Elad Yom-Tov

An improved P300-based brain-computer interface

A brain-computer interface (BCI) is a system for direct communication between brain and computer. The BCI developed in this work is based on a BCI described by Farwell and Donchin in 1988, which allows a subject to communicate one of 36 symbols presented on a 6 /spl times/ 6 matrix. The system exploits the P300 component of event-related brain potentials (ERP) as a medium for communication. The processing methods distinguish this work from Donchins work. In this work, independent component analysis (ICA) was used to separate the P300 source from the background noise. A matched filter was used together with averaging and threshold techniques for detecting the existence of P300s. The processing method was evaluated offline on data recorded from six healthy subjects. The method achieved a communication rate of 5.45 symbols/min with an accuracy of 92.1% compared to 4.8 symbols/min with an accuracy of 90% in Donchins work. The online interface was tested with the same six subjects. The average communication rate achieved was 4.5 symbols/min with an accuracy of 79.5% as apposed to the 4.8 symbols/min with an accuracy of 56% in Donchins work. The presented BCI achieves excellent performance compared to other existing BCIs, and allows a reasonable communication rate, while maintaining a low error rate.

Learning to estimate query difficulty: including applications to missing content detection and distributed information retrieval

In this article we present novel learning methods for estimating the quality of results returned by a search engine in response to a query. Estimation is based on the agreement between the top results of the full query and the top results of its sub-queries. We demonstrate the usefulness of quality estimation for several applications, among them improvement of retrieval, detecting queries for which no relevant content exists in the document collection, and distributed information retrieval. Experiments on TREC data demonstrate the robustness and the effectiveness of our learning algorithms.

What makes a query difficult

This work tries to answer the question of what makes a query difficult. It addresses a novel model that captures the main components of a topic and the relationship between those components and topic difficulty. The three components of a topic are the textual expression describing the information need (the query or queries), the set of documents relevant to the topic (the Qrels), and the entire collection of documents. We show experimentally that topic difficulty strongly depends on the distances between these components. In the absence of knowledge about one of the model components, the model is still useful by approximating the missing component based on the other components. We demonstrate the applicability of the difficulty model for several uses such as predicting query difficulty, predicting the number of topic aspects expected to be covered by the search results, and analyzing the findability of a specific domain.

Milepost GCC: Machine Learning Enabled Self-tuning Compiler

Tuning compiler optimizations for rapidly evolving hardware makes porting and extending an optimizing compiler for each new platform extremely challenging. Iterative optimization is a popular approach to adapting programs to a new architecture automatically using feedback-directed compilation. However, the large number of evaluations required for each program has prevented iterative compilation from widespread take-up in production compilers. Machine learning has been proposed to tune optimizations across programs systematically but is currently limited to a few transformations, long training phases and critically lacks publicly released, stable tools. Our approach is to develop a modular, extensible, self-tuning optimization infrastructure to automatically learn the best optimizations across multiple programs and architectures based on the correlation between program features, run-time behavior and optimizations. In this paper we describe Milepost GCC, the first publicly-available open-source machine learning-based compiler. It consists of an Interactive Compilation Interface (ICI) and plugins to extract program features and exchange optimization data with the cTuning.org open public repository. It automatically adapts the internal optimization heuristic at function-level granularity to improve execution time, code size and compilation time of a new program on a given architecture. Part of the MILEPOST technology together with low-level ICI-inspired plugin framework is now included in the mainline GCC. We developed machine learning plugins based on probabilistic and transductive approaches to predict good combinations of optimizations. Our preliminary experimental results show that it is possible to automatically reduce the execution time of individual MiBench programs, some by more than a factor of 2, while also improving compilation time and code size. On average we are able to reduce the execution time of the MiBench benchmark suite by 11% for the ARC reconfigurable processor. We also present a realistic multi-objective optimization scenario for Berkeley DB library using Milepost GCC and improve execution time by approximately 17%, while reducing compilation time and code size by 12% and 7% respectively on Intel Xeon processor.

Models of user engagement

Our research goal is to provide a better understanding of how users engage with online services, and how to measure this engagement. We should not speak of one main approach to measure user engagement --- e.g. through one fixed set of metrics --- because engagement depends on the online services at hand. Instead, we should be talking of models of user engagement. As a first step, we analysed a number of online services, and show that it is possible to derive effectively simple models of user engagement, for example, accounting for user types and temporal aspects. This paper provides initial insights into engagement patterns, allowing for a better understanding of the important characteristics of how users repeatedly interact with a service or group of services.

Classification of finger activation for use in a robotic prosthesis arm

Hand amputees would highly benefit from a robotic prosthesis, which would allow the movement of a number of fingers. In this paper we propose using the electromyographic signals recorded by two pairs of electrodes placed over the arm for operating such prosthesis. Multiple features from these signals are extracted whence the most relevant features are selected by a genetic algorithm as inputs for a simple classifier. This method results in a probability of error of less than 2%.

Feature selection for the classification of movements from single movement-related potentials

Classification of movement-related potentials recorded from the scalp to their corresponding limb is a crucial task in brain-computer interfaces based on such potentials. Many features can be extracted from raw electroencephalographic signals to be used for classification, but the utilization of irrelevant or superfluous features is detrimental to the performance of classification algorithms. It is, therefore, necessary to select a small number of relevant features for the classification task. This paper demonstrates the use of two feature selection methods to choose a small number (10-20) of relevant features from a bank containing upward of 1000 features. One method is based on information theory and the other on the use of genetic algorithms. We show that the former is poorly suited for the aforementioned classification task and discuss the probable reasons for this. However, using a genetic algorithm on data recorded from five subjects we demonstrate that it is possible to differentiate between the movements of two limbs with a classification accuracy of 87% using as little as 10 features without subject training. With the addition of a simple coding scheme, this method can be applied to multiple limb classification and a 63% classification accuracy rate can be reached when attempting to distinguish between three limbs.

Postmarket Drug Surveillance Without Trial Costs: Discovery of Adverse Drug Reactions Through Large-Scale Analysis of Web Search Queries

Background Postmarket drug safety surveillance largely depends on spontaneous reports by patients and health care providers; hence, less common adverse drug reactions—especially those caused by long-term exposure, multidrug treatments, or those specific to special populations—often elude discovery. Objective Here we propose a low cost, fully automated method for continuous monitoring of adverse drug reactions in single drugs and in combinations thereof, and demonstrate the discovery of heretofore-unknown ones. Methods We used aggregated search data of large populations of Internet users to extract information related to drugs and adverse reactions to them, and correlated these data over time. We further extended our method to identify adverse reactions to combinations of drugs. Results We validated our method by showing high correlations of our findings with known adverse drug reactions (ADRs). However, although acute early-onset drug reactions are more likely to be reported to regulatory agencies, we show that less acute later-onset ones are better captured in Web search queries. Conclusions Our method is advantageous in identifying previously unknown adverse drug reactions. These ADRs should be considered as candidates for further scrutiny by medical regulatory authorities, for example, through phase 4 trials.

Detection of movement-related potentials from the electro-encephalogram for possible use in a brain-computer interface

Brain-computer interfaces are devices for enabling patients with severe motor disorders to communicate with the world. One method for operating such devices is to use movement-related potentials that are generated in the brain when the patient moves, or imagines a movement of, one of his limbs. An algorithm for detecting movement-related potentials using a small number of EEG channels was developed. This algorithm is a combination of the matched filter, a non-linear transformation previously developed as part of a similar detector, and a classifier. The algorithm was compared with previous designs of similar detectors by both theoretic analysis and off-line evaluation of performance on data recorded from five subjects. It is shown that the performance of the algorithm was superior to that of previous methods, improving the area under the receiver operating characteristic curve to 87.8%, an improvement of 25% compared with the best previously suggested detection method. Finally, the probable sources for false detections were identified, and possible ways to minimise them are proposed.

Pro-Anorexia and Pro-Recovery Photo Sharing: A Tale of Two Warring Tribes

Background There is widespread use of the Internet to promote anorexia as a lifestyle choice. Pro-anorexia content can be harmful for people affected or at risk of having anorexia. That movement is actively engaged in sharing photos on social networks such as Flickr. Objective To study the characteristics of the online communities engaged in disseminating content that encourages eating disorders (known as “pro-anorexia”) and to investigate if the posting of such content is discouraged by the posting of recovery-oriented content. Methods The extraction of pro-anorexia and pro-recovery photographs from the photo sharing site Flickr pertaining to 242,710 photos from 491 users and analyzing four separate social networks therein. Results Pro-anorexia and pro-recovery communities interact to a much higher degree among themselves than what is expected from the distribution of contacts (only 59-72% of contacts but 74-83% of comments are made to members inside the community). Pro-recovery users employ similar words to those used by pro-anorexia users to describe their photographs, possibly in order to ensure that their content appears when pro-anorexia users search for images. Pro-anorexia users who are exposed to comments from the opposite camp are less likely to cease posting pro-anorexia photographs than those who do not receive such comments (46% versus 61%), and if they cease, they do so approximately three months later. Our observations show two highly active communities, where most interaction is within each community. However, the pro-recovery community takes steps to ensure that their content is visible to the pro-anorexia community, both by using textual descriptions of their photographs that are similar to those used by the pro-anorexia group and by commenting to pro-anorexia content. The latter activity is, however, counterproductive, as it entrenches pro-anorexia users in their stance. Conclusions Our results highlight the nature of pro-anorexia and pro-recovery photo sharing and accentuate the need for clinicians to be aware of such content and its effect on their patients. Our findings suggest that some currently used interventions are not useful in helping pro-anorexia users recover. Thus, future work should focus on new intervention methods, possibly tailored to individual characteristics.