Brief

Researchers have developed GourNet, a Convolutional Neural Network model designed to detect diseases in mango leaves. Trained on the MangoLeafBD dataset, which includes eight classes (seven diseases and one healthy), GourNet achieved a 97% classification accuracy. The model utilizes data augmentation and preprocessing techniques, and its source code has been made publicly available. AI

Researchers have introduced Multiple Additive Neural Networks (MANN), a novel methodology that replaces decision trees with shallow neural networks in the Gradient Boosting framework. This approach integrates Convolutional Neural Networks (CNNs) and Capsule Neural Networks to handle both structured and unstructured data, including images and audio. MANN demonstrates improved accuracy and generalizability over traditional methods like Extreme Gradient Boosting (XGB) and offers enhanced robustness against overfitting. AI

A new study published on arXiv investigates the necessity of data augmentation and lung segmentation for AI-driven COVID-19 detection using chest X-rays. The research, which proposes a methodology called SDL-COVID, found that lung segmentation is crucial for accurate predictions. The study also demonstrated that excessive data augmentation can lead to model overfitting and decreased test accuracy. AI

Researchers have developed a novel method for interpreting Convolutional Neural Networks (CNNs) in image classification tasks by leveraging quantum annealing for feature selection. This approach identifies the most influential feature maps contributing to a model's predictions, aiming to enhance transparency and trust in AI systems. The technique encodes the feature selection problem into a quantum constrained optimization problem, which is then solved using quantum annealing. Evaluations show improved class disentanglement compared to existing explainable AI methods like GradCAM and GradCAM++. AI

Researchers have introduced a novel framework called Layer Separation Optimization to address challenges in training deep learning models with cross-entropy loss. This method aims to mitigate the strong nonconvexity issues that arise during the training of deep networks. By decomposing the complex optimization problem into smaller, more manageable subproblems using auxiliary variables, the framework theoretically provides an upper bound for the original cross-entropy loss and demonstrates improved optimization behavior in numerical experiments. AI

Researchers have developed PREMAP2, an enhanced algorithm for approximating neural network preimages, significantly improving scalability and efficiency. This new method extends the capabilities of its predecessor, PREMAP, allowing for analysis of more complex neural network architectures like convolutional neural networks. PREMAP2 can be applied to various certification tasks, including reliability, robustness, interpretability, and fairness, across different domains such as computer vision and control systems. The implementation is available as open-source software. AI

Researchers have developed a method to incorporate geographic information into motor insurance claim prediction models, even with limited location data. By utilizing environmental data from OpenStreetMap and CORINE Land Cover, along with visual features from satellite imagery, they enhanced the accuracy of zone-level claim frequency models. The study found that combining coordinates with environmental features at a 5 km scale was most beneficial for both linear and tree-based models, demonstrating that the representation of geographic context is more crucial than model complexity. AI