DAPRA released a video of the LS3 Alpha Dog as it conquers a woody hillside. More intensive testing is to begin this summer.

First there were porters, then beasts of burden, then mechanized vehicles. And now we’re back to beasts of burden – and I mean beast! Yesterday Boston Dymanics took their Alpha Dog for a walk in the woods. After putting the LS3 prototype through the rigors of warehouse testing, like trying to kick it over with all their might, they decided it was ready for a real terrain test.

The Army considers physical overburden one of its top five science and technology challenges. Soldiers in the field can end up carrying gear weighing 100 lbs and the physical strain takes its toll. The LS3 (Legged Squad Support System) Alpha Dog is DARPA’s solution to lightening the solder’s load. But more than just navigating rugged terrain, yesterday’s demonstration tested the robot’s ability to follow a person using its “eyes,” a set of sensors that enable it to distinguish between trees or other obstructions and humans. You can actually see in the video that it follows the leader pretty well. With the strength to shoulder up to 400 lbs, the Alpha Dog will definitely be a soldier’s best friend. It also serves as a mobile power source that troops can use to recharge batteries for radios and other devices. It is still quite loud, however, so sneaking up on the enemy with Alpha Dog trailing probably wouldn’t be very smart.

This summer, DARPA and Boston Dynamics will begin an 18 month test period during which they will sharpen Alpha Dog’s capabilities. Its vision will be fine-tuned to follow individuals or designated objects and its ability to autonomously navigate over and around terrain obstacles will be evaluated. One planned test will see if Alpha Dog can complete a 20-mile trek while loaded with 400 lbs and without refueling. In addition to improving its vision, hearing technology will be added so that it can respond to commands like “stop,” “sit” or “come here.” At the end of testing, the robot will be given a chance to work with real Marines as they carry out field exercises. After watching the video, I’m pretty confident Alpha Dog will be ready.

[image credits: DARPAtv via YouTube and ieee spectrum]
[video credit: DARPAtv via YouTube]
image 1: Alpha Dog
image 2: Alpha Dog
video: Alpha Dog

The Pandora of movies. Films by Eve Sussman and Rufus Corporation are clips pieced together by a computer algorithm.

Indie movie makers can be a strange bunch, pushing the envelope of their craft and often losing us along the way. In any case, if you’re going to produce something unintelligible anyway, why not let a computer do it? Eve Sussmam and the Rufus Corporation did just that. She and lead actor Jeff Wood traveled to the Kazakhstan border of the Caspian Sea for two years of filming. But instead of a movie with a beginning, middle and end, they shot 3,000 individual and unrelated clips. To the clips they added 80 voice-overs and 150 pieces of music, mixed it all together and put it in a computer. A program on her Mac G5 tower, known at Rufus as the “serendipity machine,” then splices the bits together to create a final product.

As you might imagine, the resultant film doesn’t always make sense. But that’s part of the fun! As the Rufus Corporation writes on their website, “The unexpected juxtapositions create a sense of suspense alluding to a story that the viewer composes.”

It’s a clever experiment even if some viewers end up wanting to gouge their eyes out after a sitting. And there is some method to their madness. The film, titled “whiteonwhite:algorithnoir,” is centered on a geophysicist named Holz (played by Wood) who’s stuck in a gloomy, 1970’s-looking city operated by the New Method Oil Well Cementing Company. Distinct scenes such as wire tapped conversations or a job interview for Mr. Holz are (hopefully) woven together by distinct voiceovers and dialogues. When the scenes and audio are entered into the computer they’re tagged with keywords. The program then pieces them together in a way similar to Pandora’s stringing together of like music. If a clip is tagged “white,” the computer will randomly select from tens of other clips also having the “white” tag. The final product is intended to be a kind of “dystopian futuropolis.” What that means, however, changes with each viewing as no two runs are the same.

Watching the following trailer, I actually got a sense…um, I think…of a story.

Rufus Corporation says the movie was “inspired by Suprematist quests for transcendence, pure space and artistic higher ground.” I have no idea what that means but I hope they’ve achieved it. Beautiful things can happen when computers create art. And it’s only a matter of time before people attempt the same sort of thing with novel writing. Just watching the trailer, it’s hard to tell if the movie’s any good or not. I missed the showings at the Sundance Film Festival, but even so, they probably didn’t resemble the trailer anyway. And that’s okay, because that’s the whole point.

[image credits: Rufus Corporation and PRI via YouTube]
[video credit: PRI via YouTube]
image 1: whiteonwhite
image 2: Rufus
video: whiteonwhite

The upright, folded hands is a sign of confidence. So is driving a BMW. With ConnectedDrive Connect, BMW owners can do both!

While robotic cars have a ways to go yet before rolling (themselves) out onto showroom floors, BMW is incorporating driver assistance features into its cars that drivers can use – and they can sell – sooner rather than later. Their semi-autonomous driving system, ConnectedDrive Connect, was announced last year. The car maker has finished its closed track test runs and has released a video of the car out on the Autobahn for some good old fashioned hands off driving fun.

ConnectedDrive Connect includes four types of sensors – radar, camera, laser scanners and ultrasound distance sensors – that allow it to track cars in front of it up to a distance of 50 meters. It also detects cars in adjacent lanes. A driving simulator produces driving strategies on the go. For example, the car slows if it’s moving to fast into a turn, and it taps the brakes to maintain control going downhill. Like any cruise control the driver sets the speed and the maintains a safe distance behind a car in front of it. The system can be enabled between speeds of 30 and 180 km/h (81 mph). I’d be a little worried if that truck in the next lane splashes some mud on the car that it doesn’t go pell mell into the car in front of me. The sensors are “largely resistant to dirt build-up” according to BMW, but I’d wipe down regularly just to be safe. BMW owners probably don’t need to be reminded of that anyway.

But it’s really no fun to just simply maintain a safe distance, not with a Beamer. If the car in front of you is going too slow, ConnectedDrive Connect senses your impatience, changes lanes, and leaves the slowpoke in the dust.

To date the BMW 5 Series has logged 5,000 kilometers in “highly-automated” mode on freeways. When can we expect to see BMW drivers begin using both hands to apply their makeup? The company says it’s a few years yet before CDC goes into production. One of the things they’ll have to improve is the GPS tracking. Right now they only take the car out on roads that they’ve mapped to within centimeter accuracy.

It is becoming increasingly likely that your next car will have some kind of driver assistance. Last year carmakers spent over $10 billion in advanced driver assistance systems. By 2016 that number could increase to $130 billion, according to an ABI Research projection. The rapid spread will be due largely to their incorporation into more mainstream cars rather than being an option for luxury cars. As the technology is improved production costs are decreasing, making it increasingly feasible to add driver assistance to less expensive cars. Volkswagon’s Temporary Auto Pilot (TAP) is similar to ConnectDrive Connect in that it maintains a safe distance behind cars and keeps the car from veering out of the lane. And both Volkswagon and BMW get the job done without the ungainly periscope-looking sensor on Google’s Prius.

[image credits: World Car Fans]
[video credits: motorsixty via YouTube]
image: ConnectDrive Connect
video: ConnectDrive Connect

Udacity gives students videos that they can watch at their own pace and be continually quizzed on.

This past August fellow Singularity Hub writer Aaron Saenz wrote about Udacity, the online university created by Stanford artificial intelligence professor and Google autonomous vehicle leader, Sebastian Thrun. At the time Thrun was gearing up to teach his Introduction to Artificial Intelligence course to a class of 200 at Stanford. But why teach 200 when you can teach 1,000…or 160,000? With Udacity, Thrun and fellow AI giant Peter Norvig created an online version of the course, and anyone that wanted to enroll could – for free. The homework assignments and exams would be the same as the ones given to the Stanford students, and they would be graded in the same way so online enrollees could see how they stacked up to some of the brightest students in the world. It was to be a grand experiment in education.

Now, the semester’s over. The exams have been taken, the homework’s been turned in, computers logged off and pencils set down. How’d it all turn out? Thrun spoke recently at the Digital Life Design conference about he and Norvig’s experience. As you’ll see, his students weren’t the only ones with much to learn.

Online, the course went viral. Over 100,000 people enrolled in the initial weeks. By the time the lessons began Thrun and Norvig were instructors for a class size of 160,000. With students all over the world, they enlisted the help of some 2,000 volunteer translators to translate the classes into 44 different languages. Discussion groups were set up on social networks like Facebook so students could help each other, forming what Thrun called an “entire counterculture.”

Thrun also proudly pointed out that he was teaching more students than all the students of Stanford.

The lessons themselves were very simple – at least in method if not in content. Material was explained by Thrun and Norvig as they drew on sheets of paper. Kind of like the overhead projector lessons before the days of Powerpoint, except the online students could interact with the drawings. Rather than simply lecturing to the student and asking them to regurgitate the information on exams, the online format allowed for constant quizzing. Students would be asked a question then answer it by clicking or entering values right on the drawings. They wanted the student to actively think, be constantly challenged and given constant feedback.

The flexibility that this format offers is immediately clear. If the student misses a point or doesn’t quite understand, he or she can rewind, watch it again. Get the quiz wrong, just take it again…and again if you have to.

Until you get it right.

Udacity's artificial intelligence course attracted 160,000 people from all over the world.

When the course began, however, it wasn’t like that. Initially Thrun had structured it as he had structured every other course over the past twenty years of teaching. Give the kids really hard material, then it’s sink or swim. But then he received an email from a parent who called his class a “weeder” class, and told him his daughter was dropping out. It was an epiphany for Thrun, compelling him to make a bold claim: “Grades are the failure of the education system.”

Thrun’s sudden dislike of grades is with its all-or-nothing nature. If we get a “C” on an exam we obviously haven’t mastered the material. Yet even if we get a “C”, the professor moves on to more advanced material anyway that will likely depend on the previous, unmastered material. After the email,
Thrun completely revamped Udacity to break the mold. If a student is having trouble with a problem they continue to work on it until they get it right. To Thrun, that’s still worth an A+. Imagine that, an entire class of students who can test at an A+ level. He sums up the attitude by paraphrasing a point made by Salman Khan, founder of the online Khan Academy: “When you learn to ride a bicycle, and you fail to learn a bicycle, you don’t stop to learn a bicycle, give the person a ‘D’ and move onto unicycle.”

What will education look like in the future? If other educators buy into the Udacity model it would be a sea change in the approach to education. An email Thrun received from a student in Afghanistan shows just how radically it is already changing.

I spent the last few days under incoming mortar and rocket attacks, then dodging checkpoints under questionable legal status to exfiltrate a war zone to a third world air field until things settled down. I had about an hour of fairly solid internet connectivity to be able to get the assignments done, and still managed a respectable score. This is a typical week here for me.

Okay, it’s time to address the note-taking, 800 ton gorilla in the room. Don’t we always hear that the key to a better education is to make classes smaller? How can two people possibly teach a class of 160,000 students? Obviously Thrun and Norvig didn’t grade the homework and exams by hand. What kinds of pioneering AI professors would they be if they didn’t employ their subject matter to correct it? AI programs shouldered the grading, and even handled question submissions. The multitudinous questions submitted by the students are sifted by a program and the most common ones are plucked out to be addressed. Not only does this make effectively answering questions possible, but it will highlight confusion points where the curriculum could be fine-tuned.

Will online education bring about the end of 'class as usual?'

One might logically question the format, wonder if learning by yourself on a laptop turns education into a lonely and impersonal experience. But anyone who’s watched a movie on a long plane ride knows the blissful isolation of plugging in ear phones and watching the latest action/romance/comedy/family movies. If done thoughtfully, educators can take advantage of the intimacy of just you and your laptop and create an amazingly personal learning experience despite being one of 200,000 enrolled. Here, Udacity seems to have succeeded. With its illustrations literally drawn out for the students, the lessons made one student feel that Thrun and Norvig were “personally tutoring” her. Anyone who’s sat in an auditorium, looking over the heads of 200 hundred other students at the far away professor and Powerpoint projections knows that the impersonal feel of an average university classroom has much room for improvement. Incidentally, two weeks after Thrun’s AI class began at Stanford the class attendance had dwindled down from about 200 to about 30. They preferred him online rather than in person.

Udacity is now offering up two new courses, CS 101: Building a Search Engine and CS 373: Programming a Robotic Car. Together with David Evans, a professor at the University of Virginia, Thrun will teach you how to build a search engine in just seven weeks. The Search Engine course doesn’t require any programming experience. The Robotic Car course is more advanced, but don’t be scared off. Thrun says that familiarity with linear algebra and statistics and programming experience is useful, but none of this is required.

If you have 22 minutes to spare, I think you’ll enjoy Thrun’s talk. If you don’t, just go to 15:45 and listen as the moving student testimonials come in from all over the world. They alone should convince you that Udacity is on to something great. It was made clear that Thrun thinks so, when he shocked the audience by announcing that he was leaving his tenured position at Stanford. “I feel like there’s a red pill and a blue pill,” he told them. “You can take the blue pill and go back to Stanford…but I’ve taken the red pill and I’ve seen Wonderland.”

[image credits: Udacity and Xbxg32000 via WikiCommons]
[video credits: DLDconference via YouTube]
image 1 and 2: Udacity
image 3: lecture hall
video: DLDconference

The gangly Agile Justin, ready to toss a ball to its robotic twin Rollin' Justin.

Where was Agile Justin last year when we needed him to throw out the first pitch at a Philadelphia Phillies game? The PhillieBot was booed by Phillies fans after bouncing the ball to home plate. It would have been a different story had DLR’s latest robot been there.

Last summer DLR showed us Rollin’ Justin’s amazing ability to catch. Now they’ve created a robot that can toss the ball to Rollin’. Just as Rollin’ Justin was a great test platform for robotics technologies behind high-speed perception, catching strategy, dexterity and body control, Justin’s Agile twin presents DLR programmers with the challenge of effective ball tossing – something that PhillieBot failed miserably at.

As Hizook reports, DLR started with Rolling Justin and added “1.5 faster arms through different gear ratios; completely new wheel electronics and bus architecture, which allows a 500Hz control loop over all four wheels and steering [degrees of freedom] on the mobile platform; 1kHz control loop for the arms, torso and hand [degrees of freedom].”

Watch the ball toss in the video below. Obviously Agile Justin throws like a robot, kind of sidearm/underhand, not much like a major league pitcher. The coordination between arm, torso, and wheels gives new meaning to the term “pitching mechanics.”

[image credits: hizook via YouTube and DLR]
[video credits: hizook via YouTube]
image 1: throw
image 2: Agile Justin
video: Agile Justin

The ants simulation is a fairly straightforward strategy game where each ant from a hill is depicted as a single brightly colored pixel. Ants use their programmer’s code to tell them how to explore the space around them. Gather brown food pellets (also just a single pixel) to generate new ants, attack rival ants of different colors, and seal off enemy mounds to gain victory points. The simulation ends when all but one team is destroyed, when the strength of each team causes a stalemate, or when time is up. The ant hill that has the most points at the end of the game wins. Here’s “Andrew” from the AI challenge to give you a look at a typical game with three competing ant hills. Imagine this scenario played out tens of thousands of times with different AIs controlling the ants each time and you get a good idea of what goes on at the Google AI Challenge Fall 2011 competition:

As an AI challenge, ant wrangling is pretty straight forward. You need to teach your ants how to find a path based on what they can ‘see’. You have to give them preferences for finding/retrieving food. You also have to teach them how and when to go in for the kill. Master those tasks and you’ll perform really well in the AI challenge, but that’s easier said than done. Of the 13k+ matches simulated, almost half (~48%) end because ant teams can’t manage to find food.

The reason for the high rate of failure is also one of the AI Challenge’s most interesting goals: to draw in a lot of competitors. With starter packages ready to download, almost anyone can register and get a team started with just a few minutes of work. Really, just minutes. There’s a wide variety of programming languages supported (C++, Python, Java, etc) and the AI Challenge team is willing to work with you if you prefer a language they haven’t developed yet. Even people relatively new to programming can enter the competition and try their hand at insect domination. As everything is handled online, there’s no geographic boundaries either. Teams from all over the world can submit code, see how they do, and improve their ideas as they learn from their matches.

Of course, casting the fishing net wide means you get a lot of old tires in your catch. Current results show that about 15% of games end with a lone survivor, about 15% with a stabilized war (no hill able to beat another), and about 20% reach a time limit. Those are the three outcomes that really show AI routines competing against one another. The remaining 50% of games end basically due to fundamental programming errors – ants can’t find food or they don’t attack rival hills. Even with half the games ending on error, the AI Challenge already has more than 13,000 games played. That’s in less than a week of competition! And because this is all run in simulation, the challenge website can run 4500 games a day, so we can expect a lot more meaningful match ups. By the end of this ongoing series of online ant battles we’re probably going to see a low level AI that’s REALLY good at commanding an ant hill.

After the simulation is over, every game can be reviewed step by step so that programmers can see what their ants (and competitors) were doing. Click the image to jump to the 'visualizer' and watch a game unfold at your command.

So what? Well, whatever artificial intelligence developed by the AI challenge is going to be fairly rudimentary relatively speaking. We’re not going to see anything to rival the narrow AI applications we’ve reviewed in the past nor anything in even the same ecosystem as the artificial general intelligences (AGIs) most of us think about when we talk about smart computers. But that doesn’t matter. The reason Google sponsors the AI Challenge is that competitions like this help teach the rising new generation of programmers how to think creatively and innovate their code as they learn. Ideally, they’ll build their code so it adapts during each game. That sort of education is what makes the AI Challenge worthwhile, and it’s an investment that will pay dividends to Google (and every other computer based company out there) in the years ahead.

It’s important to remember that this challenge is just one of many around the world that helps grow the size and skills of the programmer community. We’ve discussed the ongoing Ms. Pac-Man wars to see who can build the best AI ghosts or orb-gobbling heroine. You can also include robot competitions with heavy problem solving elements like Rubik’s Cube racing and Micromouse maze runs. Each of these challenges provide great opportunities to have fun and develop job skills at the same time, and they’re increasing in number and scale every year. The AI Challenge started off as a much smaller competition at the University of Waterloo. Now its open to everyone with funding from the biggest name on the internet. As it and other challenges continue to grow, they will attract many more into pursuing careers in computer science and engineering. Who knows, somewhere among these ant hills may be the programmer that eventually creates the artificial intelligences we’ve been dreaming about for centuries.

Hopefully he or she won’t include the seek and destroy skills they honed competing for Google’s attention.

[image credits: AI Challenge]
[video credit: AI Challenge via ComputerWiz222]
[source: AI Challenge]

Nao Energy STation – or NEST - allows your Nao robot to take care of its energy needs on its own.

Nao, the miniature humanoid robot favorite, has added yet another capability to its growing list of awesomeness. It hasn’t gotten better at recognizing objects and faces. Its language comprehension skills haven’t been refined. Nor has it learned any new dance moves or card games. The latest addition to Nao’s bag of tricks is more about maintenance than marvel. Aldebaran, the makers of Nao, have created NEST – Nao Energy STation – that the robot uses to charge itself. No longer will members of the Nao Developer Program keep having to check to see if the robot’s still got juice. When its battery runs low, it simply walks over to its charging station, backs up to a post, plugs in just below its neck and takes a seat while it powers up. It uses its image tracking technology to orient itself to markings on a mat attached to the post. And because Nao is so smart, if the position of the markings is shifted somehow, say on an irregular surface, it adjusts and remembers for the next time. When it’s feeling like a new robot it stands up and literally whacks the plug off its back. Check the demonstration in the following video.

If we’re going to have truly autonomous robots in the future, being able to recharge themselves will be a necessity. Now, with NeST, Nao Developers won’t find themselves out on the town with a dead robot because they forgot to charge it before they left the house. If only cell phones were so gifted.

[video credits: NaoForge via YouTube]
video: NeST

"Extra oil and vinegar please!"

Willow Garage’s gifted PR2 robot just got even smarter. Developers from the University of Tokyo and the Technical University of Munchen have collaborated to give it the ability to find an object it can’t see or isn’t even sure is there. In a way similar to humans, the PR2 can now go after an object by reasoning where it is most likely to be.

If you’re at someone else’s house and you need a bottle opener you’d probably begin by looking in the kitchen drawers as opposed to, say, the bathroom (candles, maybe, but not the bottle opener). By using “semantic search” the PR2 calculates the probability of finding the object in a number of places. If it doesn’t find it in the first place, it continues its search in other likely places.

In the following video, University of Tokyo researchers send the PR2 out for a sandwich. It looks in the refrigerator, doesn’t see one, then takes the elevator to the ground floor and orders one at Subway!

This sort of common sense is not so common among robots (and many humans too, but that’s beyond the scope of this article). Key to semantic search is knowing where sandwiches, etc are likely to be in the first place. For this reason, the researchers introduced probability maps that indicated the refrigerator and Subway are high probability sandwich locations. The maps can be updated say, if Subway closes, and like seemingly everything else modifications can be made from an iPad.

Semantic search is the latest creation from TUM’s Intelligent Autonomous Systems. The PR2 robot platform uses Robot Operating System-based software that TUM has already contributed to significantly. Semantic search can be combined with other code packets found in the treasure trove of ROS so the PR2 can cook you breakfast, fetch you a beer, or clean up after you in a smarter way. In a video we showed last year, the PR2 depended on a human to decide which cups and bottles on the table were to be cleared away for washing and which ones should stay. By increasing the specifications of semantic search the robot’s selectivity can be made finer so that, for instance, any cups that are half-full should remain.

If robots are going to work with us and for us they’re going to need the kind of common sense that semantic search gives them. If we have to tell the robot what room, what refrigerator, what shelf and what can of beer it needs to fetch we’re apt to do it ourselves.

And that’s just no fun.

[video credits: GerbilGod7 via YouTube]
video: sandwich

The writing's on the wall...or in the fine print of the world's news media. A new software claims that the recent Arab Spring revolutions could have been predicted by tracking global news "tone."

A new type of software has been shown to predict revolutions by mining news reports around the world. Retrospectively mining the news for the past 30 years the software indicates points at which the likelihood for a revolution is high. When put to the test – bingo! – the software showed spikes just before the recent Egyptian and Libyan upheavals. It was also able to sift through world news to retrospectively pinpoint Osama Bin Ladin’s location to within 200 km. In the emerging science of ‘culturomics’ that tracks cultural trends through the written word, the software was the first to demonstrate that news coverage can be used to predict future events.

The software sifted through news reports from nearly every country in the world. Major sources were from the global news databases such the US government-run Open Source Center which provides foreign open source intelligence, Britain’s equivalent BBC Monitoring, as well as the New York Times’ archive that dates back to 1945. In all, the body of data included over 100 million news articles. The story elements were woven together into a mind-boggling web of 100 trillion relationships. To crunch the massive amount of data the SGI Altix supercomputer Nautilus was enlisted. Its 1024 Intel Nehalem cores give it a total processing power of 8.2 teraflops (trillions of floating point operations per second).

The strategy for pulling out relevant information included two main techniques. ‘Sentiment mining’ involves counting the number of words in a document that are categorized as positive, such as “good” or “nice,” or negative, such as “terrible” or “horrific.” Changes in the tone of a regions’ news documents over time correlates with the sentiment of the people in that region. A major dip in tone, as when the frequency of negative terms rapidly increases, means the natives are getting restless. An unprecedented dip could spell revolution. The second technique, called ‘full-text geocoding,’ matches sentiment to geographic locations.

The dip in global tone concerning Egypt that preceded Mubarak's ouster had only been seen twice in the past 30 years.

The software was developed by Kalev Leetaru at the University of Illinois’ Institute for Computing in the Humanities, Arts and Social Science. Looking specifically at news about Egypt, Libya and Tunisia, Leetaru’s software showed compelling trends of negative tone in the decade leading up to the recent revolutions we’ve come to call the Arab Spring.

In Egypt, for instance, widespread protests that began on January 25, 2011 led to President Mubarak’s resignation on February 11. Tone monitoring was performed on 52,438 articles worldwide between January 1979 and March 2011 that contained any mention of an Egyptian city. The software selected for Egyptian cities rather than the word “Egypt” to filter out articles that only casually mentioned Egypt the way a travel guide might do. Between January 1 and January 24 of 2011, global tone about Egypt dropped to an extent that had only been seen twice in the past 30 years. The previous times were January 1991 when US planes bombarded Iraqi troops in Kuwait and the March 2003 US invasion of Iraq. According to Leetaru, the constant speculation of pundits in the days before the uprising was simply wasted breath. “Despite being hailed as a social media revolution,” he writes in the report, “monitoring the tone of only mainstream media around the world would have been enough to suggest the potential for unrest in Egypt.” He doesn’t assert that a threshold for global tone exists beyond which revolution becomes inevitable. Nor is the software powerful enough to predict progression or timing of events. Rather, tone tracking will reveal periods of “increased potential for unrest” that policy-makers could use to cut through the speculation. It’s an assessment of trends over long periods of time that are harder to spot by subjective and likely biased monitoring by governments and think tanks.

“The mere fact that the US President stood in support of Mubarak suggests very strongly that even the highest level analysis suggested that Mubarak was going to stay there,” Leetaru told the BBC News. “That is likely because you have these area experts who have been studying Egypt for 30 years, and in 30 years nothing has happened to Mubarak. If you look at this tonal curve it would tell you the world is darkening so fast and so strongly against him that it doesn’t seem possible he could survive.”

Tonal analyses pointed to similar darkenings for the leaders of Tunisia, Libya, and for the ethnic conflicts in the Balkans during the 1990s that Leetaru points out “caught many by surprise.”

Global news can also give us clues to locate individuals of interest. The software analyzed documents that contained the words “bin ladin” between 1979 and April 2011, and mapped associated locations in those documents. Many would have put their money on Bin Ladin being in Afghanistan. Tallying up locations associated with the elusive Al Qaeda leader shows that only 28 percent of news reports had him in Afghanistan while about half of them speculated he was in Pakistan. Abbottabad, which was mentioned only once following Bin Ladin’s discovery there, is less than 200 kilometers from the two cities most associated with him. Again, Leetaru acknowledges the limitations of the software. “While far from a definitive lock in Bin Ladin’s location,” he writes in the report, “global news content would have suggested Northern Pakistan in a 200 km radius around Islamabad and Peshawar as his most likely location, and that he was nearly twice as likely to be making his residence in Pakistan as Afghanistan.”

A map of locations associated with Osama Bin Ladin and their sources between 1979 and April 2011.

The current research extends the young field of ‘culturomics’ which attempts to quantify human culture “across societies and across centuries.” In collaboration with Google books, study released last year scanned a collection of digitized texts that represented 4 percent of all books ever printed. They quantified cultural trends reflected in the English language across diverse topics such as the “evolution of grammar, collective memory, the adoption of technology [and] the pursuit of fame.” The report acknowledges that this kind of high-throughput, computer-based tone monitoring is not as accurate as when tone is gauged by humans. In other efforts to predict the future through tracking trends, police in Los Angeles are currently experimenting with algorithms that predict crimes before they happen.

It’s one thing, however, to plot global tone and ask retrospectively if trends are pointed towards events that we know have already happened. It still remains to be seen whether Leetaru’s software can track events today to predict events of tomorrow. “It is obviously much easier to find precursory signs when you know where to look than to do it blindly,” Thomas Chadefaux, a political scientist at the Swiss Federal Institute of Technology said in an interview with the journal Nature. He also called Leetaru’s work “a welcome addition to a field – political science – that has cared very little about finding early warning signals for war, or making predictions at all.”

[image credits: Nature and First Monday]
image 1: Mubarak
image 2: Egypt trends
image 3: Bin Ladin location

Hi, I'm Watson. Let's talk about your latest medical claim, shall we?

Ever feel like the people reviewing your claims at the insurance company are a bunch of faceless, mindless drones? Well it turns out your paranoid cynicism is coming true – congratulations! In a historic step towards modernizing healthcare, insurance giant WellPoint is teaming up with IBM so that the Watson AI can help their staff make informed decisions. The amount of medical literature is staggering – doubling in size about every five years. No human can possibly hope to keep up. Watson, however, is able to process 200 million pages of content in just three seconds. In early 2012, some WellPoint nurses will be able to access Watson to assist them in reviewing patient cases and treatment requests. Later, WellPoint expects to roll out the service to a few oncology practices and eventually this technology could be helping medical professionals all over the world. Taste a bit of this AI-fueled medical utopia in the video from IBM below. While many will undoubtedly lament the “dehumanizing” nature of AI in healthcare, Watson represents one of the great hopes in this field. With unparalleled abilities to understand human writing in huge amounts at high speeds, Watson can place the entire global history of medical literature in the hands of your doctor. Faceless and cold? Perhaps. But better and smarter as well.

As we discussed when Watson was poised to take on Jeopardy! champions earlier this year, the IBM system is a wonder when it comes to understanding natural human language. That’s not an easy skill for a computer, yet Watson is able to process billions of documents to find correlations between key words and concepts. This allowed the AI to win at Jeopardy! and it also allows it to recommend the best solutions to medical problems by searching through records and finding the most likely causes for ailments. Early tests with Watson have shown it to process complex symptoms, recognize possible causes with uncanny range and accuracy, and suggest treatments that impress the doctors working with the AI. The big goal of Watson in medicine is to give every medical professional access to the system, not to replace human decision making but to improve upon it by keeping humans informed. Here’s a brief video from IBM that highlights that vision:

WellPoint is one of the largest insurance groups in the US (the largest by some counts) with one in nine Americans participating. They have 34 million+ members and 70+ million when counting subsidiaries. Their agreement with IBM is historic not because it’s the first time AI is being applied to medical databases (it’s not, this has been done before) but because of the scale of the company and Watson’s superior understanding of human language. Early next year when Watson is made available to WellPoint personnel it will be used to sort through patient charts, medical records, WellPoint’s history of treatments, and IBM’s library of medical textbooks and journals so that the computer can synthesize treatment recommendations using all of these sources. We’re talking about millions upon millions of documents reviewed in just seconds. Nurses will take Watson’s recommendations as guidance, but conflicts in claims (your doctor wants to give you Drug X but Watson recommends Drug Y) will still be settled by human reviewers.

From there, WellPoint expects to roll out the medical version of Watson to oncology practices to assist in complex treatment plans. No word yet on the time it will take for that application to come online, nor how long WellPoint and IBM will stay at that level before moving on towards more general uses of the technology.

Eventually, however, I think we’ll see Watson, or its successor, commonly used in all areas of the medical industry from patient interviews to ER visits to insurance claim lawsuits. WellPoint expects the nurses in the opening trial will be able to access Watson from desktop computers or even mobile devices, and I foresee a future where tablet computers (or smart phones) serve as easy to use portals to Watson medical solutions calculated in the cloud. That means that not only will Watson-like technology be available to all levels of medical professionals but it will be available almost anywhere.

Consider that possibility for a moment. Every doctor (or nurse, or claim adjuster, etc) would have access to the world’s collection of medical expertise at their fingertips. And, because Watson not only makes recommendations but also cites the sources it used to make its decisions, your medical provider could quickly perform their own research to verify or expand upon Watson’s suggestions. This technology may turn every doctor into a super doctor. If we could make it freely accessible via smart phones it could also turn even untrained humans into passable medical assistants. These are world-changing possibilities.

And they come with world-sized risks. While I’m not a pessimist, there are clear avenues for abuse or failure. What if insurance companies (or governments, or whoever) got to decide that some treatments were too costly, and kept Watson from suggesting them? What if medical professionals became so trusting of Watson that an error in the AI would be accepted as medical fact and cause misdiagnoses of patients? What if Watson (or its successor) had no real competition, making an entire industry reliant upon a single system, and again allowing an error to be propagated to millions across the world?

WellPoint (and IBM) have assured the public (1,2) that they only intend Watson to be an objective assistant, and that it won’t take financial cost-benefit into account. Likewise, they have openly stated that Watson is here to augment human judgement, not replace it. Yet the fact that these companies must come out and assuage our fears about these risks reinforces the fact that such possibilities do exist.

That doesn’t mean we should let our fears keep us from adopting Watson on a large scale in the medical industry, but we should go into this with open eyes. AI could be one of the most powerful and transformative tools in world healthcare.  It is overwhelming a positive trend, and we need to support it and grow it as much possible. But, while this technology is still young, we should be very vocal about how we, as patients and human medical professionals, want it to be developed. One way or another medical AIs are going to revolutionize the industry. If WellPoint, IBM, and the public work together we can ensure that the revolution leaves us with a much healthier world.

[image and video credit: IBM]
[sources: WellPoint press release, IBM]

$4 million can buy you a lot of AI journalism...and a new name apparently.

I have met my replacement and it is a robot. A robot obsessed with sports. AI journalism pioneer StatSheet gained a name for itself by transforming raw sports data into compelling narratives that people would want to read. To date their automated writing programs have created 100,000+ articles, 350+ free mobile apps, and 1 million unique pages using 2 billion+ pieces of statistical data. Now they are about to do much, much more. Raising $4 million in venture capital, StatSheet is changing its name to Automated Insights and expanding its visions beyond sports. Why restrict your AI writer to one field? Automated Insights will bring their data driven narratives to healthcare, finance, real estate, and many other applications. If something is too expensive, too repetitive, or too boring for a human to write, chances are that Automated Insights can have their AI do it instead. My fellow biology-based writers, our newest member is going to take the jobs we don’t want…and then maybe the rest of the jobs, too.

While StatSheet has been around since 2007, their expansive network of automatically generated sports sites really took off last fall when the company produced a unique page for every single NCAA Division I college basketball team. Since then, they’ve been busy. This spring they created a program that would use basketball statistics to help you talk trash during March Madness. This month they’ve extended their automated coverage to include all the Major League Baseball teams (bringing the total up to 375). In the next couple of weeks they’ll expand into all 32 NFL teams, and the 244 NCAA Division I college football teams as well. At their current pace they are generating 15,000 to 20,000 new articles every month. All this with just a dozen or so employees. StatSheet manages the AI, the AI creates the content.

Content which is limited, but not at all obviously created by a computer. Their automated game previews and play by play coverage are decent. Not mind-blowing editorial commentary on the sports world, just readable articles about the numbers and outcomes that so many fans find enthralling. When I reviewed StatSheet earlier I actually got confused between human and AI generated content, and CEO Robbie Allen had to straighten me out. At this level of analysis they both seem about the same. Still, there’s little real worry that such AI writing could replace the indepth human analysis that most journalists pride themselves on.

Of course it doesn’t have to. Part of the large appeal of StatSheet is that they can create content for niche fans. Who cares if your team is last in the league – the AI journalist can still provide you with the regular updates and analysis that you crave. Their database (2 billion+ data points and counting) is exhaustive enough to give you the sort of historical perspectives and long term narratives that human journalism doesn’t even attempt unless it knows it will sell well. When journalism is automated, no topic is too small, no perspective too unpopular to get attention. If there is enough data, the AI will write the story. Simple as that.

This new graphic sums it up nicely, I think. If you have data, Automated Insight will turn it into something your audience will want to read. Size is not an issue.

In their new guise as Automated Insights, the North Carolina based company is going to bring the same niche analysis to all new fields. Do you have piles of data about real estate prices? Automated Insights could transform that into compelling articles for business or public consumption. Mounds of actuarial data? Why hire dozens of human workers to sort through it when an AI can do the same quicker and more consistently. Here’s a list of potential fields that Automated Insights is looking to pursue: life sciences, healthcare, business intelligence, sales productivity, monitoring solutions, sports, finance, weather, real estate, and local interest. Each is an application space with mountains of data and many small markets which would be interested in transforming stats into narratives that are targeted to their niche concerns. We’re talking about millions of potential websites, mobile apps, etc – each specifically but automatically constructed so that a human can read a few short articles instead of pouring through tons of raw data. Automated Insights is poised to take the most boring writing known to humanity and foist it off onto a computer.

Sounds awesome.

Not to mention completely necessary. As we continue to build vast networks of data mining operations, and create immense archives of digital information we’re going to need some means of understanding everything we’ve collected. AI writing is a good solution. A machine pours over millions of statistics and translates a the relevant information into a concise narrative that you can comprehend. Such a system could help professionals everywhere stay on top of the exact amount of news that pertains to their jobs. Quick, efficient, and fairly cheap because it’s all made with AI labor.

For now, writers need not fear Automated Insights. They aren’t actually taking jobs we don’t want, they’re creating jobs that didn’t exist before. The first markets for AI writing are the markets that were too unprofitable for human labor to explore. As artificial intelligence improves however, they’ll produce content that will make them competitive with humans in a variety of fields. Journalism, sales figures, stock analysis, etc. We’ve already seen this transition start to happen with lawyers.

Looking forward, I’m intimidated by this company’s growth. Today, Automated Insight’s writing is formulaic, easily digestible, and not terribly awe-inspiring. $4 million in venture funding, however, is going to help Automated Insights not only expand its applications, but improve its algorithms. How long until it can be humorous? How long until it can cater to preferences at the individual level? How long until it’s as good as a human reporter at more complex story-telling? Sooner than we imagine, I think. And Automated Insights is far from the only group pursuing this kind of AI writing. It’s just a matter of time until we read as much automated content as we do human content.

Well, if you can’t beat the robot revolution, why not join it? Automated Insights is hiring. I’ll leave you with a quick look at the company’s appeal for applications. It’s good to know that the people who are shortening the lifespan of my career are having such a good time. (Just kidding, guys. Kudos on your funding and your work, and I look forward to seeing what you produce next.)

[image credits: Statsheet (modified), Automated Insights]
[video credits: Automated Insights]
[source: Automated Insights]

Oh, the things you hear on the internet.

I can think of few things as annoying as being forced into a conversation with an idiot. But when that idiot you’re talking to turns out to be an identical copy of yourself…well you’ve just entered into realm of meta-annoyance only accessible to artificial intelligence. In a fit of curiosity or pique, Cornell’s Creative Machine Lab decided to see what happens when an AI tries to talk to itself. They had the AI conversationalist Cleverbot briefly interact with itself and then displayed that exchange as a video using text-to-speech. The results were pretty freakin’ hilarious. Check out the discussion on God, unicorns, and robots in the video below. Seems like we need another version of the Turing Test to let us know when computers have reached humanity’s level. If an AI can’t stand to talk to itself for more than a minute, it’s not nearly narcissistic enough to be a real person.

The program Cleverbot is a web-based application that talks to people through a text interface. It’s one of many such “chatbots” you can find online, each able to respond to messages you type. Cleverbot learns to be a better conversationalist by remembering all the previous discussions it has had (20 million+ so far) and choosing which previous statements made by humans best fit the current discussion it’s having with a human. If you want, you can go to the Cleverbot site right now and participate in its learning process. When you do, I want you to keep in mind what you see in the following video from Cornell’s Creative Machines Lab. We (the internet) taught Cleverbot how to converse. If even it seems to find itself ridiculous and hard to listen to, what does that say about us?

Of course, I’m kidding when I say that Cleverbot finds itself annoying. Cleverbot doesn’t have emotions. It really is just a smart way of learning language by having conversations. While that process mimics the development we see in our children, Cleverbot doesn’t come with the hormones, senses, and environmental context that makes that education a fundamental part of being human.

Besides, hook Cleverbot up to itself a million times and you’ll create a million different conversations, some of which, I’m betting, make it seem happy, enthralled, and every other emotional state we possess. (Cornell, let me know if you actually do that, it would be wonderful to hear the results). In the end, no matter how we perceive Cleverbot’s reaction to itself, it’s simply stepping through the same algorithms it would if it were talking to anyone of us. Cleverbot is a mirror, a very intelligent mirror, but just a mirror.

So maybe what we should learn from this video is that the humanity Cleverbot reflects isn’t very illuminating. We don’t need an AI to tell us that online conversations are often random, inane, and unnecessarily aggressive (though this experiment is a hilarious reminder). What we may need is a warning that Turing Tests and other measures of artificial intelligence could put humanity in a very awkward situation much sooner than we think. Watching Cleverbot talk to itself I am left with little doubt that it has years to go before it reaches a human-level of conversational skill. Yet it is clearly an advanced platform, and one that learns from an exponentially growing online community. Even this funny, strange conversation is a good indicator that AIs will eventually be able to pretend to be humans and we won’t know the difference. In fact, there’s many anecdotal examples of that already happening in the past, and one chatbot was even able to fool a human in an actual Turing Test. Clearly we’re not at human-level conversation yet, but just as clearly we’re making our way there.

Give it time, and talking with Cleverbot will be less wacky, and more enjoyable. Hell, it might even be profound. I can’t wait to see this experiment repeated in another ten years. We may not be able to distinguish it from every other conversation we overhear on the street. Who knows, getting chatbots to talk with themselves may lead to a new kind of AI generated philosophy. I’m certainly open to any school of thought that can answer the tough questions about unicorns, robots and which of us are meanies.

[screen capture and video credits: Cornell Creative Machines Lab]
[source: CCML]

The man behind the math behind the software, Santa Clara University's George Mohler.

The police officers arrived at the parking garage in downtown Santa Cruz and spotted two women behaving suspiciously. No crime had been committed, but peering through the windows of the parked cars was sketchy enough. The officers questioned the women: one had outstanding warrants; the other was in possession of illegal drugs.

What’s strange about this scenario is that no one had called the cops. In fact, the cops didn’t even know that the women would be there, just that the probability of a crime being committed at that location, at that time of day, was especially high. In one of the first cases of ‘predictive policing,’ law enforcement were able to calculate where the criminals would be and arrest them before the crime could be committed.

Oh yeah, totally “Minority Report,” absolutely “Numb3rs.”

Except it’s not Hollywood, it’s real. In July the Santa Cruz Police Department began experimenting with an interesting bit of software developed by scientists at Santa Clara University. The researchers behind the software are like an intellectual “Oceans Eleven” team of specialists: two mathematicians, an anthropologist and a criminologist. They’ve combined their cerebral forces to come up with a mathematical model that takes crime data from the past to forecast crimes in the future. The basic math is similar to that used by seismologists to predict aftershocks following an earthquake (also a handy bit of software in southern California).

Large earthquakes are unpredictable, but the aftershocks that follow are not and their occurrence can be predicted with mathematical models. It occurred to Dr. George Mohler, one of the Santa Clara mathematicians, that criminal activity might not be random and that, similar to aftershocks, some crimes might be predicted by other crimes that precede them. The reasoning is based on the assumption that crimes are clustered – it’s what police call ‘hotspots.’ Burglaries will occur in the same area and at the same houses because the vulnerabilities of that area will be known to the burglars. Gang violence is also clustered. A gang shooting will often trigger retaliatory shootings.

Using the aftershocks-inspired algorithms Dr. Mohler and his team came up with a model, then sought to test it. In collaboration with the LAPD they plugged in data on 2,803 residential burglaries occurring within a block of the San Fernando valley 11 miles by 11 miles throughout 2004. For a given day the software calculated the top 5 percent of city blocks most likely to be burglarized. The results convinced the LAPD that, had they been using the program, they could have prevented a quarter of burglaries across the entire test region for that day.

The current, real world test of the software involves generating a map of the city areas most likely to be burglarized, the time of day they are most likely to get hit, and deploying personnel accordingly. The software is recalibrated every day when burglaries from the previous day are added to the dataset. They don’t actually expect to catch people in the act, but to deter more crimes with more effective patrolling. The test that is underway will be evaluated at six months, but already the data is encouraging. Zach Friend, crime analyst for Santa Cruz police, confirmed to the New York Times that the program led to five arrests in July. Even more impressive, compared to July 2010 burglaries, the number of July 2011 burglaries are down 27 percent. Whether or not that trend holds remains to be seen, but so far it appears that being in the wrong place at the right time works.

Artificial intelligence augments human intuition. Predictive policing software avoids human bias.

Mathematical models are only as good as their predictive power, and the ability to predict requires algorithms which are based on accurate data. Given the fact that the data supplied by the Santa Cruz Police Department wasn’t collected with mathematical algorithms in mind, I asked Dr. Mohler if there were another kind of data that he wished he was getting that simply isn’t available. His answer suggests there is, but it doesn’t come from the police. “Part of this falls on the public. Crimes…need to be reported if predictive policing is going to be as effective as possible. Once reported, it would be good to have high spatial accuracy and realistic estimates of time windows in which crime happened.”

The Santa Clara software isn’t the first of its kind. Other police departments have been experimenting with their own predictive software. But according to Dr. Mohler, comparisons show that their software outpredicts the others. And they plan to develop software that predicts crimes other than burglaries. Because gang violence begets more gang violence it is amenable to the same type of chain reaction-dependent analysis. Dr. Mohler and his colleagues have already begun working on a gang violence model using the activities of three gang rivalries in Los Angeles. Evidently retaliations commonly occur within days of and at nearly the same location as the initial attack. Dr. Mohler hopes software might be developed for still other types of crimes in the future.

One impetus for adopting predictive policing is the downturn in the US economy. As police departments are pressured to downsize it becomes that much more important to patrol intelligently and efficiently. With only 26 officers for every 10,000 residents Los Angeles is particularly short-handed (Chicago has 46). “We’re facing a situation where we have 30 percent more calls for service but 20 percent less staff than in the year 2000, and that is going to continue to be our reality,” Mr. Friend told the New York Times. “So we have to deploy our resources in a more effective way, and we thought this model would help.”

Given that the crime-fighting software is the real world version “Numb3rs,” the television show in which a genius helps police solve crimes through math, one might expect Dr. Mohler was an avid viewer. Turns out he’s only seen the show twice, but what he saw was pretty accurate. “The pilot episode concern[ed] geographic profiling and matche[d] reasonably well with what is done in practice. I’m sure this doesn’t hold throughout the course of the series, but getting people excited about math isn’t a bad thing in my opinion.” If the six month evaluation of the software shows it to be effective in decreasing crime its use will undoubtedly spread to other cities in the US and the rest of the world. If life imitates art and our streets are made safer, I imagine math might get more exciting for a lot of people.

[image credits: Daily News and Santa Clara University]
image 1: George Mohler
image 2: Density map

It may not look like much but AirRobot's AR100B is a choice drone for battlefield surveillance.

At less than a kilogram total weight, a max payload of 200 grams, and just one meter in diameter the AR100B is a wholly different breed of drone compared with the MQ-9 Reaper – the drone with the most with its 3,000+ lbs of explosives and Hellfire missiles. But what it lacks in brawn it more than makes up for in surveillance capability. At over 1,200 Afghanistan missions and counting it’s clear that the AR100B unparalleled ease of use and agility makes it a highly-valued part of ground tactics strategy.

The AR100B comes from AirRobot, the unmanned aerial vehicle (UAV) platform producers with offices all over the world including the UK and the US. As the latest upgrade of AirRobot’s UAV platforms, the AR100B boasts more powerful engines and longer duration batteries. It’s quiet and easy to control – so easy in fact troops routinely take the stick with little or no training. Its four rotors give it vertical take-off and landing (VTOL) capability, with an operational radius of 1,000 meters and max flight time of 20 minutes. Infra Red and low light cameras the AR100B can be used to collect high definition video or take photos during inspection and surveillance tasks. They give ground troops a bird’s eye view into otherwise obstructed areas to scan for possible dangers such as hidden insurgents and armored vehicles. The AR100B may not be the sexiest drone, but it may just be a soldier’s best friend.

[image credits: AirRobot-UK]
image: AR100B

Don't blame the car, blame the human driver who is obviously trying to hide his face.

As if to spite me just days after I was gushing about Google’s driverless Toyota Prius, they go and rear-end someone and cause a five car accident. But as it turns out, it wasn’t the car’s fault at all: the crash was due to human error. And faster than you can dig your driver license out of the trash, Google’s putting the word out that the robot was not at fault. A spokesperson told Business Insider, “Safety is our top priority. One of our goals is to prevent fender-benders like this one, which occurred while a person was manually driving the car.”

The story broke when a Jalopnik reader sent in a photo of the accident which took place not far from Google’s Mountain View, California headquarters. The image of a woman on her cell phone, a police officer on the scene and the forlorn Google driver leaning on his car – the street view mount sticking out like a torn brake pad – is certainly not the kind of press that’s going to help Google fulfill their dreams of an accident-free robotic car paradise.

Of course, we’ll have to wait for the police report to confirm that it was the driver’s fault with his languid 500 millisecond reaction time. But it’s not hard to believe that one human rear-ended another, and as it stands this accident is a clear-cut example of why we need robotic cars. Hopefully this little bump in the road doesn’t slow Google down too much. It will most definitely speed up thinking about and perhaps setting guidelines for litigation involving driverless cars. Who’s at fault? A related question: will all insurance premiums be the same regardless of sex, age, prior nondriverless driving record?

For my money, I’m still behind the car with my hands off of the steering wheel. The fact that it was a human being’s fault means the robot’s still got 160,000+ miles of accident-free driving behind it. How many humans can make the same claim?

[image credit: Jalopnik]
image: Google accident