Blog: UMass Amherst Computer Scientist Leads the Way to the Next Revolution in Artificial Intelligence

UMass Amherst Computer Scientist Leads the Way to the Next Revolution in Artificial Intelligence
University of Massachusetts Amherst (04/02/12) Janet Lathrop

University of Massachusetts Amherst researchers are translating the "Super-Turing" computation into an adaptable computational system that learns and evolves, using input from the environment the same way human brains do. The model "is a mathematical formulation of the brain’s neural networks with their adaptive abilities," says Amherst computer scientist Hava Siegelmann. When the model is installed in a new environment, the new Super-Turing model results in an exponentially greater set of behaviors than the classical computer or the original Turing model. The researchers say the new Super-Turing machine will be flexible, adaptable, and economical. "The Super-Turing framework allows a stimulus to actually change the computer at each computational step, behaving in a way much closer to that of the constantly adapting and evolving brain," Siegelmann says.

View Full Article

Blog: Scientists Tap the Genius of Babies and Youngsters to Make Computers Smarter

Scientists Tap the Genius of Babies and Youngsters to Make Computers Smarter
UC Berkeley News Center (03/12/12) Yasmin Anwar

University of California, Berkeley researchers are studying how babies, toddlers, and preschoolers learn in order to program computers to think more like humans. The researchers say computational models based on the brainpower of young children could give a major boost to artificial intelligence research. "Children are the greatest learning machines in the universe," says Berkeley's Alison Gopnik. "Imagine if computers could learn as much and as quickly as they do." The researchers have found that children test hypotheses, detect statistical patterns, and form conclusions while constantly adapting to changes. “Young children are capable of solving problems that still pose a challenge for computers, such as learning languages and figuring out causal relationships,” says Berkeley's Tom Griffiths. The researchers say computers programmed with children's cognitive abilities could interact more intelligently and responsively with humans in applications such as computer tutoring programs and phone-answering robots. They are planning to launch a multidisciplinary center at the campus' Institute of Human Development to pursue their research. The researchers note that the exploratory and probabilistic reasoning demonstrated by young children could make computers smarter and more adaptable.

View Full Article

Blog: IBM Says Future Computers Will Be Constant Learners

IBM Says Future Computers Will Be Constant Learners
IDG News Service (02/17/12) Joab Jackson

Tomorrow's computers will constantly improve their understanding of the data they work with, which will help them provide users with more appropriate information, predicts IBM fellow David Ferrucci, who led the development of IBM's Watson artificial intelligence technology. Computers in the future "will not necessarily require us to sit down and explicitly program them, but through continuous interaction with humans they will start to understand the kind of data and the kind of computation we need," according to Ferrucci. He says the key to the Watson technology is that it queries both itself and its users for feedback on its answers. "As you use the system, it will follow up with you and ask you questions that will help improve its confidence of its answer," Ferrucci notes. IBM is now working with Columbia University researchers to adapt Watson so it can offer medical diagnosis and treatment. Watson could serve as a diagnostic assistant and offer treatment plans, says Columbia professor Herbert Chase. Watson also could find clinical trials for the patient to participate in. "Watson has bridged the information gap, and its potential for improving health care and reducing costs is immense," Chase says.

View Full Article

Blog: U.S. Intelligence Group Seeks Machine Learning Breakthroughs

U.S. Intelligence Group Seeks Machine Learning Breakthroughs
Network World (12/02/11) Michael Cooney

The U.S. Intelligence Advanced Research Projects Activity (IARPA) announced that it is looking for new ideas that may become the basis of cutting-edge machine-learning projects. "In many application areas, the amount of data to be analyzed has been increasing exponentially [sensors, audio and video, social network data, Web information], stressing even the most efficient procedures and most powerful processors," according to IARPA. "Most of these data are unorganized and unlabeled and human effort is needed for annotation and to focus attention on those data that are significant." IARPA's request for information asks about proposed methods for the automation of architecture and algorithm selection and combination, feature engineering, and training data scheduling, as well as compelling reasons to use such approaches in a scalable multi-modal analytic system and whether supporting technologies are readily available. IARPA says that innovations in hierarchical architectures such as Deep Belief Nets and hierarchical clustering will be needed for useful automatic machine-learning systems. It wants to identify promising areas for investment and plans to hold a machine learning workshop in March 2012.

View Full Article

Blog: How Revolutionary Tools Cracked a 1700s Code

How Revolutionary Tools Cracked a 1700s Code
New York Times (10/25/11) John Markoff

A cipher dating back to the 18th century that was considered uncrackable was finally decrypted by a team of Swedish and U.S. linguists by using statistics-based translation methods. After a false start, the team determined that the Copiale Cipher was a homophonic cipher and attempted to decode all the symbols in German, as the manuscript was originally discovered in Germany. Their first step was finding regularly occurring symbols that might stand for the common German pair "ch." Once a potential "c" and "h" were found, the researchers used patterns in German to decode the cipher one step at a time. Language translation techniques such as expected word frequency were used to guess a symbol's equivalent in German. However, there are other, more impenetrable ciphers that have thwarted even the translators of the Copiale Cipher. The Voynich manuscript has been categorized as the most frustrating of such ciphers, but one member of the team that cracked the Copiale manuscript, the University of Southern California's Kevin Knight, co-published an analysis of the Voynich document pointing to evidence that it contains patterns that match the structure of natural language.

View Full Article

Blog: Cops on the Trail of Crimes That Haven't Happened

Cops on the Trail of Crimes That Haven't Happened
New Scientist (10/12/11) Mellisae Fellet

The Santa Cruz, Calif., police department recently started field-testing Santa Clara University-developed software that analyzes where crime is likely to be committed. The software uses the locations of past incidents to highlight likely future crime scenes, enabling police to target and patrol those areas with the hope that their presence might stop the crimes from happening in the first place. The program, developed by Santa Clara researcher George Mohler, predicted the location and time of 25 percent of burglaries that occurred on any particular day in an area of Los Angeles in 2004 and 2005, using just the data on burglaries that had occurred before that day. The Santa Cruz police department is using the software to monitor 10 areas for residential burglaries, auto burglaries, and auto theft. If the program proves to be effective in thwarting crime in areas that are known for their high crime rates, it can be applied to other cities, says University of California, Los Angeles researcher Jeffrey Brantingham, who collaborated on the algorithm's development.

View Full Article

Blog: "Ghostwriting" the Torah?

"Ghostwriting" the Torah?
American Friends of Tel Aviv University (10/11/11)

Tel Aviv University (TAU) researchers have developed a computer algorithm that could help identify the different sources that contributed to the individual books of the Bible. The algorithm, developed by TAU professor Nachum Dershowitz, recognizes linguistic cues, such as word preference, to divide texts into probable author groupings. The researchers focused on writing style instead of subject or genre to avoid some of the problems that have vexed Bible scholars in the past, such as a lack of objectivity and complications caused by the multiple genres and literary forms found in the Bible. The software searches for and compares details that human scholars might have difficulty detecting, such as the frequency of the use of function words and synonyms, according to Dershowitz. The researchers tested the software by randomly mixing passages from the Hebrew books of Jeremiah and Ezekiel, and instructing the computer to separate them. The program was able to separate the passages with 99 percent accuracy, in addition to separating "priestly" materials from "non-priestly" materials. "If the computer can find features that Bible scholars haven't noticed before, it adds new dimensions to their scholarship," Dershowitz says.

View Full Article

Blog: New Mathematical Model to Enable Web Searches for Meaning

New Mathematical Model to Enable Web Searches for Meaning
University of Hertfordshire (09/23/11) Paige Upchurch

University of Hertfordshire computer scientist Daoud Clarke has developed a mathematical model based on a theory of meaning that could revolutionize artificial intelligence technologies and enable Web searches to interpret the meaning of queries. The model is based on the idea that the meaning of words and phrases is determined by the context in which they occur. "This is an old idea, with its origin in the philosophy of Wittgenstein, and was later taken up by linguists, but this is the first time that someone has used it to construct a comprehensive theory of meaning," Clarke says. The model provides a way to represent words and phrases as sequences of numbers, known as vectors. "Our theory tells you what the vector for a phrase should look like in terms of the vectors for the individual words that make up the phrase," Clarke says. "Representing meanings of words using vectors allows fuzzy relationships between words to be expressed as the distance or angle between the vectors." He says the model could be applied to new types of artificial intelligence, such as determining the exact nature of a particular Web query.

View Full Article

Blog: Computers Can See You--If You Have a Mug Shot; aulity of facial recognition systems

Computers Can See You--If You Have a Mug Shot
Wall Street Journal (09/03/11) Carl Bialik

Carnegie Mellon University (CMU) researchers recently presented data suggesting that facial recognition tools could identify individuals based on posed mug shots. The researchers demonstrated that, in principle, 33 percent of people photographed could be matched with a database of photos taken from Facebook. As part of the study, the researchers used images of 93 volunteers and compared them to Facebook photos of people on the CMU network. The results mean no one using facial-recognition software can claim "I can recognize any person in the U.S. at this very moment," says CMU's Ralph Gross. The problem is taking one image and comparing it to a wide set of images to find a single correct match. Comparing photos of just one person is easier and has achieved much more success. In a recent U.S. National Institute of Standards and Technology test, facial recognition software misidentified individuals in photographs just one percent of the time. Compared to Facebook images, closed-circuit (CC) TV images will probably be even more difficult to use with facial recognition systems, according to computer-vision experts. "Identifying faces in CCTV-quality images requires human experts," says University of Cambridge professor John Daugman.

View Full Article

Blog: Researcher Teaches Computers to Detect Spam More Accurately

Researcher Teaches Computers to Detect Spam More Accurately
IDG News Service (08/10/11) Nicolas Zeitler

Georgia Tech researcher Nina Balcan recently received a Microsoft Research Faculty Fellowship for her work in developing machine learning methods that can be used to create personalized automatic programs for deciding whether an email is spam or not. Balcan's research also can be used to solve other data-mining problems. Using supervised learning, the user teaches the computer by submitting information on which emails are spam and which are not, which is very inefficient, according to Balcan. Active learning enables the computer to analyze huge collections of unlabeled emails to generate only a few questions for the user. Active learning could potentially deliver better results than supervised learning, Balcan says. However, active learning methods are highly sensitive to noise, making this potentially difficult to achieve. Balcan plans to develop an understanding of when, why, and how different kinds of learning protocols help. "My research connects machine learning, game theory, economics, and optimization," she says.

View Full Article

Blog: Challenging the Limits of Learning [... language acquisition]

Challenging the Limits of Learning
American Friends of Tel Aviv University (01/19/11)

Researchers at Tel Aviv University have developed software that models the human mind to explore language acquisition, and report that the early results suggest that people actually learn language. The program learns basic grammar using a bare minimum of cognitive machinery, similar to what a child might have, says Tel Aviv's Roni Katzir. He used unsupervised learning to program his computer to learn simple grammar on its own, and the machine-learning technique enabled the program to see raw data and conduct a random search to find the best way to characterize what it sees. The computer looked for the simplest description of any data using the Minimum Description Length criterion. Katzir was able to explore what kinds of information the human mind can acquire and store unconsciously, and whether a computer can learn in a similar manner. He believes the research has applications in technologies such as voice-dialogue systems, or for teaching robots how to read visual images. "Many linguists today assume that there are severe limits on what is learnable," Katzir says. "I take a much more optimistic view about those limitations and the capacity of humans to learn."

View Full Article

Blog: IBM Computer Gets a Buzz on for Charity Jeopardy!

IBM Computer Gets a Buzz on for Charity Jeopardy!
Associated Press (01/13/11) Jim Fitzgerald

IBM's Watson computer beat former Jeopardy! champions Ken Jennings and Brad Rutter in a 15-question practice round in which the hardware and software system answered about half of the questions and got none of them wrong. Watson, which will compete in a charity event on Jeopardy! against Jennings and Rutter on Feb. 14-16, recently received a buzzer, the finishing touch to a system that represents a huge step in computing power. "Jeopardy! felt that in order for the game to be as fair as possible, just as a human has to physically hit a buzzer, the system also would have to do that," says IBM's Jennifer McTighe. Watson consists of 10 racks of IBM servers running the Linux operating system and has 15 terabytes of random-access memory. The system has access to the equivalent of 200 million pages of content, and can mimic the human ability to understand the nuances of human language, such as puns and riddles, and answer questions in natural language. The practice round was the first public demonstration of the computer system. IBM says Watson's technology could lead to systems that can quickly diagnose medical conditions and research legal cases, among other applications.

View Full Article

Blog: Mathematical Model Shows How Groups Split Into Factions

Mathematical Model Shows How Groups Split Into Factions
Cornell Chronicle (01/03/11) Bill Steele

A mathematical model of how social networks evolve into opposing factions under strain has been developed by Cornell researchers. Earlier models of structural balance demonstrate that under suitable conditions, a group conflict will facilitate a split into just two factions, while the new model indicates how friendships and rivalries change over time and who ends up on each side. The model consists of a simple differential equation applied to a grid of numbers that can stand for relationships between persons, countries, or corporations. Cornell's Seth Marvel says people may forge alliances based on shared values, or may consider the social effects of allying with a specific individual. "The model shows that the latter is sufficient to divide a group into two factions," Marvel says. The model traces the division of groups to unbalanced relationship triangles that trigger changes that spread throughout the entire network. All too frequently the final state is comprised of two factions, each with all favorable links among themselves and all negative connections with members of the opposing faction. The model shows that if the average strength of ties across the entire system is positive, then it evolves into a single, all-positive network.

View Full Article

Blog: UCSF Team Develops "Logic Gates" to Program Bacteria as Computers

UCSF Team Develops "Logic Gates" to Program Bacteria as Computers
UCSF News (12/08/10) Kristen Bole

University of California, San Francisco (UCSF) researchers have genetically engineered E. coli bacteria with a specific molecular circuitry that will enable scientists to program the cells to communicate and perform computations. The process can be used to develop cells that act like miniature computers that can be programmed to function in a variety of ways, says UCSF professor Christopher A. Voigt. "Here, we've taken a colony of bacteria that are receiving two chemical signals from their neighbors, and have created the same logic gates that form the basis of silicon computing," Voigt says. The technology will enable researchers to use cells to perform specific, targeted tasks, says UCSF's Mary Anne Koda-Kimble. The purpose of the research is to be able to utilize all of biology's tasks in a reliable, programmable way, Voigt says. He says the automation of biological processes will advance research in synthetic biology. The researchers also plan to develop a formal language for cellular computation that is similar to the programming languages used to write computer code, Voigt says.

View Full Article

Blog: How Rare Is that Fingerprint? Computational Forensics Provides the First Clues

How Rare Is that Fingerprint? Computational Forensics Provides the First Clues
UB News Services (12/07/10) Ellen Goldbaum

University at Buffalo researchers have developed a method to computationally determine how rare a particular fingerprint is and how likely it is to belong to a specific crime suspect. The Buffalo researchers created a probabilistic method to determine if a fingerprint would randomly match another in a database. The researchers say their study could help develop computational systems that quickly and objectively show how important fingerprints are to solving crimes. "Our research provides the first systematic approach for computing the rarity of fingerprints in a scientifically robust and reliable manner," says Buffalo professor Sargur N. Srihari. Determining the similarity between two sets of fingerprints and the rarity of a specific configuration of ridge patterns are the two main types of problems involved in fingerprint analysis, Srihari says. The Buffalo method relies on machine learning, statistics, and probability.

View Full Article

Blog: International Win for Clever Dataminer; Weka data-mining software

International Win for Clever Dataminer; Weka data-mining software
University of Waikato (08/17/09)

The first place finisher in the 2009 Student Data Mining Contest, run by the University of California, San Diego, used the Weka data-mining software to predict anomalies in e-commerce transaction data. Quan Sun, a University of Waikato computer science student, says it took about a month to find the answer. The contest drew more than 300 entries from students in North America, Europe, Asia, and Australasia. "I couldn't have done it without Weka," Sun says of the open source software that was developed at Waikato. "Weka is like the Microsoft Word of data-mining, and at least half of the competitors used it in their entries." ACM's Special Interest Group on Knowledge Discovery and Data Mining gave the Weka software its Data Mining and Knowledge Discovery Service Award in 2005. Weka has more than 1.5 million users worldwide.

View Full Article

Blog: NCSA Researchers Receive Patent for System that Finds Holes in Knowledge Bases

NCSA Researchers Receive Patent for System that Finds Holes in Knowledge Bases
University of Illinois at Urbana-Champaign (08/03/09) Dixon, Vince

Researchers at the National Center for Supercomputing Applications (NCSA) at the University of Illinois, Urbana-Champaign, have received a patent for a method of determining the completeness of a knowledge base by mapping the corpus and locating weak links and gaps between important concepts. NCSA research programmer Alan Craig and former NCSA staffer Kalev Leetaru were building databases using automatic Web crawling and needed a way of knowing when to stop adding to the collection. "So this is a method to sort of help figure that out and also direct that system to go looking for more specific pieces of information," says Craig. Using any collection of information, the technique graphs the data, analyzes conceptual distances within the graph, and identifies parts of the corpus that are missing important documents. The system then suggests what concepts may best fill those gaps, creating a link between two related concepts that might otherwise not have been found. Leetaru says this system helps users complete knowledge bases with information they are initially unaware of. Leetaru says the applications for this method are limitless, as the corpus does not have to be computer-based and the method can be applied to any situation involving a collection of data that users are not sure is complete.

View Full Article

Blog: Computers Unlock More Secrets of the Mysterious Indus Valley Script

Computers Unlock More Secrets of the Mysterious Indus Valley Script
UW News (08/03/09) Hickey, Hannah

A team of Indian and U.S. researchers, led by University of Washington professor Rajesh Rao, is attempting to decipher the script of the ancient Indus Valley civilization. Some researchers have questioned whether the script's symbols are actually a language, or are instead pictograms of political or religious icons. The researchers are using computers to extract patterns from the ancient Indus symbols. The researchers have uncovered several distinct patterns in the symbols' placement in sequences, which has led to the development of a statistical model for the unknown language. "The statistical model provides insights into the underlying grammatical structure of the Indus script," Rao says. "Such a model can be valuable for decipherment, because any meaning ascribed to a symbol must make sense in the context of other symbols that precede or follow it." Calculations show that the order of the symbols is meaningful, as taking one symbol from a sequence and changing its position creates a new sequence that has a much lower probability of belonging to the language. The researchers say the presence of such distinct rules for sequencing provides support for the theory that the unknown script represents a language. The researchers used a Markov model, a statistical model that estimates the likelihood of a future event, such as inscribing a particular symbol, based on previously observed patterns. One application uses the statistical model to fill in missing symbols on damaged artifacts, which can increase the pool of data available for deciphering the writings.

View Full Article

Blog: Can Computers Decipher a 5,000-Year-Old Language?

Can Computers Decipher a 5,000-Year-Old Language?
Smithsonian.com (07/20/09) Zax, David

One of the greatest mysteries of the ancient world is the meaning of the Indus civilization's language, and University of Washington, Seattle professor Rajesh Rao is attempting to crack the 5,000-year-old script using computational techniques. He and his colleagues postulated that such methods could reveal whether the Indus script did or did not encode language by measuring the degree of randomness in a sequence, also known as conditional entropy. Rao's team employed a computer program to measure the script's conditional entropy, and then measured the conditional entropy of several natural languages, the artificial Fortran computer programming language, and non-linguistic systems such as DNA sequences. Comparison between these various readings found that the Indus script's rate of conditional entropy bore the closest resemblance to that of the natural languages. Following the publication of the team's findings in the May edition of Science, Rao and colleagues are studying longer strings of characters than they previously examined. "If there are patterns, we could come up with grammatical rules," Rao says. "That would in turn give constraints to what kinds of language families" the Indus script might belong to.

View Full Article

Blog: Twente Researcher Develops Self-Learning Security System for Computer Networks

Twente Researcher Develops Self-Learning Security System for Computer Networks
University of Twente (07/01/09) Bruysters, Joost

University of Twente researcher Damiano Bolzoni has developed SilentDefense, an anomaly network intrusion detection system that could lead to a new generation of network security systems. There are two types of network intrusion detection systems. The first uses a database of all known attacks to identify signatures of commonly used methods, but these systems have difficulty stopping new attack methods. The second uses anomaly detection, essentially learning how the network is normally used and searching for any deviation from the standard pattern. Bolzoni says anomaly detection is not widely used because truly effective systems are not commercially available, but he says SilentDefense will rectify this shortcoming. SilentDefense is based on self-learning algorithms, which significantly improves the accuracy of the system and reduces the odds of false positives. Bolzoni says the ideal network intrusion detection system is not one type or another but a combination of the two. However, before such a system can be created, he says a better anomaly detection system needs to be developed.

View Full Article