本文为美国威斯康星大学麦迪逊分校（作者：Mark W. Craven）的博士论文，共211页。

尽管神经网络已经被用于在许多实际问题领域开发高精度的分类器，但它们所学习的模型却是众所周知难以理解的。本文研究了从训练好的神经网络中提取可理解模型的任务，从而缓解了这一局限性。

本文的主要贡献是提出了一种算法，该算法克服了以往方法的局限性，采用一种新的方法从训练好的网络中提取可理解模型。该算法称为Trepan，将任务视为归纳学习问题。给定一个经过训练的网络或任何其他学习的模型，Trepan使用查询来归纳一个决策树，该决策树近似于模型所表示的函数。与以前在这方面的工作不同，Trepan具有广泛的适用性，并且可以扩展到大型网络和具有高维输入空间的问题。

本文提出了一个实验，通过将其应用于个体网络和在分类、回归和强化学习领域训练的神经网络集合来评估Trepan。这些实验表明，Trepan能够提取出可理解的决策树，同时保持对各自网络的高度保真度。在神经网络比传统决策树算法具有更高预测精度的问题域中，Trepan算法提取的树也表现出更高的预测精度，但在复杂性方面与直接从训练数据中学习的树具有可比性。

本文的第二个贡献是一种称为BBP的算法，它构造性地归纳了简单的神经网络。该算法的基本动机与Trepan相似：学习问题域中的可理解模型，其中神经网络具有特别适当的归纳偏差。BBP算法基于一种假设提升方法，学习关联相对较少的感知器。该算法提供了一个吸引人的优点组合：它为一类相当自然的目标函数提供了可学习性保证；在各种问题领域提供了良好的预测精度；构建了语法上简单的模型，从而促进了人类对它所学知识的理解。这些算法为提高对训练神经网络所学知识的理解提供了机制。

Although neural networks have been used todevelop highly accurate classifiers in numerous real-world problem domains, themodels they learn are notoriously difficult to understand. This thesisinvestigates the task of extracting comprehensible models from trained neuralnetworks, thereby alleviating this limitation. The primary contribution of thethesis is an algorithm that overcomes the significant limitations of previous methodsby taking a novel approach to the task of extracting comprehensible models fromtrained networks. This algorithm, called Trepan, views the task as an inductivelearning problem. Given a trained network, or any other learned model, Trepanuses queries to induce a decision tree that approximates the functionrepresented by the model. Unlike previous work in this area, Trepan is broadlyapplicable as well as scalable to large networks and problems withhigh-dimensional input spaces. The thesis presents experiments that evaluateTrepan by applying it to individual networks and to ensembles of neuralnetworks trained in classification, regression, and reinforcement-learningdomains. These experiments demonstrate that Trepan is able to extract decisiontrees that are comprehensible, yet maintain high levels of fidelity to theirrespective networks. In problem domains in which neural networks providesuperior predictive accuracy to conventional decision tree algorithms, thetrees extracted by Trepan also exhibit superior accuracy, but are comparable interms of complexity, to the trees learned directly from the training data. Asecondary contribution of this thesis is an algorithm, called BBP, thatconstructively induces simple neural networks. The motivation underlying thisalgorithm is similar to that for Trepan: to learn comprehensible models inproblem domains in which neural networks have an especially appropriateinductive bias. The BBP algorithm, which is based on a hypothesis-boostingmethod, learns perceptrons that have relatively few connections. This algorithmprovides an appealing combination of strengths: it provides learnabilityguarantees for a fairly natural class of target functions; it provides goodpredictive accuracy in a variety of problem domains; and it constructssyntactically simple models, thereby facilitating human comprehension of whatit has learned. These algorithms provide mechanisms for improving theunderstanding of what a trained neural network has learned.

1. 引言

2. 项目背景

3. TREPAN算法

4. TREPAN算法的经验评估

5. TREPAN算法的解析评估

6. MofN-sws算法：一种提取M-of-N规则的局部方法

7. 基于Boosting的感知器学习算法

8. 其他相关工作

9. 结论

附录A 用Trepan提取代表树

更多精彩文章请关注公众号：