This work evaluates several novel Deep Learning (DL) architectures for the classification of gastrointestinal diseases using the Kvasir-capsule dataset. In addition to classification performance, the paper incorporates Explainable AI (XAI) techniques to provide visual insights into the decision-making processes of the models, addressing the 'black box' nature of deep learning in medical imaging diagnosis.

The rise in crime rates coupled with advancements in computer vision has increased the need for automated crime detection services. We propose a new approach for detecting suspicious behavior to prevent shoplifting using object detection based on YOLOv5 with Deep Sort to track individuals and extract bounding box coordinates as temporal features. These features are then modeled as a time-series classification problem, achieving an F1 score of 92% on the UCF Crime dataset while significantly outperforming state-of-the-art methods in inference speed.

This paper presents a systematic review of recent advances in music generation using deep learning techniques. We categorize the latest research by examining common data representations such as raw waveforms, spectrograms, and MIDI, alongside prominent architectures like Generative Adversarial Networks (GANs), Recurrent Neural Networks (RNNs), Variational Autoencoders (VAEs), and Transformer-based models.

This research conducts an in-depth comparative analysis of two prominent cloud platforms, Microsoft Azure and Amazon Web Services (AWS), focusing on their suitability for deploying object-detection algorithms. The study evaluates both quantitative metrics (upload/download times, throughput, inference time) and qualitative assessments (cost, ease of deployment). We conclude that Azure excels in latency and throughput metrics, while AWS offers advantages in upload speeds, cost-effectiveness, and a wider service catalog.

Feature selection is a critical step in data analysis, yet formal evaluation benchmarks are lacking. We propose a synthetic data toolkit that generates datasets with controlled properties—such as feature correlation, noise levels, and relevance—to rigorously evaluate the performance of feature selection algorithms. The toolkit provides a standardized environment for comparing different selection methods.

This chapter explores the creation and utilization of synthetic data for the specific purpose of feature selection. It details methodologies for generating synthetic datasets that mimic real-world complexities while retaining known ground truths, enabling more accurate validation and benchmarking of feature selection algorithms.

We address the lack of formal evaluation benchmarks for feature selection algorithms (FSAs) by proposing a framework for generating synthetic datasets. This paper outlines the design of these benchmarks, which include varying degrees of feature relevance, redundancy, and interaction, to provide a robust testing ground for FSAs.

Respiration-correlated cone-beam computed tomography (4D-CBCT) often suffers from artifacts due to sparse sampling. We propose a machine learning-based approach to interpolate intermediate X-ray projections using pre-trained video frame interpolation models (like RIFE) and a novel regression-based model. This augmentation of projection data significantly improves the quality of the reconstructed 4D-CBCT images.

This paper proposes a Bayesian filter-based method using the Relative Belief Ratio (RBR) to assess feature significance, addressing a common limitation in many filter approaches. The method provides a strength value that serves as an importance score for ranking features. Experimental results on both synthetic and real-world datasets show that the RBR-based approach performs exceptionally well for binary classification compared to popular existing filter methods.

Current pipeline inspection methods often rely on manual procedures that introduce errors. We propose the design of a hybrid rolling and flying UAV developed with a rolling frame capable of landing on and traversing pipelines to externally inspect them for cracks, leaks, or faults. The system utilizes an onboard camera and deep learning models to classify surface cracks, combining the mobility of a drone with the stability of a crawling robot.

This study examines nitrogen dioxide (NO₂) pollution in Italy, a key environmental concern linked to the UN Sustainable Development Goals. Using open-source data and machine learning models, the researchers estimated NO₂ concentrations with strong predictive accuracy. Results show that model performance improves when data is segmented by season and urbanization level. Urban areas exhibited the highest NO₂ levels, while rural regions showed much lower concentrations. The findings highlight the value of context-specific modeling and demonstrate that open-source data paired with machine learning can effectively track NO₂ pollution across diverse regions.

This paper explores the Forward-Forward (FF) algorithm as a biologically inspired, gradient-free alternative to backpropagation for training neural networks on resource-constrained edge devices. We apply FF to classify camera trap images, demonstrating its potential for decentralized and efficient learning in wildlife monitoring applications.

Intrusion detection datasets often suffer from severe class imbalance, making it difficult to detect rare attack types. We propose 'Gamma Sampling', a novel oversampling technique designed to address this imbalance. By generating synthetic samples for the minority class based on a Gamma distribution, we improve the detection rates of intrusion detection systems.

We explored data from the 2022 and 2023 Dakar Bike Rally to determine the most important factors for building a successful race strategy. Using unsupervised clustering on time series from different rally stages along with autoencoders, we analyzed the effect of teams, sponsorship, and experience on final rankings. Our findings indicate that performance in the initial stages is key to the final outcome, while factors like team affiliation and rider experience contributed less significantly than expected.