Tag Archives: who
#431836 Do Our Brains Use Deep Learning to Make ...
The first time Dr. Blake Richards heard about deep learning, he was convinced that he wasn’t just looking at a technique that would revolutionize artificial intelligence. He also knew he was looking at something fundamental about the human brain.
That was the early 2000s, and Richards was taking a course with Dr. Geoff Hinton at the University of Toronto. Hinton, a pioneer architect of the algorithm that would later take the world by storm, was offering an introductory course on his learning method inspired by the human brain.
The key words here are “inspired by.” Despite Richards’ conviction, the odds were stacked against him. The human brain, as it happens, seems to lack a critical function that’s programmed into deep learning algorithms. On the surface, the algorithms were violating basic biological facts already proven by neuroscientists.
But what if, superficial differences aside, deep learning and the brain are actually compatible?
Now, in a new study published in eLife, Richards, working with DeepMind, proposed a new algorithm based on the biological structure of neurons in the neocortex. Also known as the cortex, this outermost region of the brain is home to higher cognitive functions such as reasoning, prediction, and flexible thought.
The team networked their artificial neurons together into a multi-layered network and challenged it with a classic computer vision task—identifying hand-written numbers.
The new algorithm performed well. But the kicker is that it analyzed the learning examples in a way that’s characteristic of deep learning algorithms, even though it was completely based on the brain’s fundamental biology.
“Deep learning is possible in a biological framework,” concludes the team.
Because the model is only a computer simulation at this point, Richards hopes to pass the baton to experimental neuroscientists, who could actively test whether the algorithm operates in an actual brain.
If so, the data could then be passed back to computer scientists to work out the next generation of massively parallel and low-energy algorithms to power our machines.
It’s a first step towards merging the two fields back into a “virtuous circle” of discovery and innovation.
The blame game
While you’ve probably heard of deep learning’s recent wins against humans in the game of Go, you might not know the nitty-gritty behind the algorithm’s operations.
In a nutshell, deep learning relies on an artificial neural network with virtual “neurons.” Like a towering skyscraper, the network is structured into hierarchies: lower-level neurons process aspects of an input—for example, a horizontal or vertical stroke that eventually forms the number four—whereas higher-level neurons extract more abstract aspects of the number four.
To teach the network, you give it examples of what you’re looking for. The signal propagates forward in the network (like climbing up a building), where each neuron works to fish out something fundamental about the number four.
Like children trying to learn a skill the first time, initially the network doesn’t do so well. It spits out what it thinks a universal number four should look like—think a Picasso-esque rendition.
But here’s where the learning occurs: the algorithm compares the output with the ideal output, and computes the difference between the two (dubbed “error”). This error is then “backpropagated” throughout the entire network, telling each neuron: hey, this is how far off you were, so try adjusting your computation closer to the ideal.
Millions of examples and tweakings later, the network inches closer to the desired output and becomes highly proficient at the trained task.
This error signal is crucial for learning. Without efficient “backprop,” the network doesn’t know which of its neurons are off kilter. By assigning blame, the AI can better itself.
The brain does this too. How? We have no clue.
Biological No-Go
What’s clear, though, is that the deep learning solution doesn’t work.
Backprop is a pretty needy function. It requires a very specific infrastructure for it to work as expected.
For one, each neuron in the network has to receive the error feedback. But in the brain, neurons are only connected to a few downstream partners (if that). For backprop to work in the brain, early-level neurons need to be able to receive information from billions of connections in their downstream circuits—a biological impossibility.
And while certain deep learning algorithms adapt a more local form of backprop— essentially between neurons—it requires their connection forwards and backwards to be symmetric. This hardly ever occurs in the brain’s synapses.
More recent algorithms adapt a slightly different strategy, in that they implement a separate feedback pathway that helps the neurons to figure out errors locally. While it’s more biologically plausible, the brain doesn’t have a separate computational network dedicated to the blame game.
What it does have are neurons with intricate structures, unlike the uniform “balls” that are currently applied in deep learning.
Branching Networks
The team took inspiration from pyramidal cells that populate the human cortex.
“Most of these neurons are shaped like trees, with ‘roots’ deep in the brain and ‘branches’ close to the surface,” says Richards. “What’s interesting is that these roots receive a different set of inputs than the branches that are way up at the top of the tree.”
This is an illustration of a multi-compartment neural network model for deep learning. Left: Reconstruction of pyramidal neurons from mouse primary visual cortex. Right: Illustration of simplified pyramidal neuron models. Image Credit: CIFAR
Curiously, the structure of neurons often turn out be “just right” for efficiently cracking a computational problem. Take the processing of sensations: the bottoms of pyramidal neurons are right smack where they need to be to receive sensory input, whereas the tops are conveniently placed to transmit feedback errors.
Could this intricate structure be evolution’s solution to channeling the error signal?
The team set up a multi-layered neural network based on previous algorithms. But rather than having uniform neurons, they gave those in middle layers—sandwiched between the input and output—compartments, just like real neurons.
When trained with hand-written digits, the algorithm performed much better than a single-layered network, despite lacking a way to perform classical backprop. The cell-like structure itself was sufficient to assign error: the error signals at one end of the neuron are naturally kept separate from input at the other end.
Then, at the right moment, the neuron brings both sources of information together to find the best solution.
There’s some biological evidence for this: neuroscientists have long known that the neuron’s input branches perform local computations, which can be integrated with signals that propagate backwards from the so-called output branch.
However, we don’t yet know if this is the brain’s way of dealing blame—a question that Richards urges neuroscientists to test out.
What’s more, the network parsed the problem in a way eerily similar to traditional deep learning algorithms: it took advantage of its multi-layered structure to extract progressively more abstract “ideas” about each number.
“[This is] the hallmark of deep learning,” the authors explain.
The Deep Learning Brain
Without doubt, there will be more twists and turns to the story as computer scientists incorporate more biological details into AI algorithms.
One aspect that Richards and team are already eyeing is a top-down predictive function, in which signals from higher levels directly influence how lower levels respond to input.
Feedback from upper levels doesn’t just provide error signals; it could also be nudging lower processing neurons towards a “better” activity pattern in real-time, says Richards.
The network doesn’t yet outperform other non-biologically derived (but “brain-inspired”) deep networks. But that’s not the point.
“Deep learning has had a huge impact on AI, but, to date, its impact on neuroscience has been limited,” the authors say.
Now neuroscientists have a lead they could experimentally test: that the structure of neurons underlie nature’s own deep learning algorithm.
“What we might see in the next decade or so is a real virtuous cycle of research between neuroscience and AI, where neuroscience discoveries help us to develop new AI and AI can help us interpret and understand our experimental data in neuroscience,” says Richards.
Image Credit: christitzeimaging.com / Shutterstock.com Continue reading
#431689 Robotic Materials Will Distribute ...
The classical view of a robot as a mechanical body with a central “brain” that controls its behavior could soon be on its way out. The authors of a recent article in Science Robotics argue that future robots will have intelligence distributed throughout their bodies.
The concept, and the emerging discipline behind it, are variously referred to as “material robotics” or “robotic materials” and are essentially a synthesis of ideas from robotics and materials science. Proponents say advances in both fields are making it possible to create composite materials capable of combining sensing, actuation, computation, and communication and operating independently of a central processing unit.
Much of the inspiration for the field comes from nature, with practitioners pointing to the adaptive camouflage of the cuttlefish’s skin, the ability of bird wings to morph in response to different maneuvers, or the banyan tree’s ability to grow roots above ground to support new branches.
Adaptive camouflage and morphing wings have clear applications in the defense and aerospace sector, but the authors say similar principles could be used to create everything from smart tires able to calculate the traction needed for specific surfaces to grippers that can tailor their force to the kind of object they are grasping.
“Material robotics represents an acknowledgment that materials can absorb some of the challenges of acting and reacting to an uncertain world,” the authors write. “Embedding distributed sensors and actuators directly into the material of the robot’s body engages computational capabilities and offloads the rigid information and computational requirements from the central processing system.”
The idea of making materials more adaptive is not new, and there are already a host of “smart materials” that can respond to stimuli like heat, mechanical stress, or magnetic fields by doing things like producing a voltage or changing shape. These properties can be carefully tuned to create materials capable of a wide variety of functions such as movement, self-repair, or sensing.
The authors say synthesizing these kinds of smart materials, alongside other advanced materials like biocompatible conductors or biodegradable elastomers, is foundational to material robotics. But the approach also involves integration of many different capabilities in the same material, careful mechanical design to make the most of mechanical capabilities, and closing the loop between sensing and control within the materials themselves.
While there are stand-alone applications for such materials in the near term, like smart fabrics or robotic grippers, the long-term promise of the field is to distribute decision-making in future advanced robots. As they are imbued with ever more senses and capabilities, these machines will be required to shuttle huge amounts of control and feedback data to and fro, placing a strain on both their communication and computation abilities.
Materials that can process sensor data at the source and either autonomously react to it or filter the most relevant information to be passed on to the central processing unit could significantly ease this bottleneck. In a press release related to an earlier study, Nikolaus Correll, an assistant professor of computer science at the University of Colorado Boulder who is also an author of the current paper, pointed out this is a tactic used by the human body.
“The human sensory system automatically filters out things like the feeling of clothing rubbing on the skin,” he said. “An artificial skin with possibly thousands of sensors could do the same thing, and only report to a central ‘brain’ if it touches something new.”
There are still considerable challenges to realizing this vision, though, the authors say, noting that so far the young field has only produced proof of concepts. The biggest challenge remains manufacturing robotic materials in a way that combines all these capabilities in a small enough package at an affordable cost.
Luckily, the authors note, the field can draw on convergent advances in both materials science, such as the development of new bulk materials with inherent multifunctionality, and robotics, such as the ever tighter integration of components.
And they predict that doing away with the prevailing dichotomy of “brain versus body” could lay the foundations for the emergence of “robots with brains in their bodies—the foundation of inexpensive and ubiquitous robots that will step into the real world.”
Image Credit: Anatomy Insider / Shutterstock.com Continue reading
#431603 What We Can Learn From the Second Life ...
For every new piece of technology that gets developed, you can usually find people saying it will never be useful. The president of the Michigan Savings Bank in 1903, for example, said, “The horse is here to stay but the automobile is only a novelty—a fad.” It’s equally easy to find people raving about whichever new technology is at the peak of the Gartner Hype Cycle, which tracks the buzz around these newest developments and attempts to temper predictions. When technologies emerge, there are all kinds of uncertainties, from the actual capacity of the technology to its use cases in real life to the price tag.
Eventually the dust settles, and some technologies get widely adopted, to the extent that they can become “invisible”; people take them for granted. Others fall by the wayside as gimmicky fads or impractical ideas. Picking which horses to back is the difference between Silicon Valley millions and Betamax pub-quiz-question obscurity. For a while, it seemed that Google had—for once—backed the wrong horse.
Google Glass emerged from Google X, the ubiquitous tech giant’s much-hyped moonshot factory, where highly secretive researchers work on the sci-fi technologies of the future. Self-driving cars and artificial intelligence are the more mundane end for an organization that apparently once looked into jetpacks and teleportation.
The original smart glasses, Google began selling Google Glass in 2013 for $1,500 as prototypes for their acolytes, around 8,000 early adopters. Users could control the glasses with a touchpad, or, activated by tilting the head back, with voice commands. Audio relay—as with several wearable products—is via bone conduction, which transmits sound by vibrating the skull bones of the user. This was going to usher in the age of augmented reality, the next best thing to having a chip implanted directly into your brain.
On the surface, it seemed to be a reasonable proposition. People had dreamed about augmented reality for a long time—an onboard, JARVIS-style computer giving you extra information and instant access to communications without even having to touch a button. After smartphone ubiquity, it looked like a natural step forward.
Instead, there was a backlash. People may be willing to give their data up to corporations, but they’re less pleased with the idea that someone might be filming them in public. The worst aspect of smartphones is trying to talk to people who are distractedly scrolling through their phones. There’s a famous analogy in Revolutionary Road about an old couple’s loveless marriage: the husband tunes out his wife’s conversation by turning his hearing aid down to zero. To many, Google Glass seemed to provide us with a whole new way to ignore each other in favor of our Twitter feeds.
Then there’s the fact that, regardless of whether it’s because we’re not used to them, or if it’s a more permanent feature, people wearing AR tech often look very silly. Put all this together with a lack of early functionality, the high price (do you really feel comfortable wearing a $1,500 computer?), and a killer pun for the users—Glassholes—and the final recipe wasn’t great for Google.
Google Glass was quietly dropped from sale in 2015 with the ominous slogan posted on Google’s website “Thanks for exploring with us.” Reminding the Glass users that they had always been referred to as “explorers”—beta-testing a product, in many ways—it perhaps signaled less enthusiasm for wearables than the original, Google Glass skydive might have suggested.
In reality, Google went back to the drawing board. Not with the technology per se, although it has improved in the intervening years, but with the uses behind the technology.
Under what circumstances would you actually need a Google Glass? When would it genuinely be preferable to a smartphone that can do many of the same things and more? Beyond simply being a fashion item, which Google Glass decidedly was not, even the most tech-evangelical of us need a convincing reason to splash $1,500 on a wearable computer that’s less socially acceptable and less easy to use than the machine you’re probably reading this on right now.
Enter the Google Glass Enterprise Edition.
Piloted in factories during the years that Google Glass was dormant, and now roaring back to life and commercially available, the Google Glass relaunch got under way in earnest in July of 2017. The difference here was the specific audience: workers in factories who need hands-free computing because they need to use their hands at the same time.
In this niche application, wearable computers can become invaluable. A new employee can be trained with pre-programmed material that explains how to perform actions in real time, while instructions can be relayed straight into a worker’s eyeline without them needing to check a phone or switch to email.
Medical devices have long been a dream application for Google Glass. You can imagine a situation where people receive real-time information during surgery, or are augmented by artificial intelligence that provides additional diagnostic information or questions in response to a patient’s symptoms. The quest to develop a healthcare AI, which can provide recommendations in response to natural language queries, is on. The famously untidy doctor’s handwriting—and the associated death toll—could be avoided if the glasses could take dictation straight into a patient’s medical records. All of this is far more useful than allowing people to check Facebook hands-free while they’re riding the subway.
Google’s “Lens” application indicates another use for Google Glass that hadn’t quite matured when the original was launched: the Lens processes images and provides information about them. You can look at text and have it translated in real time, or look at a building or sign and receive additional information. Image processing, either through neural networks hooked up to a cloud database or some other means, is the frontier that enables driverless cars and similar technology to exist. Hook this up to a voice-activated assistant relaying information to the user, and you have your killer application: real-time annotation of the world around you. It’s this functionality that just wasn’t ready yet when Google launched Glass.
Amazon’s recent announcement that they want to integrate Alexa into a range of smart glasses indicates that the tech giants aren’t ready to give up on wearables yet. Perhaps, in time, people will become used to voice activation and interaction with their machines, at which point smart glasses with bone conduction will genuinely be more convenient than a smartphone.
But in many ways, the real lesson from the initial failure—and promising second life—of Google Glass is a simple question that developers of any smart technology, from the Internet of Things through to wearable computers, must answer. “What can this do that my smartphone can’t?” Find your answer, as the Enterprise Edition did, as Lens might, and you find your product.
Image Credit: Hattanas / Shutterstock.com Continue reading
#431592 Reactive Content Will Get to Know You ...
The best storytellers react to their audience. They look for smiles, signs of awe, or boredom; they simultaneously and skillfully read both the story and their sitters. Kevin Brooks, a seasoned storyteller working for Motorola’s Human Interface Labs, explains, “As the storyteller begins, they must tune in to… the audience’s energy. Based on this energy, the storyteller will adjust their timing, their posture, their characterizations, and sometimes even the events of the story. There is a dialog between audience and storyteller.”
Shortly after I read the script to Melita, the latest virtual reality experience from Madrid-based immersive storytelling company Future Lighthouse, CEO Nicolas Alcalá explained to me that the piece is an example of “reactive content,” a concept he’s been working on since his days at Singularity University.
For the first time in history, we have access to technology that can merge the reactive and affective elements of oral storytelling with the affordances of digital media, weaving stunning visuals, rich soundtracks, and complex meta-narratives in a story arena that has the capability to know you more intimately than any conventional storyteller could.
It’s no understatement to say that the storytelling potential here is phenomenal.
In short, we can refer to content as reactive if it reads and reacts to users based on their body rhythms, emotions, preferences, and data points. Artificial intelligence is used to analyze users’ behavior or preferences to sculpt unique storylines and narratives, essentially allowing for a story that changes in real time based on who you are and how you feel.
The development of reactive content will allow those working in the industry to go one step further than simply translating the essence of oral storytelling into VR. Rather than having a narrative experience with a digital storyteller who can read you, reactive content has the potential to create an experience with a storyteller who knows you.
This means being able to subtly insert minor personal details that have a specific meaning to the viewer. When we talk to our friends we often use experiences we’ve shared in the past or knowledge of our audience to give our story as much resonance as possible. Targeting personal memories and aspects of our lives is a highly effective way to elicit emotions and aid in visualizing narratives. When you can do this with the addition of visuals, music, and characters—all lifted from someone’s past—you have the potential for overwhelmingly engaging and emotionally-charged content.
Future Lighthouse inform me that for now, reactive content will rely primarily on biometric feedback technology such as breathing, heartbeat, and eye tracking sensors. A simple example would be a story in which parts of the environment or soundscape change in sync with the user’s heartbeat and breathing, or characters who call you out for not paying attention.
The next step would be characters and situations that react to the user’s emotions, wherein algorithms analyze biometric information to make inferences about states of emotional arousal (“why are you so nervous?” etc.). Another example would be implementing the use of “arousal parameters,” where the audience can choose what level of “fear” they want from a VR horror story before algorithms modulate the experience using information from biometric feedback devices.
The company’s long-term goal is to gather research on storytelling conventions and produce a catalogue of story “wireframes.” This entails distilling the basic formula to different genres so they can then be fleshed out with visuals, character traits, and soundtracks that are tailored for individual users based on their deep data, preferences, and biometric information.
The development of reactive content will go hand in hand with a renewed exploration of diverging, dynamic storylines, and multi-narratives, a concept that hasn’t had much impact in the movie world thus far. In theory, the idea of having a story that changes and mutates is captivating largely because of our love affair with serendipity and unpredictability, a cultural condition theorist Arthur Kroker refers to as the “hypertextual imagination.” This feeling of stepping into the unknown with the possibility of deviation from the habitual translates as a comforting reminder that our own lives can take exciting and unexpected turns at any moment.
The inception of the concept into mainstream culture dates to the classic Choose Your Own Adventure book series that launched in the late 70s, which in its literary form had great success. However, filmic takes on the theme have made somewhat less of an impression. DVDs like I’m Your Man (1998) and Switching (2003) both use scene selection tools to determine the direction of the storyline.
A more recent example comes from Kino Industries, who claim to have developed the technology to allow filmmakers to produce interactive films in which viewers can use smartphones to quickly vote on which direction the narrative takes at numerous decision points throughout the film.
The main problem with diverging narrative films has been the stop-start nature of the interactive element: when I’m immersed in a story I don’t want to have to pick up a controller or remote to select what’s going to happen next. Every time the audience is given the option to take a new path (“press this button”, “vote on X, Y, Z”) the narrative— and immersion within that narrative—is temporarily halted, and it takes the mind a while to get back into this state of immersion.
Reactive content has the potential to resolve these issues by enabling passive interactivity—that is, input and output without having to pause and actively make decisions or engage with the hardware. This will result in diverging, dynamic narratives that will unfold seamlessly while being dependent on and unique to the specific user and their emotions. Passive interactivity will also remove the game feel that can often be a symptom of interactive experiences and put a viewer somewhere in the middle: still firmly ensconced in an interactive dynamic narrative, but in a much subtler way.
While reading the Melita script I was particularly struck by a scene in which the characters start to engage with the user and there’s a synchronicity between the user’s heartbeat and objects in the virtual world. As the narrative unwinds and the words of Melita’s character get more profound, parts of the landscape, which seemed to be flashing and pulsating at random, come together and start to mimic the user’s heartbeat.
In 2013, Jane Aspell of Anglia Ruskin University (UK) and Lukas Heydrich of the Swiss Federal Institute of Technology proved that a user’s sense of presence and identification with a virtual avatar could be dramatically increased by syncing the on-screen character with the heartbeat of the user. The relationship between bio-digital synchronicity, immersion, and emotional engagement is something that will surely have revolutionary narrative and storytelling potential.
Image Credit: Tithi Luadthong / Shutterstock.com Continue reading