Next paper candidates: 9 Aug

[m09]

Next paper candidates

Let's propose papers to study next! All papers mentioned in the comments of this issue will be listed in the next vote.

Last session runner-up(s)

Attention is All you Need

The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data.

Getafix: Learning to Fix Bugs Automatically

Static analyzers help find bugs early by warning about recurring bug categories. While fixing these bugs still remains a mostly manual task in practice, we observe that fixes for a specific bug category often are repetitive. This paper addresses the problem of automatically fixing instances of common bugs by learning from past fixes. We present Getafix, an approach that produces human-like fixes while being fast enough to suggest fixes in time proportional to the amount of time needed to obtain static analysis results in the first place. Getafix is based on a novel hierarchical clustering algorithm that summarizes fix patterns into a hierarchy ranging from general to specific patterns. Instead of an expensive exploration of a potentially large space of candidate fixes, Getafix uses a simple yet effective ranking technique that uses the context of a code change to select the most appropriate fix for a given bug. Our evaluation applies Getafix to 1,268 bug fixes for six bug categories reported by popular static analyzers for Java, including null dereferences, incorrect API calls, and misuses of particular language constructs. The approach predicts exactly the human-written fix as the top-most suggestion between 12% and 91% of the time, depending on the bug category. The top-5 suggestions contain fixes for 526 of the 1,268 bugs. Moreover, we report on deploying the approach within Facebook, where it contributes to the reliability of software used by billions of people. To the best of our knowledge, Getafix is the first industrially-deployed automated bug-fixing tool that learns fix patterns from past, human-written fixes to produce human-like fixes.

[bzz]

Revisiting Graph Neural Networks: All We Have is Low-Pass Filters

Graph neural networks have become one of the most important techniques to solve machine learning problems on graph-structured data. Recent work on vertex classification proposed deep and distributed learning models to achieve high performance and scalability. However, we find that the feature vectors of benchmark datasets are already quite informative for the classification task, and the graph structure only provides a means to denoise the data. In this paper, we develop a theoretical framework based on graph signal processing for analyzing graph neural networks. Our results indicate that graph neural networks only perform low-pass filtering on feature vectors and do not have the non-linear manifold learning property. We further investigate their resilience to feature noise and propose some insights on GCN-based graph neural network design.

[egor-bogomolov]

Topology Adaptive Graph Convolutional Networks, 2017

Spectral graph convolutional neural networks (CNNs) require approximation to the convolution to alleviate the computational complexity, resulting in performance loss. This paper proposes the topology adaptive graph convolutional network (TAGCN), a novel graph convolutional network defined in the vertex domain. We provide a systematic way to design a set of fixed-size learnable filters to perform convolutions on graphs. The topologies of these filters are adaptive to the topology of the graph when they scan the graph to perform convolution. The TAGCN not only inherits the properties of convolutions in CNN for grid-structured data, but it is also consistent with convolution as defined in graph signal processing. Since no approximation to the convolution is needed, TAGCN exhibits better performance than existing spectral CNNs on a number of data sets and is also computationally simpler than other recent methods.

[m09]

Winner is Attention Is All You Need :tada: