Tag Archives: simple

#435601 New Double 3 Robot Makes Telepresence ...

Posted on September 12, 2019 by Android

Today, Double Robotics is announcing Double 3, the latest major upgrade to its line of consumer(ish) telepresence robots. We had a (mostly) fantastic time testing out Double 2 back in 2016. One of the things that we found out back then was that it takes a lot of practice to remotely drive the robot around. Double 3 solves this problem by leveraging the substantial advances in 3D sensing and computing that have taken place over the past few years, giving their new robot a level of intelligence that promises to make telepresence more accessible for everyone.

Double 2’s iPad has been replaced by “a fully integrated solution”—which is a fancy way of saying a dedicated 9.7-inch touchscreen and a whole bunch of other stuff. That other stuff includes an NVIDIA Jetson TX2 AI computing module, a beamforming six-microphone array, an 8-watt speaker, a pair of 13-megapixel cameras (wide angle and zoom) on a tilting mount, five ultrasonic rangefinders, and most excitingly, a pair of Intel RealSense D430 depth sensors.

It’s those new depth sensors that really make Double 3 special. The D430 modules each uses a pair of stereo cameras with a pattern projector to generate 1280 x 720 depth data with a range of between 0.2 and 10 meters away. The Double 3 robot uses all of this high quality depth data to locate obstacles, but at this point, it still doesn’t drive completely autonomously. Instead, it presents the remote operator with a slick, augmented reality view of drivable areas in the form of a grid of dots. You just click where you want the robot to go, and it will skillfully take itself there while avoiding obstacles (including dynamic obstacles) and related mishaps along the way.

This effectively offloads the most stressful part of telepresence—not running into stuff—from the remote user to the robot itself, which is the way it should be. That makes it that much easier to encourage people to utilize telepresence for the first time. The way the system is implemented through augmented reality is particularly impressive, I think. It looks like it’s intuitive enough for an inexperienced user without being restrictive, and is a clever way of mitigating even significant amounts of lag.

Otherwise, Double 3’s mobility system is exactly the same as the one featured on Double 2. In fact, that you can stick a Double 3 head on a Double 2 body and it instantly becomes a Double 3. Double Robotics is thoughtfully offering this to current Double 2 owners as a significantly more affordable upgrade option than buying a whole new robot.

For more details on all of Double 3's new features, we spoke with the co-founders of Double Robotics, Marc DeVidts and David Cann.

IEEE Spectrum: Why use this augmented reality system instead of just letting the user click on a regular camera image? Why make things more visually complicated, especially for new users?

Marc DeVidts and David Cann: One of the things that we realized about nine months ago when we got this whole thing working was that without the mixed reality for driving, it was really too magical of an experience for the customer. Even us—we had a hard time understanding whether the robot could really see obstacles and understand where the floor is and that kind of thing. So, we said “What would be the best way of communicating this information to the user?” And the right way to do it ended up drawing the graphics directly onto the scene. It’s really awesome—we have a full, real time 3D scene with the depth information drawn on top of it. We’re starting with some relatively simple graphics, and we’ll be adding more graphics in the future to help the user understand what the robot is seeing.

How robust is the vision system when it comes to obstacle detection and avoidance? Does it work with featureless surfaces, IR absorbent surfaces, in low light, in direct sunlight, etc?

We’ve looked at all of those cases, and one of the reasons that we’re going with the RealSense is the projector that helps us to see blank walls. We also found that having two sensors—one facing the floor and one facing forward—gives us a great coverage area. Having ultrasonic sensors in there as well helps us to detect anything that we can't see with the cameras. They're sort of a last safety measure, especially useful for detecting glass.

It seems like there’s a lot more that you could do with this sensing and mapping capability. What else are you working on?

We're starting with this semi-autonomous driving variant, and we're doing a private beta of full mapping. So, we’re going to do full SLAM of your environment that will be mapped by multiple robots at the same time while you're driving, and then you'll be able to zoom out to a map and click anywhere and it will drive there. That's where we're going with it, but we want to take baby steps to get there. It's the obvious next step, I think, and there are a lot more possibilities there.

Do you expect developers to be excited for this new mapping capability?

We're using a very powerful computer in the robot, a NVIDIA Jetson TX2 running Ubuntu. There's room to grow. It’s actually really exciting to be able to see, in real time, the 3D pose of the robot along with all of the depth data that gets transformed in real time into one view that gives you a full map. Having all of that data and just putting those pieces together and getting everything to work has been a huge feat in of itself.

We have an extensive API for developers to do custom implementations, either for telepresence or other kinds of robotics research. Our system isn't running ROS, but we're going to be adding ROS adapters for all of our hardware components.

Telepresence robots depend heavily on wireless connectivity, which is usually not something that telepresence robotics companies like Double have direct control over. Have you found that connectivity has been getting significantly better since you first introduced Double?

When we started in 2013, we had a lot of customers that didn’t have WiFi in their hallways, just in the conference rooms. We very rarely hear about customers having WiFi connectivity issues these days. The bigger issue we see is when people are calling into the robot from home, where they don't have proper traffic management on their home network. The robot doesn't need a ton of bandwidth, but it does need consistent, low latency bandwidth. And so, if someone else in the house is watching Netflix or something like that, it’s going to saturate your connection. But for the most part, it’s gotten a lot better over the last few years, and it’s no longer a big problem for us.

Do you think 5G will make a significant difference to telepresence robots?

We’ll see. We like the low latency possibilities and the better bandwidth, but it's all going to be a matter of what kind of reception you get. LTE can be great, if you have good reception; it’s all about where the tower is. I’m pretty sure that WiFi is going to be the primary thing for at least the next few years.

DeVidts also mentioned that an unfortunate side effect of the new depth sensors is that hanging a t-shirt on your Double to give it some personality will likely render it partially blind, so that's just something to keep in mind. To make up for this, you can switch around the colorful trim surrounding the screen, which is nowhere near as fun.

When the Double 3 is ready for shipping in late September, US $2,000 will get you the new head with all the sensors and stuff, which seamlessly integrates with your Double 2 base. Buying Double 3 straight up (with the included charging dock) will run you $4,ooo. This is by no means an inexpensive robot, and my impression is that it’s not really designed for individual consumers. But for commercial, corporate, healthcare, or education applications, $4k for a robot as capable as the Double 3 is really quite a good deal—especially considering the kinds of use cases for which it’s ideal.

[ Double Robotics ] Continue reading →

Posted in Human Robots

#435593 AI at the Speed of Light

Posted on September 11, 2019 by Android

Neural networks shine for solving tough problems such as facial and voice recognition, but conventional electronic versions are limited in speed and hungry for power. In theory, optics could beat digital electronic computers in the matrix calculations used in neural networks. However, optics had been limited by their inability to do some complex calculations that had required electronics. Now new experiments show that all-optical neural networks can tackle those problems.

The key attraction of neural networks is their massive interconnections among processors, comparable to the complex interconnections among neurons in the brain. This lets them perform many operations simultaneously, like the human brain does when looking at faces or listening to speech, making them more efficient for facial and voice recognition than traditional electronic computers that execute one instruction at a time.

Today's electronic neural networks have reached eight million neurons, but their future use in artificial intelligence may be limited by their high power usage and limited parallelism in connections. Optical connections through lenses are inherently parallel. The lens in your eye simultaneously focuses light from across your field of view onto the retina in the back of your eye, where an array of light-detecting nerve cells detects the light. Each cell then relays the signal it receives to neurons in the brain that process the visual signals to show us an image.

Glass lenses process optical signals by focusing light, which performs a complex mathematical operation called a Fourier transform that preserves the information in the original scene but rearranges is completely. One use of Fourier transforms is converting time variations in signal intensity into a plot of the frequencies present in the signal. The military used this trick in the 1950s to convert raw radar return signals recorded by an aircraft in flight into a three-dimensional image of the landscape viewed by the plane. Today that conversion is done electronically, but the vacuum-tube computers of the 1950s were not up to the task.

Development of neural networks for artificial intelligence started with electronics, but their AI applications have been limited by their slow processing and need for extensive computing resources. Some researchers have developed hybrid neural networks, in which optics perform simple linear operations, but electronics perform more complex nonlinear calculations. Now two groups have demonstrated simple all-optical neural networks that do all processing with light.

In May, Wolfram Pernice of the Institute of Physics at the University of Münster in Germany and colleagues reported testing an all-optical “neuron” in which signals change target materials between liquid and solid states, an effect that has been used for optical data storage. They demonstrated nonlinear processing, and produced output pulses like those from organic neurons. They then produced an integrated photonic circuit that incorporated four optical neurons operating at different wavelengths, each of which connected to 15 optical synapses. The photonic circuit contained more than 140 components and could recognize simple optical patterns. The group wrote that their device is scalable, and that the technology promises “access to the high speed and high bandwidth inherent to optical systems, thus enabling the direct processing of optical telecommunication and visual data.”

Now a group at the Hong Kong University of Science and Technology reports in Optica that they have made an all-optical neural network based on a different process, electromagnetically induced transparency, in which incident light affects how atoms shift between quantum-mechanical energy levels. The process is nonlinear and can be triggered by very weak light signals, says Shengwang Du, a physics professor and coauthor of the paper.

In their demonstration, they illuminated rubidium-85 atoms cooled by lasers to about 10 microKelvin (10 microdegrees above absolute zero). Although the technique may seem unusually complex, Du said the system was the most accessible one in the lab that could produce the desired effects. “As a pure quantum atomic system [it] is ideal for this proof-of-principle experiment,” he says.

Next, they plan to scale up the demonstration using a hot atomic vapor center, which is less expensive, does not require time-consuming preparation of cold atoms, and can be integrated with photonic chips. Du says the major challenges are reducing cost of the nonlinear processing medium and increasing the scale of the all-optical neural network for more complex tasks.

“Their demonstration seems valid,” says Volker Sorger, an electrical engineer at George Washington University in Washington who was not involved in either demonstration. He says the all-optical approach is attractive because it offers very high parallelism, but the update rate is limited to about 100 hertz because of the liquid crystals used in their test, and he is not completely convinced their approach can be scaled error-free. Continue reading →

Posted in Human Robots

#435589 Construction Robots Learn to Excavate by ...

Posted on September 11, 2019 by Android

Pavel Savkin remembers the first time he watched a robot imitate his movements. Minutes earlier, the engineer had finished “showing” the robotic excavator its new goal by directing its movements manually. Now, running on software Savkin helped design, the robot was reproducing his movements, gesture for gesture. “It was like there was something alive in there—but I knew it was me,” he said.

Savkin is the CTO of SE4, a robotics software project that styles itself the “driver” of a fleet of robots that will eventually build human colonies in space. For now, SE4 is focused on creating software that can help developers communicate with robots, rather than on building hardware of its own.
The Tokyo-based startup showed off an industrial arm from Universal Robots that was running SE4’s proprietary software at SIGGRAPH in July. SE4’s demonstration at the Los Angeles innovation conference drew the company’s largest audience yet. The robot, nicknamed Squeezie, stacked real blocks as directed by SE4 research engineer Nathan Quinn, who wore a VR headset and used handheld controls to “show” Squeezie what to do.

As Quinn manipulated blocks in a virtual 3D space, the software learned a set of ordered instructions to be carried out in the real world. That order is essential for remote operations, says Quinn. To build remotely, developers need a way to communicate instructions to robotic builders on location. In the age of digital construction and industrial robotics, giving a computer a blueprint for what to build is a well-explored art. But operating on a distant object—especially under conditions that humans haven’t experienced themselves—presents challenges that only real-time communication with operators can solve.

The problem is that, in an unpredictable setting, even simple tasks require not only instruction from an operator, but constant feedback from the changing environment. Five years ago, the Swedish fiber network provider umea.net (part of the private Umeå Energy utility) took advantage of the virtual reality boom to promote its high-speed connections with the help of a viral video titled “Living with Lag: An Oculus Rift Experiment.” The video is still circulated in VR and gaming circles.

In the experiment, volunteers donned headgear that replaced their real-time biological senses of sight and sound with camera and audio feeds of their surroundings—both set at a 3-second delay. Thus equipped, volunteers attempt to complete everyday tasks like playing ping-pong, dancing, cooking, and walking on a beach, with decidedly slapstick results.

At outer-orbit intervals, including SE4’s dream of construction projects on Mars, the limiting factor in communication speed is not an artificial delay, but the laws of physics. The shifting relative positions of Earth and Mars mean that communications between the planets—even at the speed of light—can take anywhere from 3 to 22 minutes.

A long-distance relationship

Imagine trying to manage a construction project from across an ocean without the benefit of intelligent workers: sending a ship to an unknown world with a construction crew and blueprints for a log cabin, and four months later receiving a letter back asking how to cut down a tree. The parallel problem in long-distance construction with robots, according to SE4 CEO Lochlainn Wilson, is that automation relies on predictability. “Every robot in an industrial setting today is expecting a controlled environment.”
Platforms for applying AR and VR systems to teach tasks to artificial intelligences, as SE4 does, are already proliferating in manufacturing, healthcare, and defense. But all of the related communications systems are bound by physics and, specifically, the speed of light.
The same fundamental limitation applies in space. “Our communications are light-based, whether they’re radio or optical,” says Laura Seward Forczyk, a planetary scientist and consultant for space startups. “If you’re going to Mars and you want to communicate with your robot or spacecraft there, you need to have it act semi- or mostly-independently so that it can operate without commands from Earth.”

Semantic control
That’s exactly what SE4 aims to do. By teaching robots to group micro-movements into logical units—like all the steps to building a tower of blocks—the Tokyo-based startup lets robots make simple relational judgments that would allow them to receive a full set of instruction modules at once and carry them out in order. This sidesteps the latency issue in real-time bilateral communications that could hamstring a project or at least make progress excruciatingly slow.
The key to the platform, says Wilson, is the team’s proprietary operating software, “Semantic Control.” Just as in linguistics and philosophy, “semantics” refers to meaning itself, and meaning is the key to a robot’s ability to make even the smallest decisions on its own. “A robot can scan its environment and give [raw data] to us, but it can’t necessarily identify the objects around it and what they mean,” says Wilson.

That’s where human intelligence comes in. As part of the demonstration phase, the human operator of an SE4-controlled machine “annotates” each object in the robot’s vicinity with meaning. By labeling objects in the VR space with useful information—like which objects are building material and which are rocks—the operator helps the robot make sense of its real 3D environment before the building begins.

Giving robots the tools to deal with a changing environment is an important step toward allowing the AI to be truly independent, but it’s only an initial step. “We’re not letting it do absolutely everything,” said Quinn. “Our robot is good at moving an object from point A to point B, but it doesn’t know the overall plan.” Wilson adds that delegating environmental awareness and raw mechanical power to separate agents is the optimal relationship for a mixed human-robot construction team; it “lets humans do what they’re good at, while robots do what they do best.”

This story was updated on 4 September 2019. Continue reading →

Posted in Human Robots

#435522 Harvard’s Smart Exo-Shorts Talk to the ...

Posted on August 20, 2019 by Android

Exosuits don’t generally scream “fashionable” or “svelte.” Take the mind-controlled robotic exoskeleton that allowed a paraplegic man to kick off the World Cup back in 2014. Is it cool? Hell yeah. Is it practical? Not so much.

Yapping about wearability might seem childish when the technology already helps people with impaired mobility move around dexterously. But the lesson of the ill-fated Google Glassholes, which includes an awkward dorky head tilt and an assuming voice command, clearly shows that wearable computer assistants can’t just work technologically—they have to look natural and allow the user to behave like as usual. They have to, in a sense, disappear.

To Dr. Jose Pons at the Legs + Walking Ability Lab in Chicago, exosuits need three main selling points to make it in the real world. One, they have to physically interact with their wearer and seamlessly deliver assistance when needed. Two, they should cognitively interact with the host to guide and control the robot at all times. Finally, they need to feel like a second skin—move with the user without adding too much extra mass or reducing mobility.

This week, a US-Korean collaboration delivered the whole shebang in a Lululemon-style skin-hugging package combined with a retro waist pack. The portable exosuit, weighing only 11 pounds, looks like a pair of spandex shorts but can support the wearer’s hip movement when needed. Unlike their predecessors, the shorts are embedded with sensors that let them know when the wearer is walking versus running by analyzing gait.

Switching between the two movement modes may not seem like much, but what naturally comes to our brains doesn’t translate directly to smart exosuits. “Walking and running have fundamentally different biomechanics, which makes developing devices that assist both gaits challenging,” the team said. Their algorithm, computed in the cloud, allows the wearer to easily switch between both, with the shorts providing appropriate hip support that makes the movement experience seamless.

To Pons, who was not involved in the research but wrote a perspective piece, the study is an exciting step towards future exosuits that will eventually disappear under the skin—that is, implanted neural interfaces to control robotic assistance or activate the user’s own muscles.

“It is realistic to think that we will witness, in the next several years…robust human-robot interfaces to command wearable robotics based on…the neural code of movement in humans,” he said.

A “Smart” Exosuit Hack
There are a few ways you can hack a human body to move with an exosuit. One is using implanted electrodes inside the brain or muscles to decipher movement intent. With heavy practice, a neural implant can help paralyzed people walk again or dexterously move external robotic arms. But because the technique requires surgery, it’s not an immediate sell for people who experience low mobility because of aging or low muscle tone.

The other approach is to look to biophysics. Rather than decoding neural signals that control movement, here the idea is to measure gait and other physical positions in space to decipher intent. As you can probably guess, accurately deciphering user intent isn’t easy, especially when the wearable tries to accommodate multiple gaits. But the gains are many: there’s no surgery involved, and the wearable is low in energy consumption.

Double Trouble
The authors decided to tackle an everyday situation. You’re walking to catch the train to work, realize you’re late, and immediately start sprinting.

That seemingly easy conversion hides a complex switch in biomechanics. When you walk, your legs act like an inverted pendulum that swing towards a dedicated center in a predictable way. When you run, however, the legs move more like a spring-loaded system, and the joints involved in the motion differ from a casual stroll. Engineering an assistive wearable for each is relatively simple; making one for both is exceedingly hard.

Led by Dr. Conor Walsh at Harvard University, the team started with an intuitive idea: assisted walking and running requires specialized “actuation” profiles tailored to both. When the user is moving in a way that doesn’t require assistance, the wearable needs to be out of the way so that it doesn’t restrict mobility. A quick analysis found that assisting hip extension has the largest impact, because it’s important to both gaits and doesn’t add mass to the lower legs.

Building on that insight, the team made a waist belt connected to two thigh wraps, similar to a climbing harness. Two electrical motors embedded inside the device connect the waist belt to other components through a pulley system to help the hip joints move. The whole contraption weighed about 11 lbs and didn’t obstruct natural movement.

Next, the team programmed two separate supporting profiles for walking and running. The goal was to reduce the “metabolic cost” for both movements, so that the wearer expends as little energy as needed. To switch between the two programs, they used a cloud-based classification algorithm to measure changes in energy fluctuation to figure out what mode—running or walking—the user is in.

Smart Booster
Initial trials on treadmills were highly positive. Six male volunteers with similar age and build donned the exosuit and either ran or walked on the treadmill at varying inclines. The algorithm performed perfectly at distinguishing between the two gaits in all conditions, even at steep angles.

An outdoor test with eight volunteers also proved the algorithm nearly perfect. Even on uneven terrain, only two steps out of all test trials were misclassified. In an additional trial on mud or snow, the algorithm performed just as well.

“The system allows the wearer to use their preferred gait for each speed,” the team said.

Software excellence translated to performance. A test found that the exosuit reduced the energy for walking by over nine percent and running by four percent. It may not sound like much, but the range of improvement is meaningful in athletic performance. Putting things into perspective, the team said, the metabolic rate reduction during walking is similar to taking 16 pounds off at the waist.

The Wearable Exosuit Revolution
The study’s lightweight exoshorts are hardly the only players in town. Back in 2017, SRI International’s spin-off, Superflex, engineered an Aura suit to support mobility in the elderly. The Aura used a different mechanism: rather than a pulley system, it incorporated a type of smart material that contracts in a manner similar to human muscles when zapped with electricity.

Embedded with a myriad of sensors for motion, accelerometers and gyroscopes, Aura’s smartness came from mini-computers that measure how fast the wearer is moving and track the user’s posture. The data were integrated and processed locally inside hexagon-shaped computing pods near the thighs and upper back. The pods also acted as the control center for sending electrical zaps to give the wearer a boost when needed.

Around the same time, a collaboration between Harvard’s Wyss Institute and ReWalk Robotics introduced a fabric-based wearable robot to assist a wearer’s legs for balance and movement. Meanwhile, a Swiss team coated normal fabric with electroactive material to weave soft, pliable artificial “muscles” that move with the skin.

Although health support is the current goal, the military is obviously interested in similar technologies to enhance soldiers’ physicality. Superflex’s Aura, for example, was originally inspired by technology born from DARPA’s Warrior Web Program, which aimed to reduce a soldier’s mechanical load.

That said, military gear has had a long history of trickling down to consumer use. Similar to the way camouflage, cargo pants, and GORE-TEX trickled down into the consumer ecosphere, it’s not hard to imagine your local Target eventually stocking intelligent exowear.

Image and Video Credit: Wyss Institute at Harvard University. Continue reading →

Posted in Human Robots

#435474 Watch China’s New Hybrid AI Chip Power ...

Posted on August 7, 2019 by Android

When I lived in Beijing back in the 90s, a man walking his bike was nothing to look at. But today, I did a serious double-take at a video of a bike walking his man.

No kidding.

The bike itself looks overloaded but otherwise completely normal. Underneath its simplicity, however, is a hybrid computer chip that combines brain-inspired circuits with machine learning processes into a computing behemoth. Thanks to its smart chip, the bike self-balances as it gingerly rolls down a paved track before smoothly gaining speed into a jogging pace while navigating dexterously around obstacles. It can even respond to simple voice commands such as “speed up,” “left,” or “straight.”

Far from a circus trick, the bike is a real-world demo of the AI community’s latest attempt at fashioning specialized hardware to keep up with the challenges of machine learning algorithms. The Tianjic (天机*) chip isn’t just your standard neuromorphic chip. Rather, it has the architecture of a brain-like chip, but can also run deep learning algorithms—a match made in heaven that basically mashes together neuro-inspired hardware and software.

The study shows that China is readily nipping at the heels of Google, Facebook, NVIDIA, and other tech behemoths investing in developing new AI chip designs—hell, with billions in government investment it may have already had a head start. A sweeping AI plan from 2017 looks to catch up with the US on AI technology and application by 2020. By 2030, China’s aiming to be the global leader—and a champion for building general AI that matches humans in intellectual competence.

The country’s ambition is reflected in the team’s parting words.

“Our study is expected to stimulate AGI [artificial general intelligence] development by paving the way to more generalized hardware platforms,” said the authors, led by Dr. Luping Shi at Tsinghua University.

A Hardware Conundrum
Shi’s autonomous bike isn’t the first robotic two-wheeler. Back in 2015, the famed research nonprofit SRI International in Menlo Park, California teamed up with Yamaha to engineer MOTOBOT, a humanoid robot capable of driving a motorcycle. Powered by state-of-the-art robotic hardware and machine learning, MOTOBOT eventually raced MotoGPTM world champion Valentino Rossi in a nail-biting match-off.

However, the technological core of MOTOBOT and Shi’s bike vastly differ, and that difference reflects two pathways towards more powerful AI. One, exemplified by MOTOBOT, is software—developing brain-like algorithms with increasingly efficient architecture, efficacy, and speed. That sounds great, but deep neural nets demand so many computational resources that general-purpose chips can’t keep up.

As Shi told China Science Daily: “CPUs and other chips are driven by miniaturization technologies based on physics. Transistors might shrink to nanoscale-level in 10, 20 years. But what then?” As more transistors are squeezed onto these chips, efficient cooling becomes a limiting factor in computational speed. Tax them too much, and they melt.

For AI processes to continue, we need better hardware. An increasingly popular idea is to build neuromorphic chips, which resemble the brain from the ground up. IBM’s TrueNorth, for example, contains a massively parallel architecture nothing like the traditional Von Neumann structure of classic CPUs and GPUs. Similar to biological brains, TrueNorth’s memory is stored within “synapses” between physical “neurons” etched onto the chip, which dramatically cuts down on energy consumption.

But even these chips are limited. Because computation is tethered to hardware architecture, most chips resemble just one specific type of brain-inspired network called spiking neural networks (SNNs). Without doubt, neuromorphic chips are highly efficient setups with dynamics similar to biological networks. They also don’t play nicely with deep learning and other software-based AI.

Brain-AI Hybrid Core
Shi’s new Tianjic chip brought the two incompatibilities together onto a single piece of brainy hardware.

First was to bridge the deep learning and SNN divide. The two have very different computation philosophies and memory organizations, the team said. The biggest difference, however, is that artificial neural networks transform multidimensional data—image pixels, for example—into a single, continuous, multi-bit 0 and 1 stream. In contrast, neurons in SNNs activate using something called “binary spikes” that code for specific activation events in time.

Confused? Yeah, it’s hard to wrap my head around it too. That’s because SNNs act very similarly to our neural networks and nothing like computers. A particular neuron needs to generate an electrical signal (a “spike”) large enough to transfer down to the next one; little blips in signals don’t count. The way they transmit data also heavily depends on how they’re connected, or the network topology. The takeaway: SNNs work pretty differently than deep learning.

Shi’s team first recreated this firing quirk in the language of computers—0s and 1s—so that the coding mechanism would become compatible with deep learning algorithms. They then carefully aligned the step-by-step building blocks of the two models, which allowed them to tease out similarities into a common ground to further build on. “On the basis of this unified abstraction, we built a cross-paradigm neuron scheme,” they said.

In general, the design allowed both computational approaches to share the synapses, where neurons connect and store data, and the dendrites, the outgoing branches of the neurons. In contrast, the neuron body, where signals integrate, was left reconfigurable for each type of computation, as were the input branches. Each building block was combined into a single unified functional core (FCore), which acts like a deep learning/SNN converter depending on its specific setup. Translation: the chip can do both types of previously incompatible computation.

The Chip
Using nanoscale fabrication, the team arranged 156 FCores, containing roughly 40,000 neurons and 10 million synapses, onto a chip less than a fifth of an inch in length and width. Initial tests showcased the chip’s versatility, in that it can run both SNNs and deep learning algorithms such as the popular convolutional neural network (CNNs) often used in machine vision.

Compared to IBM TrueNorth, the density of Tianjic’s cores increased by 20 percent, speeding up performance ten times and increasing bandwidth at least 100-fold, the team said. When pitted against GPUs, the current hardware darling of machine learning, the chip increased processing throughput up to 100 times, while using just a sliver (1/10,000) of energy.

Although these stats are great, real-life performance is even better as a demo. Here’s where the authors gave their Tianjic brain a body. The team combined one chip with multiple specialized networks to process vision, balance, voice commands, and decision-making in real time. Object detection and target tracking, for example, relied on a deep neural net CNN, whereas voice commands and balance data were recognized using an SNN. The inputs were then integrated inside a neural state machine, which churned out decisions to downstream output modules—for example, controlling the handle bar to turn left.

Thanks to the chip’s brain-like architecture and bilingual ability, Tianjic “allowed all of the neural network models to operate in parallel and realized seamless communication across the models,” the team said. The result is an autonomous bike that rolls after its human, balances across speed bumps, avoids crashing into roadblocks, and answers to voice commands.

General AI?
“It’s a wonderful demonstration and quite impressive,” said the editorial team at Nature, which published the study on its cover last week.

However, they cautioned, when comparing Tianjic with state-of-the-art chips designed for a single problem toe-to-toe on that particular problem, Tianjic falls behind. But building these jack-of-all-trades hybrid chips is definitely worth the effort. Compared to today’s limited AI, what people really want is artificial general intelligence, which will require new architectures that aren’t designed to solve one particular problem.

Until people start to explore, innovate, and play around with different designs, it’s not clear how we can further progress in the pursuit of general AI. A self-driving bike might not be much to look at, but its hybrid brain is a pretty neat place to start.

*The name, in Chinese, means “heavenly machine,” “unknowable mystery of nature,” or “confidentiality.” Go figure.

Image Credit: Alexander Ryabintsev / Shutterstock.com Continue reading →

Posted in Human Robots