Researchers developed a high-performance photonic spiking neural network that surpasses traditional digital systems with its ultrafast performance and low power consumption. The new network achieved excellent classification accuracies of over 94%, outperforming benchmark results with small training sets.
Researchers developed a modified bandit Q-learning algorithm that aims to learn optimal Q values for every state-action pair, balancing exploitation and exploration. The scheme relies on photonic systems to enhance learning quality, accelerating parallel learning through conflict-free decision-making.
Researchers developed a neural network-based system, PAT, for snapshot compressive imaging. It achieves comparable image quality to CASSI and holds strong promise due to advancements in AI processing capabilities.
A new quantum-classical hybrid algorithm has been developed to accelerate dynamic mode decomposition for high-dimensional time series analysis. The algorithm can operate with a small number of samples and has a quantum advantage in the analysis of high-dimensional time series.
Researchers propose leveraging high-value information to overcome statistical and computational challenges in reinforcement learning. By accessing valuable observations, agents can improve strategies without trial and error, making the learning process more efficient and effective.
A joint research team has developed a novel approach combining machine learning with quantum-classical computational molecular design to accelerate the discovery of efficient OLED emitters. The optimal OLED emitter discovered is a deuterated derivative of Alq₃, which is both extremely efficient at emitting light and synthesizable.
Researchers use photonics to accelerate convolutional neural networks by harnessing light's unique properties, reducing power consumption and increasing efficiency. This approach enables real-time image processing and scalable solutions for complex images.
Researchers summarize existing compiler technologies in deep learning co-design and propose a new framework, the Buddy Compiler, to address performance bottlenecks in current AI applications. The study highlights the importance of hardware-software co-design in achieving optimal efficiency and effectiveness in deep learning systems.
This field uses trial-and-error learning with natural selection to solve complex reinforcement learning tasks, but requires significant computational resources. Researchers can enhance its efficiency by improving encoding, sampling, search operators, algorithmic frameworks, and evaluation methods.
A team of researchers developed an unsupervised entity alignment framework to improve knowledge graph search, avoiding human labor. The framework outperformed most competitors on precision and recall, scoring higher overall across multiple datasets.
A new neural network, CD-GAN, uses common sense knowledge to enhance text descriptions and generate images of birds at three resolution levels. The system achieved competitive scores against other image generation methods, producing vivid and natural-looking images.
A comprehensive review of e-nose methods and algorithms aims to improve smell detection capabilities. The study highlights limitations of current gas sensors and provides an outlook on algorithm design.
A new theorem helps system designers make informed decisions about consistency and availability trade-offs in networks. The CAL theorem provides a quantitative relationship between these factors, allowing for more efficient and reliable network designs.
Researchers successfully connected human brain to computer for simple computational imaging tasks, using ghost imaging to reconstruct images of an object behind a wall. The use of real-time feedback from the visual cortex improved both imaging speed and image quality.
Researchers developed a deep learning approach to recognize and predict motion using vector-based relative change in position. The method, VecNet+LSTM, scored higher than other frameworks in recognizing motion and predicting future movements. This study has implications for machine learning in video analysis and artificial intelligence.
A new inverse rendering framework enables fast and efficient cloud tomography, allowing for accurate analysis of atmospheric dynamics and energy balance. The path recycling and sorting algorithm speeds up the process, overcoming computational limitations.
Researchers have developed a new method to encode single-cell tomograms using 3D Zernike polynomials, resulting in a data compression ratio of 22.9 and saving over 95% of space. The strategy is efficient on various experimental data types and can be applied to lab-on-a-chip systems.
A team of researchers at the University of Connecticut created freeform illuminators that enable flexible illumination design and calibration using a blood-coated sensor. The newly developed technology simplifies microscopy experiments by reducing size, increasing density, and adjusting angle of illumination.
Researchers have successfully applied speckle illumination to photoacoustic microscopy, reducing tissue damage and improving image reconstruction. The technique harnesses the power of structured illumination methods initially developed for optical microscopy, allowing for more efficient imaging with acoustic detection.
According to Li, machine intelligence is based on a combination of matter, energy, structure, and time, which he calls
Researchers developed an efficient algorithm that combines classical and quantum correlation functions to improve super-resolution microscopy. The algorithm, called 'super deconvolution imaging,' results in increased spatial frequency content, reduced mean squared errors, and faster imaging speeds.
Researchers developed a new ghost imaging algorithm to address image quality limitations in electron microscopy, achieving improved resolution and contrast using lower flux illumination. The approach enables robust transmission electron microscopy imaging with reduced sample damage.
A new neural network method has been developed to measure the 3D distance between two differently-colored spots in microscopic samples. This approach combines PSF engineering with machine learning to overcome accuracy limitations due to diffraction and provides accurate measurements for studying cellular processes.
The field of intelligent computing has made significant progress, focusing on autonomous perception, information gathering, analysis, and reasoning. Key findings include the need to develop more flexible systems that can accommodate diverse forms of intelligence through computing.
Researchers developed intelligent programmable meta-imagers that generate learned illumination patterns to pre-select relevant details during measurement, improving high-accuracy sensing with reduced measurements. The system adapts to different types and levels of noise, outperforming conventional compressed sensing.
Ferroelectric materials have shown promising solutions for intelligent computing, including low-power logic devices, high-performance memory cells, and neuromorphic devices. These advancements can break the 'heat wall', 'memory wall', and von Neumann bottleneck, respectively.
Researchers have developed a new approach to phase retrieval in coherent X-ray imaging, using a complexity parameter to guide the algorithm. This methodology reduces artifacts and improves solution quality, resulting in higher-resolution images of micro- and nano-sized objects.
Deep learning empowers cell image analysis by automating tasks such as segmentation and tracking, improving research efficiency. However, there are challenges in data quantity, quality, and confidence that remain to be addressed.
Researchers propose a passive optical device, called an optical inverter, to undo the effects of multimode optical fibers on spatial information. The inverter can achieve single-shot wide-field imaging and super-resolution imaging through MMFs, enabling potential applications in micro-endoscopes and optical microscopy.
Researchers developed SDCBench, a benchmark suite that evaluates workload co-location in datacenters, enabling cloud tenants to understand performance isolation ability and choose their best-fitted cloud services. The tool also helps cloud providers improve service quality to increase revenue.
Graph computing studies the human world's graphs to analyze and compute them, uncovering hidden information in large-scale data. Key applications include real-time epidemiology analysis and targeted advertising.
Speckle-correlation imaging technique extends its field of view by considering a limited memory effect. Researchers proposed an algorithm to extrapolate the correlation in reconstruction, estimating object and decay simultaneously.
A new deep learning-based framework reduces autofocusing and image acquisition times for virtual staining of unlabeled tissue sections. The approach achieves high-quality staining with minimal loss of image sharpness and contrast.
Researchers propose a new solution called labeled von Neumann architecture (LvNA) to tackle challenges in cloud computing. LvNA incorporates label-powered control mechanisms to differentiate, isolate, and prioritize user-defined application requests, mitigating resource contentions and guaranteeing performance.
Researchers analyzed LNS's tail latency and low entropy benefits compared to mTCP and Linux network stacks. The study revealed that fulldatapath prioritized processing and full-path zero-copy are primary factors for high performance, improving tail latency by up to 5.5 times.
Current AI models are restricted by a lack of experience in real-world environments, despite achieving significant advancements in virtual settings. Researchers are now exploring ways to bridge this gap with foundation models that can operate in physical spaces.
Researchers have developed a novel global-to-local design approach to compose heterogeneous swarms of robots, enabling them to achieve collective behavior. The system allows users to define target behaviors by changing the number and position of distribution's modes, enabling swarms to adapt autonomously.