RIA: Constructive Neural Network Learning Algorithms for Pattern Classification

小行星9409580 本研究的主要目标是系统地探索一类用于模式分类的神经网络的构造性学习算法，该算法可能：1）消除对ad-hoc的、通常不适当的网络结构选择的需要; 2）产生接近于最小的网络，其复杂性与由训练数据隐含指定的特定任务的复杂性相称; 3）提供真实的世界模式分类任务的预期情况复杂度的经验估计; 4）提供自然的方式来权衡一些性能度量（例如，学习时间）相对于其他（例如，网络规模）; 5）提供方便的方法来整合领域知识（当可用时）以加速学习。

9409580 Honavar The primary objectives of the proposed research is a systematic exploration of a class of constructive learning algorithms for neural networks for pattern classification that can potentially: 1) Eliminate the need for ad-hoc, often inappropriate choices of network architecture; 2) Yield close to minimal networks whose complexity is commensurate with the complexity of the particular task implicitly specified by the training data; 3) Provide empirical estimates of expected-case complexity of real- world pattern classification tasks; 4) Offer natural ways to trade off some performance measures (e.g., learning time) against others (e.g., network size); 5) Provide convenient ways to incorporate domain knowledge (when available) to speed up learning.