Pruning studies up to date focused on uncovering a smaller network by removing redundant units, and fine-tuning to compensate accuracy drop as a result. In this study, unlike the others, we propose an approach to uncover a smaller network that is competent only in a specific task, similar to top-down attention mechanism in human visual system. This approach doesn’t require fine-tuning and is proposed as a fast and effective alternative of training from scratch when the network focuses on a specific task in the dataset. Pruning starts from the output and proceeds towards the input by computing neuron importance scores in each layer and propagating them to the preceding layer. In the meantime, neurons determined as worthless are pruned. We applied our approach on three benchmark datasets: MNIST, CIFAR-10 and ImageNet. The results demonstrate that the proposed pruning method typically reduces computational units and storage without harming accuracy significantly.
Keywords: Deep learning | Computer vision | Network pruning | Network compressing | Top-down attention | Perceptual visioning

The abundance of accessible educational data, supported by the technology-enhanced learning platforms, provides opportunities to mine learning behavior of students, addressing their issues, optimizing the educational environment, and enabling data-driven decision making. Virtual learning environments complement the learning analytics paradigm by effectively providing datasets for analysing and reporting the learning process of students and its reflection and contribution in their respective performances. This study deploys a deep artificial neural network on a set of unique handcrafted features, extracted from the virtual learning environments clickstream data, to predict at-risk students providing measures for early intervention of such cases. The results show the proposed model to achieve a classification accuracy of 84%–93%. We show that a deep artificial neural network outperforms the baseline logistic regression and support vector machine models. While logistic regression achieves an accuracy of 79.82%–85.60%, the support vector machine achieves 79.95%–89.14%. Aligned with the existing studies - our findings demonstrate the inclusion of legacy data and assessment-related data to impact the model significantly. Students interested in accessing the content of the previous lectures are observed to demonstrate better performance. The study intends to assist institutes in formulating a necessary framework for pedagogical support, facilitating higher education decision-making process towards sustainable education.
Keywords: Learning analytics | Predicting success | Educational data | Machine learning | Deep learning | Virtual learning environments (VLE)

The detection and classification of bruises of pears based on thermal images have been investigated. A simple thermal imaging system in the long-wavelength ranges (8–14 μm) assembledμwith hot air equipment was constructed to capture cleaner images. Higher velocity and temperature of the air reduced the time required to obtain a clean image, but the images were not sufficient able to discriminate the slight and invisible variation of bruises over consecutive days. The grey-level co-occurrence matrix of the thermal images were analysed, and the slight differences in the pears over consecutive days were presented in the form of a line chart. A traditional deep learning algorithm commonly used in classification of big data sets was modified to one suitable for classification of a small sample data set of thermasl images (3246 samples were used as the training data set and 1125 were used as a test data set) collected from 300 pears over 10 days. The best test prediction accuracy obtained was 99.25%.
Keywords: Detection and classification | Thermal images | Grey-level co-occurrence matrix | Deep learning

Weight-sharing is one of the pillars behind Convolutional Neural Networks and their successes. However, in physical neural systems such as the brain, weight-sharing is implausible. This discrepancy raises the fundamental question of whether weight-sharing is necessary. If so, to which degree of precision? If not, what are the alternatives? The goal of this study is to investigate these questions, primarily through simulations where the weight-sharing assumption is relaxed. Taking inspiration from neural circuitry, we explore the use of Free Convolutional Networks and neurons with variable connection patterns. Using Free Convolutional Networks, we show that while weight-sharing is a pragmatic optimization approach, it is not a necessity in computer vision applications. Furthermore, Free Convolutional Networks match the performance observed in standard architectures when trained using properly translated data (akin to video). Under the assumption of translationally augmented data, Free Convolutional Networks learn translationally invariant representations that yield an approximate form of weight-sharing.
Keywords: Deep learning | Convolutional neural networks | Weight-sharing | Biologically plausible architectures

Context: COSMIC is a widely used functional size measurement (FSM) method that supports software development effort estimation. The FSM methods measure functional product size based on functional requirements. Unfortu- nately, when the description of the product’s functionality is often abstract or incomplete, the size of the product can only be approximated since the object to be measured is not yet fully described. Also, the measurement performed by human-experts can be time-consuming, therefore, it is worth considering automating it. Objective: Our objective is to design a new prediction model capable of approximating COSMIC-size of use cases based only on their names that is easier to train and more accurate than existing techniques. Method: Several neural-network architectures are investigated to build a COSMIC size approximation model. The accuracy of models is evaluated in a simulation study on the dataset of 437 use cases from 27 software develop- ment projects in the Management Information Systems (MIS) domain. The accuracy of the models is compared with the Average Use-Case approximation (AUC), and two recently proposed two-step models —Average Use-Case Goal-aware Approximation (AUCG) and Bayesian Network Use-Case Goal AproxImatioN (BN-UCGAIN). Results: The best prediction accuracy was obtained for a convolutional neural network using a word-embedding model trained on Wikipedia + Gigaworld. The accuracy of the model outperformed the baseline AUC model by ca. 20%, and the two-step models by ca. 5–7%. In the worst case, the improvement in the prediction accuracy is visible after estimating 10 use cases. Conclusions: The proposed deep learning model can be used to automatically approximate COSMIC size of software applications for which the requirements are documented in the form of use cases (or at least in the form of use- case names). The advantage of the model is that it does not require collecting historical data other than COSMIC size and names of use cases.
Keywords: Functional size approximation | Approximate software sizing methods | COSMIC | Deep learning | Word embeddings | Use cases