Human brain is adept at continually acquiring new skills through structured patterns without forgetting previous learning, a feat that neural networks struggle to emulate. When these networks are exposed to new information, they tend to experience a significant decline in performance on tasks they were previously trained on, a phenomenon known as catastrophic forgetting. We introduce a novel approach to representation learning that utilizes a hierarchical structure. In our proposed method, individual classes are encapsulated and dynamically routed to maintain relationships between similar classes at different levels. Our model develops into a tree-like architecture, with each phase adding nodes that contain capsules of semantically related classes. During training, only the nodes representing new classes and their children are trained, while the rest of the tree remains frozen. This effectively preserves the representations of older classes.

Understanding issue threads is an essential aspect of software maintenance and development, aiding developers in effectively addressing and managing software-related issues. These threads typically contain an issue description, comments discussing possible solutions, and often culminate in a pull request where the proposed changes are elaborated. Even though they are crucial, understanding issue threads can be a lot of work because they are often long and complex, particularly in big projects. This paper, therefore, aims to automate the process of issue thread summarization using advanced AI models, specifically the GPT-3.5-Turbo, reducing the time spent and improving the efficiency of the interpretation process. Our approach taps into the potential of the zero-shot learning methodology, enabling the model to produce context-specific summaries without reliance on prior examples. Additionally, we have developed an algorithm that determines the most effective length for these summaries, which enhances their clarity and relevance. The performance of the model is assessed using automated metrics, including ROUGE and BART scores, for extractive and abstractive summary evaluation respectively. Further, we may like to add that summaries of around 30% to 40% of the total size of the issue thread appears to be sufficient, though it varies slightly from case to case. The model’s successful generation of brief, clear, and pertinent summaries not only boosts team communication and project management but also lays the groundwork for its future integration into a comprehensive tool for simplified exploration and comprehension of complex software repositories.

Abstract Segmenting scientific abstracts into discourse categories like background, objective, method, result, and conclusion is useful in many downstream tasks like search, recommendation and summarization. This task of classifying each sentence in the abstract into one of a given set of discourse categories is called sequential sentence classification. Existing machine learning‐based approaches to this problem consider the content of only the abstract to obtain the neural representation of each sentence, which is then labelled with a discourse category. But this ignores the semantic information offered by the discourse labels themselves. In this paper, we propose LIHT, Label Informed Hierarchical Transformers – a method for sequential sentence classification that explicitly and hierarchically exploits the semantic information in the labels to learn label‐aware neural sentence representations. The hierarchical model helps to capture not only the fine‐grained interactions between the discourse labels and the words in the abstract at the sentence level but also the potential dependencies that may exist in the label sequence. Thus, LIHT generates label‐aware contextual sentence representations that are then labelled with a conditional random field. We evaluate LIHT on three publicly available datasets, namely, PUBMED‐RCT, NICTA‐PIBOSO and CSAbstract. The incremental gain in F1‐score in all the three cases over the respective state‐of‐the‐art approaches is around . Though the gains are modest, LIHT establishes a new performance benchmark for this task and is a novel technique of independent interest. We also perform an ablation study to identify the contribution of each component of LIHT in the observed performance, and a case study to visualize the roles of the different components of our model.

Contextualised word representations (e.g., ELMo and BERT) have been shown to outperform static representations (e.g., Word2vec, Fasttext, and GloVe) for many NLP tasks. In this paper, we investigate the use of contextualised embeddings for code search and classification, an area receiving less attention. We construct CodeELMo by training ELMo from scratch and fine tuning CodeBERT embeddings using masked language modeling based on natural language (NL) texts related to software development concepts and programming language (PL) texts consisting of method comment pairs from open source code bases. The dimensionality of the Finetuned Code BERT embeddings is reduced using linear transformations and augmented with a CodeELMo representation to develop CodeELBE – a lowdimensional contextualised software code representation. Results for binary classification and retrieval tasks show that CodeELBE1considerably improves retrieval performance on standard deep code search datasets compared to CodeBERT and baseline BERT models.

Abstract Approaches to evaluate comments based on whether they increase code comprehensibility for software maintenance tasks are important, but largely missing. We proposeCommentfor automated classification and quality evaluation of code comments of C codebases based on how they can help to understand existing code. We conduct surveys and document developers' perceptions on the type of comments that prove useful to maintaining software in the form of comment categories. A total of 20,206 comments have been collected from open‐sourceGithubprojects and annotated with assistance from industry experts. We develop features to semantically analyze comments to locate concepts related to categories of usefulness. Additionally, features based on code and comment correlation are designed to infer whether the comment is also consistent and not superfluous. Using neural networks, comments are classified asuseful,partially useful, andnot usefulwith precision and recall scores of 86.27% and 86.42%, respectively. The proposed framework for comment quality evaluation incorporates industry practices and adds significant value to companies wanting to formulate better code commenting strategies. Furthermore, large codebases can be de‐cluttered by removing comments not helpful in maintaining code.

The task of keyphrase generation, unlike extraction, aims to generate the phrases which succinctly capture the key information of the source text, that are even absent in the document (i.e., do not match any contiguous sub-sequence of source text). Despite the significant progress achieved by sequence-to-sequence (seq2seq) models in modelling such high entropy task, they are limited by their deterministic modelling capability which limits the generation of a diverse set of keyphrases. To address the above limitation, in this paper, we propose to incorporate Conditional Variational Autoencoder (CoVA) into seq2seq models for its ability to represent a set of keyphrases as a probabilistic distribution which improves the diversity of the generated keyphrases. We model the probabilistic distribution using a hierarchical latent structure where a global latent variable tries to model the diversity among the keyphrases and local latent variables control the generation of each keyphrase to make them coherent. Experimental results on four benchmark datasets of research papers demonstrate the effectiveness of our proposed approach in achieving a large improvement in diversity along with modest gains in quality with respect to previous models.

Singing Voice Detection (SVD) has been an active area of research in music information retrieval (MIR). Currently, two deep neural network-based methods, one based on CNN and the other on RNN, exist in literature that learn optimized features for the voice detection (VD) task and achieve state-of-the-art performance on common datasets. Both these models have a huge number of parameters (1.4M for CNN and 65.7K for RNN) and hence not suitable for deployment on devices like smartphones or embedded sensors with limited capacity in terms of memory and computation power. The most popular method to address this issue is known as knowledge distillation in deep learning literature (in addition to model compression) where a large pre-trained network known as the teacher is used to train a smaller student network. Given the wide applications of SVD in music information retrieval, to the best of our knowledge, model compression for practical deployment has not yet been explored. In this paper, efforts have been made to investigate this issue using both conventional as well as ensemble knowledge distillation techniques.

Keyphrases in a research paper succinctly capture the primary content of the paper and also assist in indexing the paper at a concept level. Given the huge rate at which scientific papers are published today, it is important to have effective ways of automatically extracting keyphrases from a research paper. In this paper, we present a novel method, Syntax and Semantics Aware Keyphrase Extraction (SaSAKE), to extract keyphrases from research papers. It uses a transformer architecture, stacking up sentence encoders to incorporate sequential information, and graph encoders to incorporate syntactic and semantic dependency graph information. Incorporation of these dependency graphs helps to alleviate long-range dependency problems and identify the boundaries of multi-word keyphrases effectively. Experimental results on three benchmark datasets show that our proposed method SaSAKE achieves state-of-the-art performance in keyphrase extraction from scientific papers.

We present the Fitness Based Layer Rank Selection (FLRS) Algorithm for Accelerating Convolutional Neural Networks by CANDECOMP/PARAFAC (CP) Decompositions. FLRS selects the layers and corresponding ranks based on a parameter fitness factor. The advantage of the proposed FLRS algorithm is that it does not require retraining iteratively during rank selection. The experimental results show that VGG-16 Network can be replaced by an approximate network where the convolutional layers are replaced by a sequence of four convolutional layers with smaller kernels. The approximated network has less than one-fifth of the original model parameters and performs less than one-fifth of the total number of computations as compared to the original model with an accuracy drop of less than 1% across SVHN, CIFAR-10 and CALTECH-101 datasets.

Regardless of attempts to extract knowledge from code bases to aid in program comprehension, there is an absence of a framework to extract and integrate knowledge to provide a near-complete multifaceted understanding of a program. To bridge this gap, we propose SMARTKT (Smart Knowledge Transfer) to extract and transfer knowledge related to software development and application-specific characteristics and their interrelationships in form of a knowledge graph. For an application, the knowledge graph provides an overall understanding of the design and implementation and can be used by an intelligent natural language query system to convert the process of knowledge transfer into a developer-friendly Google-like search. For validation, we develop an analyzer to discover concurrency-related design aspects from runtime traces in a machine learning framework and obtain a precision and recall of around 97% and 95% respectively. We extract application-specific knowledge from code comments and obtain 72% match against human-annotated ground truth.

Automatic segmentation of tumor in low dose scans like the Cone Beam Computed Tomography (CBCT) is quite challenging. We use a semi-automatic approach to segment tumor from non tumor using the classical level-set formulation. A pipeline of techniques, mainly involving gradient-based level-sets (GB) and Local Rank Transform (LRT) is used to achieve the tumor segmentation. Since CBCT images are prone to noise, the edge strength at the tumor and non-tumor boundary is very low. To improve the edge strength in the CBCT image, we propose to use the edges obtained from the LRT-attractor of the image. The gradient-based level-sets with LRT-attractor (GBLA) is a non-linear technique that helps in strengthening the latent tumor and non-tumor boundary. We compare the GBLA level-sets with the GB level-sets technique, and report our results on 307 volumes of 45 patients. It was found that average precision is improved by 10% when using GBLA.

Program comprehension is a major challenge for system maintenance. Reverse engineering has been employed for control-flow analysis of applications but not much work has been done for comprehending concurrent non-deterministic behavior of multi-threaded applications. We present D-CUBE, built using dynamic instrumentation APIs, which plugs in during execution and infers various thread models like concurrency, safety, data access, thread-pool state, exception model etc. for multi-threaded applications at runtime. We extract run-time events traced according to pre-specified logic and feed them to decision trees for inference. We use 3 benchmark suites (LOC: 50-3200) -- CDAC Pthreads benchmark [1] (18 Cases), Open POSIX Test-Suites [2] (21 Cases) and PARSEC 3.0 benchmarks [3] (3 Cases) for accuracy and volume testing and validate our approach by comparing the documented behavior of test-suites with D-CUBE's output models. We achieve over 90% accuracy. D-CUBE produces graphical event-traces with every inference for quick and effective comprehension of large code.

Representation schemes play an important role in the fields of Computer Vision, Graphics, Image Processing, CAD/CAM etc. Various representation schemes have been discussed in the literature for both 2D and 3D. In this paper, we are presenting a scheme of representation using the concept of octagonal distances. They are called Medial Circle Representation (MCR) and Medial Sphere Representation (MSR) in 2D and 3D, respectively. Storage requirement, computational complexity, merits and demerits of the representation schemes are discussed.