Many apologizes for cross-posting
SCI2002 Sixth World Multiconference on Systemics, Cybernetics and Informatics July 14-18, 2002 ~ Orlando, Florida Special Session: Learning with confidence http://laren.dsi.unimi.it/SCI2002/ Call for papers Leaving the asymptotic learnability results of early sixties, for instance from E. Gold or A. Gill, modern theories consider learning as a statistical operation, possibly based on highly structured sample values, possibly done in a very poor probabilistic framework. In this scenario the target of our learning task is generally a function that is a random object, and we want to frame its variability within a set of possible realizations with satisfactory confidence. Under a computational perspective this problem reads in terms of sample complexity for a given accuracy (a relevant measure of the width of the realization set) and In the aim of locating the learning task in the one or other side of the exponential complexity divide, former results came from rather elementary probabilistic modeling based on binomial experiments and sharp bounds such as those coming from Chernoff inequality. Subsequent comparisons of the algorithms efficiency on a same learning task lead to the employment of more sophisticated statistical tools to identify very accurate confidence intervals, in relation with both sample properties - such as their distribution law or error rate - and structural constraints - such as the allowed complexity of the statistics. These theoretical improvements allow, for instance, to distinguish between different degrees of the polynomials describing sample complexities of algorithms for learning a monotone DNF under proper probability hypotheses on the example space. Many efforts have also been devoted to the confidence intervals for the shape of continuous functions, with results concerning trained neural networks as well.The session aims at collecting contributions by researchers involved in these topics. The special perspective is the exploitation of relations between the randomness of the training examples and their mutual dependence exactly denoted by the function we want discovering from them. Submissions A 2000 characters abstract should be submitted in electronic format (preferably in PDF, but PostScript or MS Word are also acceptable formats) to [EMAIL PROTECTED] within February 23, 2002, using as subject-line "SCI2002 Special session submission". After notification of acceptance the authors will have to submit within April 5, 2002 an extended abstract not exceeding the length of six pages. Please do not send your papers to SCI2002 secretariat. All papers must be presented by one of the authors, who must pay the registration fee. For more information about the general conference please see http://www.iiisci.org/sci2002/. Session Chair Bruno Apolloni Dipartimento di Scienze dell'Informazione Universita' degli Studi di Milano Via Complico 39/41, I-20153 Milano - Italy Phone: +39 02 503 16284 Fax: +39 02 503 16288 E-mail: [EMAIL PROTECTED] confidence.
