Learning Theory (LT)


Monday, March 21st, 16h30-18h10

Paper ID   Title
LT-1 Evolution versus Learning in Temporal Neural Networks
LT-2 Minimization of Empirical Error over Perceptron Networks
LT-3 Interval Basis Neural Networks
LT-4 Learning from Randomly-Distributed Inaccurate Measurements
LT-5 Combining topological and cardinal directional relation information in QSR
LT-6 An Evidence Theoretic Ensemble Design Technique

Title: Evolution versus Learning in Temporal Neural Networks
Author(s): Hedi Soula,
Guillaume Beslon,
Joel Favrel
In this paper, we study the difference between two ways of setting synaptic weights in a "temporal" neural network. Used as a controller of a simulated mobile robot, the neural network is alternatively evolved through an evolutionary algorithm or trained via an hebbian reinforcement learning rule. We compare both approaches and argue that in the last instance only the learning paradigm is able to exploit meaningfully the temporal features of the neural network.

Title: Minimization of Empirical Error over Perceptron Networks
Author(s): Vera Kurkova
Abstract: Supervised learning by perceptron networks is investigated as an approximate minimization of empirical error functional. Rates of convergence of suboptimal solutions obtainable using networks with n hidden units to optimal solutions that require the same number of hidden units as the size of the training set are derived in terms of a variational norm. It is shown that fast rates can be guaranteed when the data defining the empirical error can be interpolated by a function with a Sobolev-type norm which does not grow exponentially with the input dimension d.

Title: Interval Basis Neural Networks
Author(s): A. Horzyk
Abstract: The paper introduces a new type of ontogenic neural networks called Interval Basis Neural Networks (IBNN). IBNNs configure the whole topology and compute weights after the a priori knowledge collected form training data. After the statistical analysis the training data of the same class are grouped by intervals separately for all input features. This IBNNs feature make possible to computed all network parameters without training. Moreover the IBNN take into account the distances between patterns of the same classes and builds the well-approximating model especially on the borders between the classes. Futhermore the IBNNs are insensitive for differences in quantity of patterns represented classes. The IBNNs always correctly classify training data and very good generalize other data.

Title: Learning from Randomly-Distributed Inaccurate Measurements
Author(s): John Eidson,
Bruce Hamilton,
Valery Kanevsky
Abstract: Traditional measurement systems are designed with tight control over the time and place of measurement of the device or envi-ronment under test. This is true whether the measurement sys-tem uses a centralized or a distributed architecture. Currently there is considerable interest in using mobile consumer devices as measurement platforms for testing large dispersed systems. There is also growing activity in developing concepts of ubiqui-tous measurement, such as “smart dust”. Under these conditions the times and places of measurement are random, which raises the question of the validity and interpretation of the acquired data. This paper presents a mathematical analysis that shows it is possible under certain conditions to establish dependence be-tween error bounds and confidence probability on models built using data acquired in this manner.

Title: Combining topological and cardinal directional relation information in QSR
Author(s): Haibin Sun
Abstract: Combining different knowledge representation languages is one of the main topics in Qualitative Spatial Reasoning (QSR). In this paper, we combine well-known RCC8 calculus (RCC8) and cardinal direction calculus (CDC) based on regions and give the interaction tables for the two calculi. The interaction tables can be used as a tool in solving constraint satisfaction problems (CSP) and consistency checking procedure of QSR for combined spatial knowledge.

Title: An Evidence Theoretic Ensemble Design Technique
Author(s): Hakan Altincay
Abstract: Ensemble design techniques based on resampling the training set are successfully used to improve the classification accuracies of the base classifiers. In Boosting technique, each training set is obtained by drawing samples with replacement from the available training set according to a weighted distribution which is iteratively updated for generating new classifiers for the ensemble. The resultant classifiers are accurate in different parts of the input space mainly specified the sample weights. In this study, a dynamic integration of boosting based ensembles is proposed so as to take into account the heterogeneity of the input sets. In this approach, a Dempster-Shafer theory based framework is developed to consider the training sample distribution in the restricted input space of each test sample. The effectiveness of the proposed technique is compared to AdaBoost algorithm using nearest mean type base classifier.