Lifeboat Foundation: Safeguarding Humanity
Toggle light / dark theme

Blog

Category: robotics/AI – Page 17

Google launches speech dataset for African languages

Google has collaborated with African universities and research institutions to launch WAXAL, an open-source speech database designed to support the development of voice-based artificial intelligence for African languages.

African institutions, including Makerere University in Uganda, the University of Ghana, Digital Umuganda in Rwanda, and the African Institute for Mathematical Sciences (AIMS), participated in the data collection for this initiative. The dataset provides foundational data for 21 Sub-Saharan African languages, including Hausa, Luganda, Yoruba, and Acholi.

WAXAL is designed to support the development of speech recognition systems, voice assistants, text-to-speech tools, and other voice-enabled applications across sectors such as education, healthcare, agriculture, and public services.

A new robotic system could perform delicate eye surgery

Retinal vein occlusion (RVO) is a severe disease that occurs when a vein in the light-sensitive layer at the back of the eye (i.e., the retina) becomes blocked, which results in a loss of vision. There are currently a few medical interventions that address RVO, including the periodic injection of medications that block the abnormal growth of blood vessels or of steroids, which reduce swelling and inflammation.

A promising procedure for the treatment of RVO is retinal vein cannulation (RVC). This is a very delicate surgical intervention that requires surgeons to insert a tiny needle into the blocked vein with high precision, delivering clot-dissolving drugs or medications that control the abnormal growth of blood vessels.

Given that retinal veins targeted for cannulation are similar in thickness to a human hair, manually inserting a needle inside them with high precision is very challenging. Robots could potentially assist surgeons in performing RVO procedures, ensuring that needles are inserted correctly and without damaging the patients’ retina.

AI tool predicts brain age, cancer survival and other disease signals from unlabeled brain MRIs

Mass General Brigham investigators have developed a robust new artificial intelligence (AI) foundation model that is capable of analyzing brain MRI datasets to perform numerous medical tasks, including identifying brain age, predicting dementia risk, detecting brain tumor mutations and predicting brain cancer survival. The tool, known as BrainIAC, outperformed other, more task-specific AI models and was especially efficient when limited training data were available.

Results are published in Nature Neuroscience.

“BrainIAC has the potential to accelerate biomarker discovery, enhance diagnostic tools and speed the adoption of AI in clinical practice,” said corresponding author Benjamin Kann, MD, of the Artificial Intelligence in Medicine (AIM) Program at Mass General Brigham. “Integrating BrainIAC into imaging protocols could help clinicians better personalize and improve patient care.”

AI tool can predict which trauma patients need blood transfusions before they reach the hospital

Severe bleeding is one of the most common and preventable causes of death after traumatic injury, yet currently available tools have poor ability to determine which patients urgently need blood transfusions. A new multinational study, just published in Lancet Digital Health, suggests artificial intelligence (AI) may help close that gap.

Researchers have developed and validated machine-learning models that can accurately predict whether trauma patients will require blood transfusions, using only information available before they reach the hospital such as vital signs, injury patterns, and medication history.

Co-author Prof Patricia Maguire from University College Dublin (UCD), Director of UCD AI Healthcare Hub and UCD Institute for Discovery, said, “These findings show that AI-driven decision support could enable earlier and more precise identification of patients at highest risk of hemorrhagic shock, using data already available to emergency services. This has clear potential to support more timely transfusion decisions, although prospective evaluation will be needed before clinical implementation.”

Decoding the shadows: Vehicle recognition software uncovers unusual traffic behavior

Researchers at the Department of Energy’s Oak Ridge National Laboratory have developed a deep learning algorithm that analyzes drone, camera, and sensor data to reveal unusual vehicle patterns that may indicate illicit activity, including the movement of nuclear materials. The work is published in the journal Future Transportation.

The software monitors routine traffic over time to establish a baseline for “patterns of life,” enabling detection of deviations that could signal something out of place. For example, a surge in overnight truck traffic at a facility which is normally only visited during the day could reveal illegal shipments.

The research builds on a previous ORNL-developed technology for recognizing specific vehicles from side views. Researchers improved the structure of this software’s deep learning network to provide much broader capabilities than any existing recognition systems, said ORNL’s Sally Ghanem, lead researcher.

AI-powered compressed imaging system developed for high-speed scenes

A research team from the Xi’an Institute of Optics and Precision Mechanics (XIOPM) of the Chinese Academy of Sciences, along with collaborators from the Institute National de la Recherche Scientifique, Canada, and Northwest University, has developed a single-shot compressed upconversion photoluminescence lifetime imaging (sCUPLI) system for high-speed imaging.

High-fidelity recovery from complex inverse problems remains a key challenge in compressed high-speed imaging. Deep learning has revolutionized the reconstruction, but pure end-to-end “black-box” networks often suffer from structural artifacts and high costs. To address these issues, the team from XIOPM propose a multi-prior physics-enhanced neural network (mPEN) in an article published in Ultrafast Science.

By integrating mPEN with compressed optical streak ultra-high-speed photography (COSUP), the researchers developed the sCUPLI system. This system utilized an encoding path for temporal shearing and a prior path to record unencoded integral images. It effectively suppressed artifacts and corrected spatial distortion by synergistically correcting multiple complementary priors including physical models, sparsity constraints, and deep image priors.

A ‘crazy’ dice proof leads to a new understanding of a fundamental law of physics

Right now, molecules in the air are moving around you in chaotic and unpredictable ways. To make sense of such systems, physicists use a law known as the Boltzmann distribution, which, rather than describe exactly where each particle is, describes the chance of finding the system in any of its possible states. This allows them to make predictions about the whole system even though the individual particle motions are random. It’s like rolling a single die: Any one roll is unpredictable, but if you keep rolling it again and again, a pattern of probabilities will emerge.

Developed in the latter half of the 19th century by Ludwig Boltzmann, an Austrian physicist and mathematician, this Boltzmann distribution is used widely today to model systems in many fields, ranging from AI to economics, where it is called “multinomial logit.”

Now, economists have taken a deeper look at this universal law and come up with a surprising result: The Boltzmann distribution, their mathematical proof shows, is the only law that accurately describes unrelated, or uncoupled, systems.

/* */