Explanation-Based Neural Network Learning

这项研究试图将机器学习的两种主要范式：归纳学习和分析学习结合起来。基于实例和神经网络学习等归纳方法可以可靠地从噪声数据中学习简单的函数，但需要大量的训练样本才能将其扩展到非常复杂的函数。相比之下，分析性方法，如基于解释的学习，可以从更少的数据中学习复杂的函数，但依赖于学习者强大的先验知识。目前机器学习中的许多研究都试图结合这两种方法的优点，以获得从近似的先验知识和观察数据中学习更正确的概括的方法。提出的研究采用了一种新的方法来解决这个问题：统一神经网络学习和基于解释的学习。更具体地说，这项研究将建立在最近开发的基于解释的神经网络(EBNN)学习方法的基础上。初步研究表明，如果有准确的领域知识，EBNN可以从较少的例子中获得比纯归纳学习更好的泛化能力，并且随着学习者的先验知识质量的提高，EBNN会优雅地退化。这项研究将更充分地探索神经网络和基于解释的方法相结合的空间，重点关注诸如扩大到更复杂的学习任务，可以从基于神经网络的解释中提取的替代类型的信息，在从非常强到非常弱的先验知识的整个频谱上的稳健操作，以及域理论和目标函数的替代表示。EBNN学习将应用于两个不同的任务领域。如果成功，这项研究可能会产生更多实际问题的学习方法，并导致对符号方法和神经网络方法之间的对应关系有更清晰的理解。

This research seeks to combine the two primary paradigms for machine learning: inductive and analytical learning. Inductive methods such as instance-based and neural network learning can reliably learn simple functions from noisy data, but require vast numbers of training examples in order to scale up to very complex functions. In contrast, analytical methods such as explanation-based learning can learn complex functions from much less data, but rely upon strong prior knowledge on the part of the learner. Much current research in machine learning seeks to combine the best of both approaches, to obtain methods that learn more correct generalizations from approximate prior knowledge together with observed data. The proposed research takes a novel approach to this problem: unifying neural network learning and explanation- based learning. More specifically, this research will build on the recently developed explanation-based neural network (EBNN) learning method. Preliminary research has demonstrated experimentally that EBNN can generalize better from fewer examples than pure inductive learning if accurate domain knowledge is available, and that it degrades gracefully with the quality of the learner's prior knowledge. This research will explore more fully the space of combined neural net and explanation-based methods, focusing on issues such as scaling up to more complex learning tasks, alternative types of information that can be extracted from explanations based on neural networks, operating robustly over the entire spectrum from very strong to very weak prior knowledge, and alternative representations for the domain theory and target function. EBNN learning will be applied to two different task domains. If successful, this research could produce learning methods that scale up to more practical problems, and lead to a clearer understanding of the correspondence between symbolic and neural network approaches.