Tag Archives: inspired
#432181 Putting AI in Your Pocket: MIT Chip Cuts ...
Neural networks are powerful things, but they need a lot of juice. Engineers at MIT have now developed a new chip that cuts neural nets’ power consumption by up to 95 percent, potentially allowing them to run on battery-powered mobile devices.
Smartphones these days are getting truly smart, with ever more AI-powered services like digital assistants and real-time translation. But typically the neural nets crunching the data for these services are in the cloud, with data from smartphones ferried back and forth.
That’s not ideal, as it requires a lot of communication bandwidth and means potentially sensitive data is being transmitted and stored on servers outside the user’s control. But the huge amounts of energy needed to power the GPUs neural networks run on make it impractical to implement them in devices that run on limited battery power.
Engineers at MIT have now designed a chip that cuts that power consumption by up to 95 percent by dramatically reducing the need to shuttle data back and forth between a chip’s memory and processors.
Neural nets consist of thousands of interconnected artificial neurons arranged in layers. Each neuron receives input from multiple neurons in the layer below it, and if the combined input passes a certain threshold it then transmits an output to multiple neurons above it. The strength of the connection between neurons is governed by a weight, which is set during training.
This means that for every neuron, the chip has to retrieve the input data for a particular connection and the connection weight from memory, multiply them, store the result, and then repeat the process for every input. That requires a lot of data to be moved around, expending a lot of energy.
The new MIT chip does away with that, instead computing all the inputs in parallel within the memory using analog circuits. That significantly reduces the amount of data that needs to be shoved around and results in major energy savings.
The approach requires the weights of the connections to be binary rather than a range of values, but previous theoretical work had suggested this wouldn’t dramatically impact accuracy, and the researchers found the chip’s results were generally within two to three percent of the conventional non-binary neural net running on a standard computer.
This isn’t the first time researchers have created chips that carry out processing in memory to reduce the power consumption of neural nets, but it’s the first time the approach has been used to run powerful convolutional neural networks popular for image-based AI applications.
“The results show impressive specifications for the energy-efficient implementation of convolution operations with memory arrays,” Dario Gil, vice president of artificial intelligence at IBM, said in a statement.
“It certainly will open the possibility to employ more complex convolutional neural networks for image and video classifications in IoT [the internet of things] in the future.”
It’s not just research groups working on this, though. The desire to get AI smarts into devices like smartphones, household appliances, and all kinds of IoT devices is driving the who’s who of Silicon Valley to pile into low-power AI chips.
Apple has already integrated its Neural Engine into the iPhone X to power things like its facial recognition technology, and Amazon is rumored to be developing its own custom AI chips for the next generation of its Echo digital assistant.
The big chip companies are also increasingly pivoting towards supporting advanced capabilities like machine learning, which has forced them to make their devices ever more energy-efficient. Earlier this year ARM unveiled two new chips: the Arm Machine Learning processor, aimed at general AI tasks from translation to facial recognition, and the Arm Object Detection processor for detecting things like faces in images.
Qualcomm’s latest mobile chip, the Snapdragon 845, features a GPU and is heavily focused on AI. The company has also released the Snapdragon 820E, which is aimed at drones, robots, and industrial devices.
Going a step further, IBM and Intel are developing neuromorphic chips whose architectures are inspired by the human brain and its incredible energy efficiency. That could theoretically allow IBM’s TrueNorth and Intel’s Loihi to run powerful machine learning on a fraction of the power of conventional chips, though they are both still highly experimental at this stage.
Getting these chips to run neural nets as powerful as those found in cloud services without burning through batteries too quickly will be a big challenge. But at the current pace of innovation, it doesn’t look like it will be too long before you’ll be packing some serious AI power in your pocket.
Image Credit: Blue Planet Studio / Shutterstock.com Continue reading
Posted in Human Robots
#431999 Brain-Like Chips Now Beat the Human ...
Move over, deep learning. Neuromorphic computing—the next big thing in artificial intelligence—is on fire.
Just last week, two studies individually unveiled computer chips modeled after information processing in the human brain.
The first, published in Nature Materials, found a perfect solution to deal with unpredictability at synapses—the gap between two neurons that transmit and store information. The second, published in Science Advances, further amped up the system’s computational power, filling synapses with nanoclusters of supermagnetic material to bolster information encoding.
The result? Brain-like hardware systems that compute faster—and more efficiently—than the human brain.
“Ultimately we want a chip as big as a fingernail to replace one big supercomputer,” said Dr. Jeehwan Kim, who led the first study at MIT in Cambridge, Massachusetts.
Experts are hopeful.
“The field’s full of hype, and it’s nice to see quality work presented in an objective way,” said Dr. Carver Mead, an engineer at the California Institute of Technology in Pasadena not involved in the work.
Software to Hardware
The human brain is the ultimate computational wizard. With roughly 100 billion neurons densely packed into the size of a small football, the brain can deftly handle complex computation at lightning speed using very little energy.
AI experts have taken note. The past few years saw brain-inspired algorithms that can identify faces, falsify voices, and play a variety of games at—and often above—human capability.
But software is only part of the equation. Our current computers, with their transistors and binary digital systems, aren’t equipped to run these powerful algorithms.
That’s where neuromorphic computing comes in. The idea is simple: fabricate a computer chip that mimics the brain at the hardware level. Here, data is both processed and stored within the chip in an analog manner. Each artificial synapse can accumulate and integrate small bits of information from multiple sources and fire only when it reaches a threshold—much like its biological counterpart.
Experts believe the speed and efficiency gains will be enormous.
For one, the chips will no longer have to transfer data between the central processing unit (CPU) and storage blocks, which wastes both time and energy. For another, like biological neural networks, neuromorphic devices can support neurons that run millions of streams of parallel computation.
A “Brain-on-a-chip”
Optimism aside, reproducing the biological synapse in hardware form hasn’t been as easy as anticipated.
Neuromorphic chips exist in many forms, but often look like a nanoscale metal sandwich. The “bread” pieces are generally made of conductive plates surrounding a switching medium—a conductive material of sorts that acts like the gap in a biological synapse.
When a voltage is applied, as in the case of data input, ions move within the switching medium, which then creates conductive streams to stimulate the downstream plate. This change in conductivity mimics the way biological neurons change their “weight,” or the strength of connectivity between two adjacent neurons.
But so far, neuromorphic synapses have been rather unpredictable. According to Kim, that’s because the switching medium is often comprised of material that can’t channel ions to exact locations on the downstream plate.
“Once you apply some voltage to represent some data with your artificial neuron, you have to erase and be able to write it again in the exact same way,” explains Kim. “But in an amorphous solid, when you write again, the ions go in different directions because there are lots of defects.”
In his new study, Kim and colleagues swapped the jelly-like switching medium for silicon, a material with only a single line of defects that acts like a channel to guide ions.
The chip starts with a thin wafer of silicon etched with a honeycomb-like pattern. On top is a layer of silicon germanium—something often present in transistors—in the same pattern. This creates a funnel-like dislocation, a kind of Grand Canal that perfectly shuttles ions across the artificial synapse.
The researchers then made a neuromorphic chip containing these synapses and shot an electrical zap through them. Incredibly, the synapses’ response varied by only four percent—much higher than any neuromorphic device made with an amorphous switching medium.
In a computer simulation, the team built a multi-layer artificial neural network using parameters measured from their device. After tens of thousands of training examples, their neural network correctly recognized samples 95 percent of the time, just 2 percent lower than state-of-the-art software algorithms.
The upside? The neuromorphic chip requires much less space than the hardware that runs deep learning algorithms. Forget supercomputers—these chips could one day run complex computations right on our handheld devices.
A Magnetic Boost
Meanwhile, in Boulder, Colorado, Dr. Michael Schneider at the National Institute of Standards and Technology also realized that the standard switching medium had to go.
“There must be a better way to do this, because nature has figured out a better way to do this,” he says.
His solution? Nanoclusters of magnetic manganese.
Schneider’s chip contained two slices of superconducting electrodes made out of niobium, which channel electricity with no resistance. When researchers applied different magnetic fields to the synapse, they could control the alignment of the manganese “filling.”
The switch gave the chip a double boost. For one, by aligning the switching medium, the team could predict the ion flow and boost uniformity. For another, the magnetic manganese itself adds computational power. The chip can now encode data in both the level of electrical input and the direction of the magnetisms without bulking up the synapse.
It seriously worked. At one billion times per second, the chips fired several orders of magnitude faster than human neurons. Plus, the chips required just one ten-thousandth of the energy used by their biological counterparts, all the while synthesizing input from nine different sources in an analog manner.
The Road Ahead
These studies show that we may be nearing a benchmark where artificial synapses match—or even outperform—their human inspiration.
But to Dr. Steven Furber, an expert in neuromorphic computing, we still have a ways before the chips go mainstream.
Many of the special materials used in these chips require specific temperatures, he says. Magnetic manganese chips, for example, require temperatures around absolute zero to operate, meaning they come with the need for giant cooling tanks filled with liquid helium—obviously not practical for everyday use.
Another is scalability. Millions of synapses are necessary before a neuromorphic device can be used to tackle everyday problems such as facial recognition. So far, no deal.
But these problems may in fact be a driving force for the entire field. Intense competition could push teams into exploring different ideas and solutions to similar problems, much like these two studies.
If so, future chips may come in diverse flavors. Similar to our vast array of deep learning algorithms and operating systems, the computer chips of the future may also vary depending on specific requirements and needs.
It is worth developing as many different technological approaches as possible, says Furber, especially as neuroscientists increasingly understand what makes our biological synapses—the ultimate inspiration—so amazingly efficient.
Image Credit: arakio / Shutterstock.com Continue reading
Posted in Human Robots
#431958 The Next Generation of Cameras Might See ...
You might be really pleased with the camera technology in your latest smartphone, which can recognize your face and take slow-mo video in ultra-high definition. But these technological feats are just the start of a larger revolution that is underway.
The latest camera research is shifting away from increasing the number of mega-pixels towards fusing camera data with computational processing. By that, we don’t mean the Photoshop style of processing where effects and filters are added to a picture, but rather a radical new approach where the incoming data may not actually look like at an image at all. It only becomes an image after a series of computational steps that often involve complex mathematics and modeling how light travels through the scene or the camera.
This additional layer of computational processing magically frees us from the chains of conventional imaging techniques. One day we may not even need cameras in the conventional sense any more. Instead we will use light detectors that only a few years ago we would never have considered any use for imaging. And they will be able to do incredible things, like see through fog, inside the human body and even behind walls.
Single Pixel Cameras
One extreme example is the single pixel camera, which relies on a beautifully simple principle. Typical cameras use lots of pixels (tiny sensor elements) to capture a scene that is likely illuminated by a single light source. But you can also do things the other way around, capturing information from many light sources with a single pixel.
To do this you need a controlled light source, for example a simple data projector that illuminates the scene one spot at a time or with a series of different patterns. For each illumination spot or pattern, you then measure the amount of light reflected and add everything together to create the final image.
Clearly the disadvantage of taking a photo in this is way is that you have to send out lots of illumination spots or patterns in order to produce one image (which would take just one snapshot with a regular camera). But this form of imaging would allow you to create otherwise impossible cameras, for example that work at wavelengths of light beyond the visible spectrum, where good detectors cannot be made into cameras.
These cameras could be used to take photos through fog or thick falling snow. Or they could mimic the eyes of some animals and automatically increase an image’s resolution (the amount of detail it captures) depending on what’s in the scene.
It is even possible to capture images from light particles that have never even interacted with the object we want to photograph. This would take advantage of the idea of “quantum entanglement,” that two particles can be connected in a way that means whatever happens to one happens to the other, even if they are a long distance apart. This has intriguing possibilities for looking at objects whose properties might change when lit up, such as the eye. For example, does a retina look the same when in darkness as in light?
Multi-Sensor Imaging
Single-pixel imaging is just one of the simplest innovations in upcoming camera technology and relies, on the face of it, on the traditional concept of what forms a picture. But we are currently witnessing a surge of interest for systems that use lots of information but traditional techniques only collect a small part of it.
This is where we could use multi-sensor approaches that involve many different detectors pointed at the same scene. The Hubble telescope was a pioneering example of this, producing pictures made from combinations of many different images taken at different wavelengths. But now you can buy commercial versions of this kind of technology, such as the Lytro camera that collects information about light intensity and direction on the same sensor, to produce images that can be refocused after the image has been taken.
The next generation camera will probably look something like the Light L16 camera, which features ground-breaking technology based on more than ten different sensors. Their data are combined using a computer to provide a 50 MB, re-focusable and re-zoomable, professional-quality image. The camera itself looks like a very exciting Picasso interpretation of a crazy cell-phone camera.
Yet these are just the first steps towards a new generation of cameras that will change the way in which we think of and take images. Researchers are also working hard on the problem of seeing through fog, seeing behind walls, and even imaging deep inside the human body and brain.
All of these techniques rely on combining images with models that explain how light travels through through or around different substances.
Another interesting approach that is gaining ground relies on artificial intelligence to “learn” to recognize objects from the data. These techniques are inspired by learning processes in the human brain and are likely to play a major role in future imaging systems.
Single photon and quantum imaging technologies are also maturing to the point that they can take pictures with incredibly low light levels and videos with incredibly fast speeds reaching a trillion frames per second. This is enough to even capture images of light itself traveling across as scene.
Some of these applications might require a little time to fully develop, but we now know that the underlying physics should allow us to solve these and other problems through a clever combination of new technology and computational ingenuity.
This article was originally published on The Conversation. Read the original article.
Image Credit: Sylvia Adams / Shutterstock.com Continue reading
Posted in Human Robots