With the increasing impacts of climate change, there is a growing demand for accessible tools that can provide reliable future climate information to support planning, finance, and other decision-making applications. Large language models (LLMs), such as GPT-4o, present a promising approach to bridging the gap between complex climate data and the general public, offering a way for non-specialist users to obtain essential climate insights through natural language interaction. However, an essential challenge remains underexplored: Evaluating the ability of LLMs to provide accurate and reliable future climate predictions, which is crucial for applications that rely on anticipating climate trends. In this study, we investigated the capability of GPT-4o in predicting rainfall at short-term (15-day) and long-term (12-month) scales. We designed a series of experiments to assess GPT’s performance under different conditions, including scenarios with and without expert data inputs. Our results indicated that GPT, when operating independently, tended to generate conservative forecasts, often reverting to historical averages in the absence of clear trend signals. This study highlights the potential and challenges of applying LLMs for future climate predictions, providing insights into their integration with climate-related applications and indicating directions for enhancing their predictive capabilities in the field.

Deep learning-based object detection models perform well under daytime conditions but face significant challenges at night, primarily because they are predominantly trained on daytime images. Additionally, training with nighttime images presents another challenge: Even human annotators struggle to accurately label objects in low-light conditions. This issue is particularly pronounced in transportation applications, such as detecting vehicles and other objects of interest on rural roads at night, where street lighting is often absent, and headlights may introduce undesirable glare. In this study, we addressed these challenges by introducing a novel framework for labeling-free data augmentation, leveraging synthetic data generated by the Car Learning to Act (CARLA) simulator for day-to-night image style transfer. Specifically, the framework incorporated the efficient attention Generative Adversarial Network for realistic day-to-night style transfer and used CARLA-generated synthetic nighttime images to help the model learn the vehicle headlight effect. To evaluate the efficacy of the proposed framework, we fine-tuned the state-of-the-art object detection model with an augmented dataset curated for rural nighttime environments, achieving significant improvements in nighttime vehicle detection. This novel approach was simple yet effective, offering a scalable solution to enhance deep learning-based detection systems in low-visibility environments and extended the applicability of object detection models to broader real-world contexts.

Reliable and interpretable traffic crash modeling is essential for understanding causality and improving road safety. This study introduced a novel approach to predicting crash types by utilizing a comprehensive dataset fused from multiple sources, including weather data, crash reports, high-resolution traffic information, pavement geometry, and facility characteristics. An essential part of our proposed approach was a feature group tabular transformer (FGTT) model, which organizes disparate data into meaningful feature groups, represented as tokens. These group-based tokens serve as rich semantic components, enabling effective identification of collision patterns and interpretation of causal mechanisms. The FGTT model was compared with widely used tree ensemble models, including random forest, XGBoost, and CatBoost, demonstrating better predictive performance. Furthermore, the attention heatmaps from the FGTT model revealed key influential factor interactions, providing fresh insights into the underlying causality of distinct crash types.

The non-contact blood pressure (BP) monitoring technology based on millimeter wave radar (mmWave) has been widely concerned for its advantages of non-invasive and real-time continuous monitoring. In recent years, studies have employed deep learning technologies to process mmWave radar, providing high-accuracy monitoring and high computing resource requirements. In this paper, we propose an edge-assisted framework for mmWave radar-based blood pressure monitoring to meet high accuracy and low latency application requirements because edge computing can provide a more powerful computing capability closer to users. However, it is non-trivial to effectively run such an edge-assisted mmWave radar-based blood pressure monitoring upon multiple users due to limited edge server resources. To solve this problem, we identify an opportunity to optimize the inference efficiency by adjusting key system parameters, such as sampling interval and input signal sequence length. This adjustment helps to reduce the inference latency and resource contention, especially in resource-constrained edge computing environments. By adaptively configuring these parameters for multiple users, we aim to strike a balance between a high accuracy and a low latency. First, we formulate the problem as an online learning problem and propose a deep reinforcement learning-based method to solve it. Finally, we implement a testbed to evaluate the performance of our method. Extensive experimental results show that our method outperforms the baselines, achieving a latency reduction of up to 70.3% and improving a reward by up to 29.7%, while maintaining an accuracy loss within 5%.

Campus as a Living Lab at the National University of Singapore is an initiative to bridge technologies from academia to industry across the valley of death [1]. The valley of death is the gap between research & development and translation & commercialization where many favorable innovations disappear. It poses significant challenges for academics, researchers, developers, and startups. If carefully resourced and shepherded, it will see real-world deployment for commercially viable products or services from academic laboratories. The university has put forth its campus grounds, infrastructure, and equipment to support this. The university has also developed its digital twin to assess each project’s feasibility, which will be test bedding technologies under the living lab framework. This paper will discuss in detail the digital twin, evaluation criteria, validation processes, and the two ongoing projects. It also will describe how the digital twin was accurately modeled and created, as well as how it was used to assist in the feasibility studies of each project. To cross the valley of death, the university has established the strategies and partnership framework to successfully bring research-driven innovations into impactful market-ready technologies and capabilities. We will invite more companies to adopt academic research for real-world deployment.

The process of genre classification involves the identification of distinctive stylistic elements and musical characteristics that define a particular genre. It assists in developing a comprehensive understanding of the historical context, cultural influences, and musical evolution of a particular genre. This study was conducted to resolve the challenges of classifying Ethiopian music genres according to their melodic structures using deep learning techniques. The main objective was to develop a deep learning model for effective audio classification into six genres classes of Ethiopian music: Ancihoye Lene, Ambassel Major, Ambassel Minor, Bati, Tizita Major, and Tizita Minor. To achieve this, we first prepared a dataset consisting of 3952 audio recordings, which includes 533 tracks from Ethiopian Orthodox church music and 3419 samples of secular Ethiopian music. A total of 46 unique features, namely chroma short-time Fourier transform (STFT), root mean square error (RMSE), spectral centroid, spectral bandwidth, roll-off, zero crossing rate, and mel frequency cepstral coefficient (MFCC) 1 up to MFCC40, were extracted both at middle-level and low-level audio features from each sample, focusing on aspects suggested by Ethiopian music experts and preliminary experiments that highlighted the importance of tonality features. A 30-second segment of audio recordings was selected for feature extraction, resulting in datasets formatted in both CSV and JSON for further processing. We proposed deep learning algorithms namely convolutional neural networks (CNN), recurrent neural networks (RNN), a parallel RNN–CNN architecture, and long short-term memory (LSTM) networks for our classification by developing models. Our experiments revealed that the LSTM model achieved the best performance, reaching a classification accuracy of 97% using 40 MFCC features extracted from audio datasets.

This paper explored self-supervised feature extraction in the K-views algorithm for texture analysis, which can reduce the reliance on labeled data. The study compared the supervised, and self-supervised methods, focusing on pair-wise feature comparison to assess the similarity and dissimilarity between texture classes. Among the techniques evaluated, Siamese networks achieved the highest performance, while self-supervised methods like rotation prediction demonstrated competitive results and scalability. Experimental results on KTH-TIPS, Kylberg, and UIUC datasets show the potential of self-supervised approaches in improving texture classification by leveraging feature representations without the need for manually labeled data for the K-views algorithm.