Image reconstruction—the process of recovering clear images from incomplete or noisy data—has been advancing rapidly through deep learning. Yet most existing approaches rely on costly supervised training and lack theoretical transparency. A new survey maps the rise of unsupervised deep learning for image reconstruction, from traditional denoising-based priors to modern diffusion models. These methods learn structured visual information directly from unlabeled data, and have achieved impressive performance across various fields, including biomedical imaging and remote sensing. The study shows how unsupervised learning based image reconstruction unites neural network efficiency with solid mathematical foundations to achieve both interpretability and flexibility, offering a blueprint for next-generation imaging systems.

Researchers at the University of Melbourne have developed a new AI-based traffic signal control system called M²SAC that improves both fairness and efficiency at urban intersections. Unlike traditional systems focused only on cars, M²SAC accounts for pedestrians, buses, and other users. A key innovation is the phase mask mechanism, which dynamically adjusts green light timings to reduce delays. Tested on real Melbourne traffic data, the model outperformed existing methods, cutting congestion and balancing traffic flow more equitably. The approach supports smarter, fairer, and more inclusive transport systems for modern cities.

To address this challenge, researchers at Korea Advanced Institute of Science and Technology (KAIST) and Donghai Laboratory developed a new model called ProChunkFormer, which reconstructs vehicle trajectories from sparse and noisy GPS data, enabling more accurate mobility analysis and intelligent transportation planning.

The heterogeneity causes spatiotemporal inconsistencies in multimodal data, posing challenges for existing methods in multimodal feature extraction and alignment. First, in the temporal dimension, the microsecond-level temporal resolution of event data is significantly higher than the millisecond-level resolution of RGB data, resulting in temporal misalignment and making direct multimodal fusion infeasible. To address this issue, the researchers design an Event Correction Module (ECM) that temporally aligns asynchronous event streams with their corresponding image frames through optical-flow-based warping. The ECM is jointly optimized with the downstream object detection network to learn task-ware event representations.

In collaboration with universities across the world, Nicholas Hedger (University of Reading) and Tomas Knapen (Netherlands Institute for Neuroscience & Vrije Universiteit Amsterdam) explored the depths of the human experience. They discovered how the brain translates the visual world around us into touch, thereby creating a physical embodied world for us to experience. “This aspect of human experience is a fantastic area for AI development.”

An innovation from the University of Chicago Pritzker School of Molecular Engineering represents a major advance in stretchable, bendable digital displays for next-generation electronics and medical devices. The lab had previously created a high-efficiency electroluminescent material that could stretch to more than twice its original length while still emitting a fluorescent pattern. But two layers—the cathode layer and the electron transport layer—had to remain rigid for the screen to work. In the new publication, an innovative aluminum gel and a novel family of conductive polymers overcame those two last rigid hurdles.

KTU, Lithuania, scientists have developed a method to convert textile waste into both alternative fuel and high-value cement additives. Their findings show that textile ash can replace up to 7.5% of conventional cement and increase material strength by up to 16%, while recycled polyester fibres improve concrete performance by 15–20%. The research offers a sustainable pathway for managing textile waste and reducing CO₂ emissions in cement production.