Tomorrow is Today

Researchers Using Computer Analysis to Cut Through Obfuscation in Financial Reports 

COMPANY financial reports don’t usually make for thrilling reading, but with the ability to make or break fortunes, they come under intense scrutiny. Now software that can extract information from the nuanced language of such reports could provide investors with the edge they need to stay ahead of the competition.

“Financial statements carry important information about the health of reporting companies,” says Chao-Lin Liu at National Chengchi University in Taipei, Taiwan. But companies habitually downplay negative aspects by using ambiguous language and burying nuggets of information in pages of droning prose.

Text-mining techniques generally concentrate on single words: counting the number of negative or positive words in a body of text can give an indication of the overall tone, for example. But it is impossible to say whether certain words taken in isolation - such as “increased” - are positive or negative, says team member Yuan-Chen Chang. So the team designed an algorithm to recognise meaningful phrases instead.

To do this, Liu and his colleagues use statistical models to automatically identify what they call opinion patterns - subjective phrases paired with an opinion holder. For example, the sentence “The Company believes the profits could be adversely affected” contains the opinion holder “The Company” and the subjective phrases “believes” and “could be adversely affected”.

(via Mine your language: Software decodes company reports - tech - 02 November 2012 - New Scientist)

Algorithm Mines Contact Data to Build Map of Social Relationships, Even Hidden Ones

…The ability to automatically create circles from a user’s contacts list is certainly valuable, The algorithm also has the ability to add new contacts to appropriate circles.

An important limitation, however, is the scalability of the approach. McAuley and Leskovec admit their algorithm is not particularly efficient, taking about an hour to identify ten circles from a list of 1000 Facebook contacts. That’s a lot of hours of processing for Facebook’s 1 billion users. However, they say that the technique should be quicker as broader patterns become clear once all users contacts have been taken into account.

For example, it may be possible to identify the set of all people on Facebook who went to a particular university. Then one person’s circle might consist of the intersection between this set and their contact list. Just how much of a speed up this would allow isn’t clear though.

Another important question for the future is how well in principle automatically-generated circles can be made to match ground truth circles, using only the information available in contact profiles and so on. It may be that many circles are created using information that users do not explicitly make available on social networks, such as a circle of ‘best friends’. If that’s the case, then these algorithms will never be able to reconstruct the ground truth circles perfectly. But perhaps this doesn’t matter if they provide a reasonable approximation to ground truth circles that users can tinker with at their leisure.

Another interesting approach is to look for patterns of links between contacts that users do not turn into circles—in other words connections between people that users have not recognised or want to keep hidden. Such a pattern might be linked with criminal activity, for example, or point to marketing information that could be sold.

(via Algorithm Predicts Circles of Friends Using Contacts Data | MIT Technology Review)

DARPA Funded Research Aims to Automate Video Surveillance, Teach Computers to Identify Suspicious Behavior

This approach relies heavily on advances by machine vision researchers, who have made remarkable strides in last few decades in recognizing stationary and moving objects and their properties.

It’s the same vein of work that led to Google’s self-driving cars, face recognition software used on Facebook and Picasa, and consumer electronics like Microsoft’s Kinect.

When it works well, machine vision can detect objects and people — call them nouns — that are on the other side of the camera’s lens. But to figure out what these nouns are doing, or are allowed to do, you need the computer science equivalent of verbs.

…that’s where Oltramari and Lebiere have built on the work of other Carnegie Mellon researchers to create what they call a “cognitive engine” that can understand the rules by which nouns and verbs are allowed to interact. Their cognitive engine incorporates research, called activity forecasting, …which tries to understand what humans will do by calculating which physical trajectories are most likely.

They say their software “models the effect of the physical environment on the choice of human actions.”

Both projects are components of Carnegie Mellon’s Mind’s Eye architecture, a DARPA-created project that aims to develop smart cameras for machine-based visual intelligence. Predicts Oltramari: “This work should support human operators and automatize video-surveillance, both in military and civil applications.”

(via U.S. looks to replace human surveillance with computers | Security & Privacy - CNET News)

Teaching Robots To Identify Potentially Useful Objects, Use Them As Tools

GA Tech Professor Mike Stilman, who studies robot navigation among movable obstacles, is studying the cognitive processes that enable humans to grab arbitrary objects and find creative new uses for them.

Future robots might scan their environments, see something on the ground, and quickly run through a list of possible uses for the thing. Maybe it sees a chair and realizes chairs have height, so it can use the chair to climb up to reach a window to escape a burning room. Or maybe it sees a metal pipe and realizes it can work as a lever to lift something heavy, freeing a human trapped under it. Stilman said the goal is to create machines that operate like MacGyver, the ’80s TV character.

This would be helpful for military robots working alongside human personnel—so the Office of Naval Research is funding Stilman’s work with a three-year, $900,000 grant. The first step is to build a “hybrid reasoning system,” which will use physics-based algorithms and a learning system to teach robots how to recognize and use various objects. They plan to test it using a headless humanoid robot called Golem Krang, seen above pretend-rescuing a trapped seaman. This research on a MacGyver-bot is one of its kind, according to the Navy.

(via ‘MacGyver’ Robot Could Make Use Of Objects It Finds | Popular Science)

Robots That Learn Through Demonstration Instead of Programming

Neil Dantam at the Georgia Institute of Technology in Atlanta and colleagues are developing a system in which robots learn by example. Instead of programming a robot to carry out a task, a demonstration could suffice.

However, the robot does not simply mimic the human’s actions. The prototype system uses an Xbox Kinect camera to observe the human performing an activity, then breaks that activity into a sequence of key actions necessary to carry out the task.

These actions are converted into a general set of instructions – much like those found in an IKEA furniture instruction manual – that can be interpreted even by non-humanoid robots. “You want to somehow capture the important aspects of the human’s motion and transfer that to the robot,” says Dantam. “Think about how you’d tell someone how to make a cake,” he says. The way people follow those steps may vary. “An adult might bend down over the counter to work while a child may stand on tiptoes.”

(via Robot learns using IKEA-style instructions - tech - 10 October 2012 - New Scientist)

AI Developed to Improve Relevance of Commentator Patter During Live Sports

WHEN watching sport on TV …a good commentator can make all the difference, peppering a play-by-play account of the action with expert knowledge and anecdotes. But even the best commentator’s repertoire is limited.

…Created by sports fan Greg Lee at the University of Alberta in Canada, [AI Program] Scores can tap into a stash of sporting stories to find relevant anecdotes that a commentator might not have thought of.

…The system works by matching the features of a live event - such as the teams, key players, the score and the remaining time - against a database of available stories. Once stories that include some of those features are found it selects the few that are most relevant and suggests them to a human commentator.

The challenge is in evaluating the relevance of candidate stories and ranking them. Lee’s system uses machine-learning techniques to do this. The most important feature was the teams involved and the second was the difference in number of runs.

To test the system, the researchers used it to create commentary for pre-recorded sports broadcasts and presented them to 254 volunteers, who said they found the commentary relevant and enjoyable. 

(via AI sports commentator knows all the best stories - tech - 04 October 2012 - New Scientist)

UPDATE: Google’s Cat-Video Identifying Neural Net Now Working to Improve Voice Commands, Image Search, Google Glass and Self-Driving Cars

Google is now using these neural networks to recognize speech more accurately, a technology increasingly important to Google’s smartphone operating system, Android, as well as the search app it makes available for Apple devices.

“We got between 20 and 25 percent improvement in terms of words that are wrong,” says Vincent Vanhoucke, a leader of Google’s speech-recognition efforts. “That means that many more people will have a perfect experience without errors.”

The neural net is so far only working on U.S. English, and Vanhoucke says similar improvements should be possible when it is introduced for other dialects and languages.

Other Google products will likely improve over time with help from the new learning software. The company’s image search tools, for example, could become better able to understand what’s in a photo without relying on surrounding text. And Google’s self-driving cars and mobile computer built into a pair of glasses could benefit from software better able to make sense of more real-world data.

(via Google Puts Its Virtual Brain Technology to Work - Technology Review)

See Also: “Google and Stanford have created the [digital equivalent of the] visual cortex of an infant human”

Algorithm Analyzing Artistic Style Identifies Same Patterns as Art Historians

Computer scientists Computer scientists… have developed a program that analyzes paintings in a manner similar to how expert art historians perform their analysis, and conducted an experiment that showed that machines can outperform untrained humans in the analysis of fine art.

In the experiment, the researchers used approximately 1, 000 paintings of 34 well-known artists, and let the computer algorithm analyze the similarity between them based solely on the visual content of the paintings, and without any human guidance.

Surprisingly, the computer provided a network of similarities between painters that is largely in agreement with the perception of art historians.

I take this research with a grain of salt: the algorithm seems to have detected the patterns it was designed to detect.  I would have been more excited if it drew a new connection that human art experts found interesting.

(via Can computers understand art? | KurzweilAI)

Agricultural Robots Can Tell The Difference Between Plants and Weeds, Can Thin Crops

Lettuce Bot uses a camera to image the plants beneath it. Machine learning algorithms then identify which ones are desirable and which are weeds. It can work with iceberg and romaine lettuces.

Once a plant is identified as a weed, a target spray, which is mounted behind the camera, will then shoot a targeted spray of an organic compound, such as hot steam or hot organic oil, at the plant and the plant will quickly die,” the company told Startup Lab.

The plant-classification algorithm is 98 to 99 percent accurate, and the kill mechanism is accurate to a quarter of an inch when the prototype is moving a 1 mph. The firm wants it to move at 3 mph while keeping it on target. Blue River says its machines will be more efficient than other means of weed-killing, and will work well in organic fields or those that have chemical-resistant weeds.

(via Down on the farm, Lettuce Bot is quietly slaying weeds | Cutting Edge - CNET News)

Researchers Developing Software to interpret Doodled Images

[A] new algorithm developed at Brown University and the Technical University of Berlin [is] the first computer application designed for “semantic understanding” of abstract drawings, and the research team says it could improve search applications and sketch-based interfaces…

The program can identify simple abstract sketches 56 percent of the time, compared to humans’ 73 percent average. Even those sorely lacking in verisimilitude can be detected, which is the key breakthrough here.

Computers can already recognize accurate sketches, like a police sketch of a suspect compared to photos of a face, for example. But for the type of abstract sketches we all grow up with, it’s a different challenge.

[I]f you’re asked to draw a rabbit, you would probably draw something with buck teeth, huge ears and exaggerated whiskers. Other people would easily recognize this cartoonish representation… But it doesn’t actually resemble the real thing in any meaningful way, so a computer would have no idea what it is …there are subtle tricks and meanings that a human can distinguish, but which present a tough challenge for something built in the black-and-white, ones-and-zeroes world.

[Researchers] devised a list of everyday things people might feel like doodling, settled on 250 categories and used Amazon’s Mechanical Turk crowdsourcing platform to hire some sketch artists. They took 20,000 unique sketches and fed them into existing machine-learning algorithms to train the system. The project culminated in a fun real-time computer Pictionary, where the system tries to recognize objects as the person draws them.

These are drawings of dogs, and what the computer thought they actually were.

To expand their data set, the team is thinking about gamifying this concept into something you can play on iOS or Android devices…

The goal would be improved sketch-based search, the researchers say. That could improve computer accessibility for speech, movement or literacy-impaired people — and it could work in any language, too. 

Video of the team pesenting the project at SIGGRAPH

(via Computer Learns to Recognize Badly Drawn Animals | Popular Science)