Epistatic interactions between genes are believed to be a critical component in the genetic architecture of complex diseases. Genome Wide Association Studies (GWAS) may be able to detect such genetic interactions indirectly, via the identification of associated SNP markers. Major obstacles to progress in this area are: the unknown nature of epistatic interactions, little understanding of the capabilities of different filtering methods, and the computational difficulties for exhaustive analysis. A common platform enabling various detection methods is needed to avoid practical issues such as software compatibility and portability, incompatible input and output formats and varying demands on computational resources. We developed a highly optimised GPU system capable of exhaustively analysing all SNP-pairs in typical GWAS data (0.5M SNPs, 5K samples) in a few minutes on a standard desktop computer. A number of programming elements provided by a functional interface can be used to construct user-defined statistical tests to efficiently score every SNP pair. As a proof of principle, we have implemented 8 methods from the literature via our interface. We have applied all of them using a single GPU to exhaustively scan the 7 popular WTCCC case-control GWAS datasets. We present timing results for these methods, both in their original software implementations and using our platform. Significant improvements in timing are observed, up to 10000 times for CPU implementations of the popular FastEpistasis in PLINK and up to 2 orders of magnitude for some GPU implementations in the literature. As an initial discovery we show plots for overlaps of list of selected pairs by 8 algorithms for Type 2 Diabetes, WTCCC data.

In this paper, we present a new co-training strategy that makes use of unlabelled data. It trains two predictors in parallel, with each predictor labelling the unlabelled data for training the other predictor in the next round. Both predictors are support vector machines, one trained using data from the original feature space, the other trained with new features that are derived by clustering both the labelled and unlabelled data. Hence, unlike standard co-training methods, our method does not require a priori the existence of two redundant views either of which can be used for classification, nor is it dependent on the availability of two different supervised learning algorithms that complement each other.We evaluated our method with two classifiers and three text benchmarks: WebKB, Reuters newswire articles and 20 NewsGroups. Our evaluation shows that our co-training technique improves text classification accuracy especially when the number of labelled examples are very few.

We demonstrate the practical ability of neural networks (NNs) trained in a supervised mode to extract useful control "knowledge" from a large, high-dimensional empirical database, and then to deliver almost optimal control in "real time". In particular, this paper describes experiments with NN-based controllers for allocating bandwidth capacity in a telecommunications network (SDH). This system was proposed in order to overcome a "real time" response constraint. Two basic architectures, each consisting of a combination of two methods, are evaluated: (1) a feedforward network-heuristic combination and (2) a feedforward network-recurrent network combination. These architectures are compared against a linear programming (LP) optimizer as a benchmark. This LP optimizer was also used as a teacher to label the data samples for the feedforward NN training algorithm. NN-based solutions are very accurate (/spl sim/98% of optimal throughput) and, in contrast to the algorithmic approach, can be delivered in "real time". It is found that while the "human" generated heuristics (greedy search optimization) fail to find a solution in approximately 30% of cases, the best NN fails only in 4.9% of cases. Moreover, it has been found that in spite of the very high dimensionality of the problem (55 inputs and 126 outputs), the solution can be delivered by surprisingly compact NNs, with as little as around 1000 synaptic weights. This proves that on this occasion the NNs were able to extract simple but powerful "heuristics" hidden in the complex sets of numerical data.

We discuss a model of consistent learning with an additional re(cid:173) striction on the probability distribution of training samples, the target concept and hypothesis class. We show that the model pro(cid:173) vides a significant improvement on the upper bounds of sample complexity, i.e. the minimal number of random training samples allowing a selection of the hypothesis with a predefined accuracy and confidence. Further, we show that the model has the poten(cid:173) tial for providing a finite sample complexity even in the case of infinite VC-dimension as well as for a sample complexity below VC-dimension. This is achieved by linking sample complexity to an "average" number of implement able dichotomies of a training sample rather than the maximal size of a shattered sample, i.e. VC-dimension.

Introduces a class of simple polynomial neural network classifiers, called mask perceptrons. A series of algorithms for practical development of such structures is outlined. It relies on ordering of input attributes with respect to their potential usefulness and heuristic driven generation and selection of hidden units (monomial terms) in order to combat the exponential explosion in the number of higher-order monomial terms to choose from. Results of tests for two popular machine learning benchmarking domains (mushroom classification and faulty LED-display), and for two nonstandard domains (spoken digit recognition and article category determination) are given. All results are compared against a number of other classifiers. A procedure for converting a mask perceptron to a classical logic production rule is outlined and shown to produce a number of 100% percent accurate simple rules after training on 6-20% of a database.