Barry Boehm

Iordanis Fostiropoulos, Barry Boehm

Quantization has been applied to multiple domains in Deep Neural Networks (DNNs). We propose Depthwise Quantization (DQ) where $\textit{quantization}$ is applied to a decomposed sub-tensor along the $\textit{feature axis}$ of weak statistical dependence. The feature decomposition leads to an exponential increase in $\textit{representation capacity}$ with a linear increase in memory and parameter cost. In addition, DQ can be directly applied to existing encoder-decoder frameworks without modification of the DNN architecture. We use DQ in the context of Hierarchical Auto-Encoder and train end-to-end on an image feature representation. We provide an analysis on cross-correlation between spatial and channel features and we propose a decomposition of the image feature representation along the channel axis. The improved performance of the depthwise operator is due to the increased representation capacity from implicit feature decoupling. We evaluate DQ on the likelihood estimation task, where it outperforms the previous state-of-the-art on CIFAR-10, ImageNet-32 and ImageNet-64. We progressively train with increasing image size a single hierarchical model that uses 69% less parameters and has a faster convergence than the previous works.

Daniel Link, Pooyan Behnamghader, Ramin Moazeni et al.

The stakeholders of a system are legitimately interested in whether and how its architecture reflects their respective concerns at each point of its development and maintenance processes. Having such knowledge available at all times would enable them to continually adjust their systems structure at each juncture and reduce the buildup of technical debt that can be hard to reduce once it has persisted over many iterations. Unfortunately, software systems often lack reliable and current documentation about their architecture. In order to remedy this situation, researchers have conceived a number of architectural recovery methods, some of them concern-oriented. However, the design choices forming the bases of most existing recovery methods make it so none of them have a complete set of desirable qualities for the purpose stated above. Tailoring a recovery to a system is either not possible or only through iterative experiments with numeric parameters. Furthermore, limitations in their scalability make it prohibitive to apply the existing techniques to large systems. Finally, since several current recovery methods employ non-deterministic sampling, their inconsistent results do not lend themselves well to tracking a systems course over several versions, as needed by its stakeholders. RELAX (RELiable Architecture EXtraction), a new concern-based recovery method that uses text classification, addresses these issues efficiently by (1) assembling the overall recovery result from smaller, independent parts, (2) basing it on an algorithm with linear time complexity and (3) being tailorable to the recovery of a single system or a sequence thereof through the selection of meaningfully named, semantic topics. An intuitive, informative architectural visualization rounds out RELAX's contributions. RELAX is illustrated on a number of existing open-source systems and compared to other recovery methods.

Junyan Cheng, Iordanis Fostiropoulos, Barry Boehm

Transformer architectures have been successfully used in learning source code representations. The fusion between a graph representation like Abstract Syntax Tree (AST) and a source code sequence makes the use of current approaches computationally intractable for large input sequence lengths. Source code can have long-range dependencies that require larger sequence lengths to model effectively. Current approaches have a quadratic growth in computational and memory costs with respect to the sequence length. Using such models in practical scenarios is difficult. In this work, we propose the conditioning of a source code snippet with its graph modality by using the graph adjacency matrix as an attention mask for a sparse self-attention mechanism and the use of a graph diffusion mechanism to model longer-range token dependencies. Our model reaches state-of-the-art results in BLEU, METEOR, and ROUGE-L metrics for the code summarization task and near state-of-the-art accuracy in the variable misuse task. The memory use and inference time of our model have linear growth with respect to the input sequence length as compared to the quadratic growth of previous works.

Junyan Cheng, Iordanis Fostiropoulos, Barry Boehm

As opposed to natural languages, source code understanding is influenced by grammatical relationships between tokens regardless of their identifier name. Graph representations of source code such as Abstract Syntax Tree (AST) can capture relationships between tokens that are not obvious from the source code. We propose a novel method, GN-Transformer to learn end-to-end on a fused sequence and graph modality we call Syntax-Code-Graph (SCG). GN-Transformer expands on Graph Networks (GN) framework using a self-attention mechanism. SCG is the result of the early fusion between a source code snippet and the AST representation. We perform experiments on the structure of SCG, an ablation study on the model design, and the hyper-parameters to conclude that the performance advantage is from the fused representation. The proposed methods achieve state-of-the-art performance in two code summarization datasets and across three automatic code summarization metrics (BLEU, METEOR, ROUGE-L). We further evaluate the human perceived quality of our model and previous work with an expert-user study. Our model outperforms the state-of-the-art in human perceived quality and accuracy.

Daniel Link, Pooyan Behnam, Ramin Moazeni et al.

In order to maintain a system, it is beneficial to know its software architecture. In the common case that this architecture is unavailable, architecture recovery provides a way to recover an architectural view of the system. Many different methods and tools exist to provide such a view. While there have been taxonomies of different recovery methods and surveys of their results along with measurements of how these results conform to expert's opinions on the systems, there has not been a survey that goes beyond an automatic comparison and instead seeks to answer questions about the viability of individual methods in given situations, the quality of their results and whether these results can be used to indicate and measure the quality and quantity of architectural changes. For our case study, we look at the results of recoveries of versions of Android and Apache Hadoop obtained by running PKG, ACDC and ARC.

Tim Menzies, Ye Yang, George Mathew et al.

Context:More than half the literature on software effort estimation (SEE) focuses on comparisons of new estimation methods. Surprisingly, there are no studies comparing state of the art latest methods with decades-old approaches. Objective:To check if new SEE methods generated better estimates than older methods. Method: Firstly, collect effort estimation methods ranging from "classical" COCOMO (parametric estimation over a pre-determined set of attributes) to "modern" (reasoning via analogy using spectral-based clustering plus instance and feature selection, and a recent "baseline method" proposed in ACM Transactions on Software Engineering).Secondly, catalog the list of objections that lead to the development of post-COCOMO estimation methods.Thirdly, characterize each of those objections as a comparison between newer and older estimation methods.Fourthly, using four COCOMO-style data sets (from 1991, 2000, 2005, 2010) and run those comparisons experiments.Fifthly, compare the performance of the different estimators using a Scott-Knott procedure using (i) the A12 effect size to rule out "small" differences and (ii) a 99% confident bootstrap procedure to check for statistically different groupings of treatments). Results: The major negative results of this paper are that for the COCOMO data sets, nothing we studied did any better than Boehm's original procedure. Conclusions: When COCOMO-style attributes are available, we strongly recommend (i) using that data and (ii) use COCOMO to generate predictions. We say this since the experiments of this paper show that, at least for effort estimation,how data is collected is more important than what learner is applied to that data.