based on what is recognized using the wealth of GPIOs available through Arduino Shields, as well as  DeviceToDevice or DeviceToCloud communications, and more. 

So what can it do?

This is a great question, as even we have not fully explored the full range of the BrainCard/CM1K’s capabilities. Almost every day we are coming up with new applications for the technology, which is one of our quandaries, and is where YOU come in. It’s also why we are choosing to announce ourselves to the world via Indiegogo.

A simple list of known capabilities 

Object recognition

Using the KIS vesion or an off-the-shelf image sensor of your own and teach your BrainCard to recognize shapes, colors, objects, signs, people and animals.

Stereoscopic vision

With two image sensors attached, along with a CPU, your project can work in stereoscopic vision! The processor can triangulate distance and the CM1K can recognize what it’s looking at. Add some motors to the image sensors and it can track things too.

Audio RecognitionAttach a microphone and teach the BrainCard to recognize a noise, a voice, YOUR voice or other audio signals like a bird song or a dog.

Vibration and motionAttach a MEMS (Micro Electrical Mechanical Systems) device and teach the BrainCard to recognize vibrations or physical motion.

Bio signals
BrainCard can recognize data from any Bio-signal source – such as:

Electroencephalogram (EEG), Electrocardiogram (ECG), Electromyogram (EMG), Mechanomyogram (MMG), Electrooculography (EOG), Galvanic skin response (GSR), Magnetoencephalogram (MEG).

Text and Numbers
You can run your data through the BrainCard in any form — from text to binary to DNA sequences — and teach it to recognize patterns, which will allow it to detect anomalies, identify clusters and make predictions.

There are MANY MORE applications we just haven't tried yet...

Flexibility
If you go crazy while teaching and fill all 1024 neurons on a chip, don’t panic. BrainCard provides an expansion bus to stack more CM1K chips in boards of two, thereby increasing the number of modules (subject to availability) you can teach by increments of 2048 (1x CM1K equals 1,024 neurons). This expansion can be done at any time to its maximum of 8,192 (plus the original 1024 on the BrainCard), and will not impact your teaching allowing you to experiment to your heart’s content.

Maturity
The NeuroMem CM1K technology has already found many applications in industry and has been working in the real world since 2007 – so we know everything we’re claiming above is 100% true, because most of these applications have been built somewhere.

What we need, and what you getThis Indiegogo campaign has been launched with one aim: To generate the volume and revenue we need to manufacture the maker version of the CM1K technology — the BrainCard.

By supporting this Indiegogo project you will be a part of the first chapter of a much bigger story: We aim to change the way the world computes with neural network technology. We're looking to raise at least $200k to start manufacturing in volume, which will make the BrainCard as cheap as possible.

We're beginning with 1000 chips that we already have in inventory which were originally ordered by an industrial client. After that, we will aim to start manufacturing on a mass production line, and this will take approximately six months. So, those first 1000 purchasers will be the only ones able to experience the unique capabilities of the BrainCard until mid-2015.

The first 1000 BrainCard's will cost $199 and are what we call IWIN (I Want It Now), or $219 for a version including an image sensor (the IS version) - so 500 of each version.

If we don't reach the goal, all the money raised will be aimed at manufacturing as many BrainCards as we can, so that it can be more affordable for the masses.

This is why we're turning to the maker community — we'd like to crowdsource our research and development through YOU!

The impact Neural networks should be everywhere by now, in your phone, in wearable technology. The NeuroMem technology is mature and the market needs exist. This project has the ability to propel neuromorphic technology into the mainstream consciousness by showing electronics manufacturers what can be done with it.

This is why we're turning to the maker community — we'd like to crowdsource our research and development through YOU!

Risks and challengesThe core of the NeuroMem/NeuromorThings team has been in place for 16 years and has plenty of research and industrial customers already using the CM1K chip, so this is not a typical “prototype” project.

We have a full supply chain already in place for both the board and for mounting the chips. We also have a wealth of knowledge in developing board-level and semiconductor technologies — all of which makes the risks to you a bare minimum.

We just need your support to complete prototyping/testing and to begin volume manufacturing. The first 1000 IWIN BrainCards will have exclusive access to the technology for the three months it takes us to make the new batch of chips.

Once we begin mass manufacturing the BrainCard, we will begin our long development roadmap on its successors and other neuromorthings.

After the first run of IWIN devices, the rest of the time will be dedicated to mounting the chips to the boards and testing them. With enough support we can get production runs up to very large numbers per month very quickly.

Shipping
Shipping a technology product is fraught with issues like export restrictions. We've tried to make it as simple as possible and built shipping as a perk.

In the US, Mexico and Canada? included

Rest of World? $30 Shipping & Packing

Due to the technical nature of the BrainCard it can be liable to Export Restrictions in certain countries under United States Law. If you are unsure if you are effected - please contact us at: info@neuromorthings.com and put "Export" in the subject line and we'll do everything we can to help.

Other Ways You Can HelpCan't buy a BrainCard? How about giving us a High $5? High 5'ers will all feature on the website and be written into NeuromorThings lore... it's a program for those interested in the technology and who want to help but who can't spring for their own BrainCard.

Got no cash at all? No problem - simply SPREAD THE WORD! Tell everyone you know about us and help us that way instead, on Facebook, on Twitter - wherever.

Every little bit helps!

Export regulations:
It might occurs, in certain rare cases that your country is under export embargo and we cannot ship because of the nature of the technology included in the BrainCard.If this exceptional situation occurs your money will be fully refunded.

Find This Campaign On

Facebook Twitter YouTube Website
 

Team

• 
  Matt Grimshaw
  

  VP Communications

  Verified 346 friends
• 
  Guy Paillet
  

  Co-Founder

  Verified
• 
  Anne Menendez
  

  Co-Founder

  Verified
• 
  Philippe Lambinet
  

  CEO

  Verified

Do you think this campaign contains prohibited content? Let us know.