ABSTRACT. Sometime during the Summer of 2011, several of us at UC Davis were quite puzzled and shocked
to discover that software is "natural", viz., just as repetitive and predictable as
natural language corpora; In fact much more so!  By now, this early experiment has been
replicated many times, in many ways, and various applications  of naturalness have been
developed. But why is this? Is it just because of programming language  syntax? or is it due to
something else,  like conscious programmer  choice? How can we study this question?
Are there are other  "natural" corpora (other than software)  that  are similar to software?

ABSTRACT. The area of deep learning has had many remarkable successes in the
recent past. However, deep neural networks have some well-documented
weaknesses: they are data-hungry, brittle, opaque to users, hard to
analyze, and not easily constrained with axiomatic knowledge about the
world. In this talk, I will argue that ideas from programming
languages and program synthesis can offer a way of overcoming these
weaknesses. Specifically, one can use higher-level programming
languages, possibly with access to neural "subroutines", to describe
statistical models normally described neurally. Such programmatic
models are more easily understood by humans, can more easily
accommodate axiomatic knowledge about the world, and are easier to
analyze using symbolic techniques. The use of "neurosymbolic", as
opposed to purely symbolic, models can lead to lower complexity of
learning. The compositionality inherent in modern languages can allow
transfer of knowledge across learning tasks. Discovering such
programmatic models from data is a form of program synthesis, and can
perhaps also be described as "high-level machine learning". Early
experience with the problem suggests that the literature on
language-integrated program synthesis can offer powerful tools for
solving this problem. At the same time, the problem is different from
traditional synthesis in key ways, and opens up many new technical
challenges.

ABSTRACT. The rich structure of source code presents an interesting challenge for machine learning methods. Recently, Graph Neural Networks (GNN) have shown promising results in code understanding and code generation tasks. In this talk, I will briefly discuss two neural models that employ GNNs: one of them for catching variable misuse bugs and the other for generating code expressions. Finally, I will discuss some of the open challenges that GNNs face on many source code-related tasks.

ABSTRACT. In this talk, I will address the questions of

1. how we specify arbitrary tasks to a learning system
2. how we interpret its behaviour, and finally
3. how do we verify or debug it to ensure that its behaviour is consistent with the task specification.

I will also describe my initial attempts to make progress on these questions through program synthesis and verification.

ABSTRACT. Building effective program analysis tools is a challenging endeavor: analysis designers must balance multiple competing objectives, including scalability, fraction of false alarms, and the possibility of missed bugs. Not all of these design decisions are optimal when the analysis is applied to a new program with different coding idioms, environment assumptions, and quality requirements. Furthermore, the alarms produced are typically accompanied by limited information such as their location and abstract counter-examples. We present a framework Difflog that fundamentally extends the deductive reasoning rules that underlie program analyses with numerical weights. Each alarm is now naturally accompanied by a score, indicating quantities such as the confidence that the alarm is a real bug, the anticipated severity, or expected relevance of the alarm to the programmer. To the analysis user, these techniques offer a lens by which to focus their attention on the most important alarms and a uniform method for the tool to interactively generalize from human feedback. To the analysis designer, these techniques offer novel ways to automatically synthesize analysis rules in a data-driven style. Difflog shows large reductions in false alarm rates and missed bugs in large, complex programs, and it advances the state-of-the-art in synthesizing non-trivial analyses.

ABSTRACT. A major challenge when using techniques from Natural Language Processing for supervised learning on computer program source code is that many words in code are neologisms. Reasoning over such an unbounded vocabulary is not something NLP methods are typically suited for. We introduce a deep model that contends with an unbounded vocabulary (at training or test time) by embedding new words as nodes in a graph as they are encountered and processing the graph with a Graph Neural Network.

ABSTRACT. Programmers make rich use of natural language in the source code they write through identifiers and comments. Source code identifiers are selected from a pool of tokens which are strongly related to the meaning, naming conventions, and context. These tokens are often combined to produce more precise and obvious designations. Such multi-part identifiers count for 97% of all naming tokens in the Public Git Archive - the largest dataset of Git repositories to date. We introduce a bidirectional LSTM recurrent neural network to detect subtokens in source code identifiers. We trained that network on 41.7 million distinct splittable identifiers collected from 182,014 open source projects in Public Git Archive, and show that it outperforms several other machine learning models. The proposed network can be used to improve the upstream models which are based on source code identifiers, as well as improving developer experience allowing writing code without switching the keyboard case.