Teaching Computer Programming using Deep Recursive Neural Networks

编程是科学、技术、工程、数学和许多其他领域所需的一项关键技能。然而，学习编程是具有挑战性的，学习者在实际练习中有特别的困难。为了支持学生进行这样的实际应用练习，个体化教学将是有帮助的。不幸的是，提供细粒度和个性化的指导超出了人类教师的资源，特别是在拥有数万名参与者的大规模开放在线课程中。因此，我们需要一个自动化的机制来支持学生的实际编程任务。在这个项目中，我们建议通过开发一种新的神经网络架构来解决这一挑战，该架构可以将学生当前的程序编码为向量，预测这个向量应该如何改变，以沿着有能力的学生会采取的路径更接近正确的解决方案，并将改变的向量解码回一个代表学生应该如何进行的程序。通过对比学生当前的计划和他们预测的下一个状态，我们可以生成帮助学生解决任务的提示。在编码和解码中，我们打算将领域知识以手边编程语言的语法形式结合起来。因此，我们将程序的空间缩小到语法正确的空间，从而简化了编码和解码，提高了我们的提示对学生的实用性。此外，我们提出的编码器-解码器架构可以将当前表示学习研究的边界扩展到树状结构数据，例如生物学中的RNA二级结构，化学分子或自然语言的语法树，从而能够将成功的深度学习模型应用于全新类型的数据。

Programming is a key skill required in science, technology, engineering, mathematics, and multiple other fields. However, learning to program is challenging and learners have particular difficulties in practical exercises. To support students in such practical application exercises, individualized instruction would be helpful. Unfortunately, providing fine-grained and individualized instruction exceeds the resources of human instructors, especially in massive open on-line courses with tens of thousands of participants. Therefore, we require an automatized mechanism to support students in practical programming tasks.In this project we propose to address this challenge by developing a novel neural network architecture which can encode a student's current program as a vector, predict how this vector should change to get closer to a correct solution along a path that a capable student would take, and decode the changed vector back into a program that represents how the student should proceed. By contrasting the student's current program with their predicted next state, we can generate hints which help the student to solve the task.In both encoding and decoding, we intend to incorporate domain knowledge in the form of a grammar of the programming language at hand. Thus, we reduce the space of programs to those which are syntactically correct, which simplifies encoding and decoding and enhances the utility of our hints to students. Additionally, our proposed encoder-decoder architecture could extend the boundaries of current representation learning research to tree-structured data, such as RNA secondary structures in biology, chemical molecules, or syntax trees of natural language, thus enabling to apply successful deep learning models to entirely new kinds of data.