Contains fulltext : 318879.pdf (Publisher’s version ) (Open Access)

For software maintenance and evolution, insight into the codebase is crucial. One way to enhance insight is the application of static analysis to extract and visualize program-specific relations from the code itself, such as call graphs and inheritance trees. Yet, software often contains in-house libraries: unique, domain-specific libraries whose usage is typically scattered throughout the codebase. To provide sufficient insight into the usage of those libraries, the static analysis must be customized with domain-specific information. In this paper, we propose a method to enhance insight into the usage of in-house libraries by producing custom overviews. Furthermore, we describe three exploratory case studies targeting industrial C++ and Ada codebases, in which the method was developed, evolved, and validated. The method prescribes how to create custom overviews using static analysis iteratively, starting from a user-provided, initial specification of proper library usage using code patterns. As a safeguard, the method includes cross-checks to detect code fragments that deviate from proper library usage. Whenever such a deviating library usage is found, the code owners determine whether that deviating library usage should be added to the specification of proper library usage or the code fragment should be made compliant. The latter alternative makes both the codebase more regular and keeps the custom static analysis simpler. The method creates custom overviews that reveal opportunities to improve the usage of the in-house libraries, e.g., the removal of domain-specific redundant code which cannot be detected using generic tools, such as compilers and linters. We observed that industrial codebases are regular enough to create custom overviews using static analysis in the three exploratory case studies. Furthermore, we observed that the cross-checks, which detect deviating library usage, ensure the validity and completeness of the custom overviews. We conclude that producing custom overviews for in-house libraries using the method is valuable and feasible.

In this paper, we describe a case study at Philips Healthcare MRI focusing on evolutionary couplings, i.e., a technique to infer relationships among modules by analyzing their history of changes in the source code archive. In this case study, we failed to transfer CouplingViewer, a tool implementing the current state-of-art in evolutionary couplings, to industry. According to the industrial experts an important industrial requirement was not met: the signal-to-noise ratio was too low.

The way in which a system's software archive is partitioned influences the evolvability of that system. The partition of a software archive, e.g. subsystem decomposition, is mostly assessed by looking at the static (include, call) relations between the parts. In the literature history information is also taken into account to assess the partition. In this paper we describe our history-based approach to (automatically) assess the extent in which a certain partition allows its parts to evolve independently. We use the assumption that software entities which co-evolved often in the past are likely to be modified together in the near future as well. Hence, the elements of such a set should in principle belong to the same part. Our approach, therefore, identifies sets of co-evolving software entities, where each set has elements from more than one part of the archive. We illustrate our approach with a case study of a large software system that evolved during more than a decade, and has over 7 million lines of code.

In this article we introduce partial retraining, an algorithm to determine the relevance of the input variables of a trained neural network. We place this algorithm in the context of other approaches to relevance determination. Numerical experiments on both artificial and real-world problems show that partial retraining outperforms its competitors, which include methods based on constant substitution, analysis of weight magnitudes, and "optimal brain surgeon".

Intractable distributions present a common difficulty in inference within the probabilistic knowledge representation framework and variational methods have recently been popular in providing an approximate solution. In this article, we describe a perturbational approach in the form of a cumulant expansion which, to lowest order, recovers the standard Kullback-Leibler variational bound. Higher-order terms describe corrections on the variational approach without incurring much further computational cost. The relationship to other perturbational approaches such as TAP is also elucidated. We demonstrate the method on a particular class of undirected graphical models, Boltzmann machines, for which our simulation results confirm improved accuracy and enhanced stability during learning.

In this article, we introduce a measure of optimality for architecture selection algorithms for neural networks: the distance from the original network to the new network in a metric defined by the probability distributions of all possible networks. We derive two pruning algorithms, one based on a metric in parameter space and the other based on a metric in neuron space, which are closely related to well-known architecture selection algorithms, such as GOBS. Our framework extends the theoretically range of validity of GOBS and therefore can explain results observed in previous experiments. In addition, we give some computational improvements for these algorithms.