Optical AI chips have struggled to implement a crucial algorithm used to train neural networks–backpropagation. But in a new paper in the journal Science, a team from Stanford University has described the first ever implementation of the training approach on a photonic chip.
The algorithm combines classical beam physics equations with machine-learning techniques to reduce the need for extensive data processing.
When the linear accelerator at SLAC National Accelerator Laboratory is operational, groups of approximately one billion electrons travel through metal pipes at almost the speed of light. These electron groups form the accelerator’s particle beam, which is utilized to investigate the atomic behavior of molecules, innovative materials, and numerous other subjects.
However, determining the actual appearance of a particle beam as it moves through an accelerator is challenging, leaving scientists with only a rough estimate of how the beam will behave during an experiment.
This post is also available in: 
עברית (Hebrew)
Robert William, who was wrongfully identified by an AI algorithm and subsequently arrested, is suing the Detroit Police Department for the traumatizing experience he and his family had experienced.
Back in January of 2020, Robert Williams, a Black man, was arrested in front of his wife and children for a robbery committed at a Shinola store in 2018.
A new algorithm developed by Chinese company Baidu Research is dramatically faster than prior methods and shown to boost the antibody response of mRNA vaccines by up to 128 times.
Baidu Research is the research arm of Baidu, one of the largest technology companies in China. Established in 2014, it has since then been involved in various research activities such as natural language processing (NLP), machine learning, computer vision, robotics, and other areas of artificial intelligence.
A small team of chemists at the Russian Academy of Sciences, has found that metal atoms, not nanoparticles, play the key role in catalysts used in fine organic synthesis. In the study, reported in the Journal of the American Chemical Society, the group used multiple types of electron microscopy to track a region of a catalyst during a reaction to learn more about how it was proceeding.
Prior research has shown that there are two main methods for studying a reaction. The first is the most basic: As ingredients are added, the reaction is simply observed and/or measured. This can be facilitated through use of high-speed cameras. This approach will not work with nanoscale reactions, of course. In such cases, chemists use a second method: They attempt to capture the state of all the components before and after the reaction and then compare them to learn more about what happened.
This second approach leaves much to be desired, however, as there is no way to prove that the objects under study correspond with one another. In recent years, chemists have been working on a new approach: Following the action of a single particle during the reaction. This new method has proven to have merit but it has limitations as well—it also cannot be used for reactions that occur in the nanoworld. In this new effort, the researchers used multiple types of electron microscopy coupled with machine-learning algorithms.
Researchers from the University of Aberdeen develop an AI algorithm to detect planetary craters with high accuracy, efficiency, and flexibility.
A team of scientists from the University of Aberdeen has developed a new algorithm that could revolutionize planetary studies. The new technology enables scientists to detect planetary craters and accurately map their surfaces using different data types, according to a release.
Continue reading “AI-powered crater detection algorithm to unlock the secrets of the universe” »
This scientist made an algorithm to predict which artists succeed–without even looking at their art.
Greater scale and symbolic models are necessary before AI and machine learning can meet big challenges like breaking the best encryption algorithms.
Researchers have taught an AI to make artificial genomes — possibly overcoming the problem of how to protect people’s genetic information while also amassing enough DNA for research.
Generative adversarial networks (GANs) pit two neural networks against each other to produce new, synthetic data that is so good it can pass for real data. Examples have been popping up all over the web — generating pictures and videos (a la “this city does not exist”). AIs can even generate convincing news articles, food blogs, or human faces (take a look here for a complete list of all the oddities created by GANs).
Now, researchers from Estonia are going more in-depth with deepfakes of human DNA. They created an algorithm that repeatedly generates the genetic code of people that don’t exist.
Advances in quantum computation for electronic structure, and particularly heuristic quantum algorithms, create an ongoing need to characterize the performance and limitations of these methods. Here we discuss some potential pitfalls connected with the use of hardware-efficient Ansätze in variational quantum simulations of electronic structure. We illustrate that hardware-efficient Ansätze may break Hamiltonian symmetries and yield nondifferentiable potential energy curves, in addition to the well-known difficulty of optimizing variational parameters. We discuss the interplay between these limitations by carrying out a comparative analysis of hardware-efficient Ansätze versus unitary coupled cluster and full configuration interaction, and of second-and first-quantization strategies to encode Fermionic degrees of freedom to qubits.
