We propose a memory-augmented transformer framework for few-shot classification in biomedical imaging. It adapts to new domains with limited samples using context-aware memory recall and transferable visual features, significantly improving accuracy in complex, low-resource medical datasets.

Neha Devika Menon, Carlos Fernando Morales, Aditya Ramesh Iyer, Jean Claude Dufour, Elena Francesca Romero

We introduce an explainable model for legal sentencing predictions. Using rule-injected attention and ontology-based constraints, it improves interpretability, ensuring alignment with legal reasoning and increasing transparency in AI-supported judicial decision systems.

James Oliver Watkins, Laura Isabelle Prescott, Rachel Anne Turner, Olivia Grace Jenkins, Benjamin Paul Maddox

This research proposes a reinforcement learning model for resource allocation in unstable IoT networks. The model dynamically adjusts compute strategies under constrained bandwidth, improving real-time responsiveness and task efficiency for edge-based intelligent devices.

Ravi Kunal Raj, Jean Luis Martinez, Priya Shanti Mehra, Maria Camila Navarro, Taro Masaki Ishida

We propose a spatiotemporal transformer model for urban crowd flow prediction. It compensates for sparse and noisy sensor data using multilevel temporal attention, enhancing movement forecasting accuracy in complex metropolitan areas with variable pedestrian and traffic conditions.

Chen Yu Lin, Zhang Tian Wei, Gao Min Zhi, Liu Sheng Bo, Xu Liang Hao

This paper introduces a multimodal object tracking model using fusion transformers. It aligns thermal, visual, and radar signals with temporal attention, enabling robust tracking across devices and sensor types, particularly in low-visibility and multi-perspective surveillance scenarios.

Robert Patrick Morrison, Emily Noelle Harris, Sarah Ann Fletcher, Thomas Bradley Rowe, James Nathaniel Foster

We introduce a curriculum-based meta-learning model for zero-shot classification in specialized domains. The system dynamically orders tasks by difficulty and knowledge overlap, facilitating concept transfer and improving performance where domain-specific labeled data is unavailable.

Liu Rong Chen, Xu Wei Zhang, Gao Liang Cheng, Huang Xiang Tao, Zhang Feng Ming

This paper presents a self-supervised learning model that builds hierarchical representations for sparse geospatial time series. It reconstructs temporal continuity and spatial patterns using auto-regressive attention blocks, enhancing long-term forecasting accuracy in satellite-based environmental monitoring systems.

Chen Wei Zhang, Liu Hao Ming, Xu Fang Zhen, Gao Lin Yu, Zhang Jun Rui