neuralnetwork
Latest
Facebook's face-recognition tech is almost as good at Stallone-spotting as you are
Facebook's long been interested in facial recognition, as the photo tag-suggestion feature that didn't go down too well in Europe shows. The Zuck's social network also gobbled up a face-recognition outfit in 2012, but it's Facebook's AI research team that's made headway recently with technology that's almost as good as us meatsacks at identifying mugs. Known as DeepFace, the system uses a "nine-layer deep neural network" that's been taught to pick up on patterns by looking at over 4 million photos of more than 4,000 people. We're not as up-to-date with complex machine learning techniques as we should be, either, but the main reason DeepFace is so accurate is its method of "frontalization" -- or, creating a front-facing portrait from a more dynamic source image.
Qualcomm's brain-like processor is almost as smart as your dog
Biological brains are sometimes overrated, but they're still orders of magnitude quicker and more power-efficient than traditional computer chips. That's why Qualcomm has been quietly funneling some of its prodigious income into a project called "Zeroth," which it hopes will one day give it a radical advantage over other mobile chip companies. According to Qualcomm's Matt Grob, Zeroth is a "biologically-inspired" processor that is modeled on real-life neurons and is capable of learning from feedback in much the same way as a human or animal brain does. And unlike some other so-called artifical intelligences we've seen, this one appears to work. The video after the break shows a Zeroth-controlled robot exploring an environment and then naturally adjusting its behavior in response not to lines of code, but to someone telling it whether it's being "good" or "bad." "Everything here is biologically realistic: spiking neurons implemented in hardware on that actual machine." Zeroth is advanced enough that Qualcomm says it's ready to work with other companies who want to develop applications to run on it. One particular focus is on building neural networks that will fit into mobile devices and enable them to learn from users who, unlike coders, aren't able or willing to instruct devices in the usual tedious manner. Grob even claims that, when a general-purpose Zeroth neural network is trained to do something specific, such as recognizing and tracking an object, it can already accomplish that task better than an algorithm designed solely for that function. Check out the source link to see Grob's full talk and more demo videos -- especially if you want to confirm your long-held suspicion that dogs are scarily good at math.
Facebook developing brain-like AI to find deeper meaning in feeds and photos
Facebook's current News Feed ranking isn't all that clever -- it's good at surfacing popular updates, but it can miss lower-profile updates that are personally relevant. The company may soon raise the News Feed's IQ, however, as it recently launched an artificial intelligence research group. The new team hopes to use deep learning AI, which simulates a neural network, to determine which posts are genuinely important. The technology could also sort a user's photos, and it might even select the best shots. While the AI work has only just begun, the company tells MIT Technology Review that it should release some findings to the public; those breakthroughs in social networking could help society as a whole.
Robotic legs simulate our neural system, lurch along in the most human-like way so far
We've seen some pretty wonky bipedal robots before, but scientists at the University of Arizona have gone straight to the source -- us -- to make one with a more human-like saunter. It turns out it's not just our skull-borne computer that controls gait: a simple neural network in the lumber area of our spine, called the central pattern generator (CPG), also fires to provide the necessary rhythm. By creating a basic digital version of that and connecting some feedback sensors in the legs, a more natural human stride (without balance) was created -- and on top of that it didn't require the tricky processing used in other striding bots. Apparently this throws light on why babies can make that cute walking motion even before they toddle in earnest, since the necessary CPG system comes pre-installed from birth. That means the study could lead to new ways of stimulating that region to help those with spinal cord injuries re-learn to walk, and produce better, less complex walking robots to boot. Judging by the video, it's a good start, but there's still a ways to go before they can mimic us exactly -- you can watch it after the break.
DNA-based artificial neural network is a primitive brain in a test tube (video)
Many simpler forms of life on this planet, including some of our earliest ancestors, don't have proper brains. Instead they have networks of neurons that fire in response to stimuli, triggering reactions. Scientists from Caltech have actually figured out how to create such a primitive pre-brain using strands of DNA. Researchers, led by Lulu Qian, strung together DNA molecules to create bio-mechanical circuits. By sequencing the four bases of our genetic code in a particular way, they were able to program it to respond differently to various inputs. To prove their success the team quizzed the organic circuit, essentially playing 20 questions, feeding it clues to the identity of a particular scientist using more DNA strands. The artificial neural network nailed answer every time. Check out the PR and pair of videos that dig a little deeper into the experiment after the break.
British researchers design a million-chip neural network 1/100 as complex as your brain
If you want some idea of the complexity of the human brain, consider this: a group of British universities plans to link as many as a million ARM processors in order to simulate just a small fraction of it. The resulting model, called SpiNNaker (Spiking Neural Network architecture), will represent less than one percent of a human's gray matter, which contains 100 billion neurons. (Take that, mice brains!) Yet even this small scale representation, researchers believe, will yield insight into how the brain functions, perhaps enabling new treatments for cognitive disorders, similar to previous models that increased our understanding of schizophrenia. As these neural networks increase in complexity, they come closer to mimicking human brains -- perhaps even developing the ability to make their own Skynet references.
Schizophrenic computer may help us understand similarly afflicted humans
Although we usually prefer our computers to be perfect, logical, and psychologically fit, sometimes there's more to be learned from a schizophrenic one. A University of Texas experiment has doomed a computer with dementia praecox, saddling the silicon soul with symptoms that normally only afflict humans. By telling the machine's neural network to treat everything it learned as extremely important, the team hopes to aid clinical research in understanding the schizophrenic brain -- following a popular theory that suggests afflicted patients lose the ability to forget or ignore frivolous information, causing them to make illogical connections and paranoid jumps in reason. Sure enough, the machine lost it, and started spinning wild, delusional stories, eventually claiming responsibility for a terrorist attack. Yikes. We aren't hastening the robot apocalypse if we're programming machines to go mad intentionally, right?
