Seeing through clouds. "Over the last few years, several cloud removal techniques have been developed: Sentinel-2 Cloudless, Mapbox Cloudless Atlas and Google's Cloudless Satellite Map. These techniques are sifting through multiple years of images to create cloud-free scenes." eng.ruumi.io/post/seeing-throu

Video from Mars. NASA put a bunch of cameras on the spacecraft so all of us could experience the landing as close to actually being there as possible. We can look at the parachute inflating, the descent with the heat shield, the separation of the rover and descent stage, the touchdown at the landing site, with the engines blowing dust around on the surface. We can look upward from the rover to the descent stage. youtube.com/watch?v=gYQwuYZbA6

The mystery of where fast radio bursts (FRBs) come from looks like it's solved. One of the greatest mysteries in astronomy. A fast radio burst was observed, and the source was tracked down to a magnetar. The magnetar goes by the unmemorable name "J1935+2154". It's in the Milky Way about 30,000 light years from here, located in the sky in the constellation Vulpecula. english.cas.cn/head/202102/t20

Photons can fuse. Photons -- particles of light, not protons. This is something I did not know was possible. But apparently theoretical physics predicted it and it has now been observed at the LHC. When photons fuse, they form leptons. Electrons are leptons, but they are not the only subatomic particle in the category. There are also muons and tau particles. "If two photons collide, the result could be an electron-positron pair or a muon-antimuon pair." press.ifj.edu.pl/en/news/2021/

The insane engineering of the Perseverance rover. Instead of having instruments for determining the chemical composition of the Martian soil, the system stores sample tubes where they will be transferred to a spot where a future rover will retrieve them in 2026, transferring them to a return rocket that will bring them to orbit, where they will be transferred to another spacecraft that will return the Martian soil back to Earth. youtube.com/watch?v=yqqaW8DCc-

"DNA origami" for placing molecules on semiconductor chip not only by location but also in a specific orientation. "As a proof-of-concept, engineers arranged more than 3,000 glowing moon-shaped nanoscale molecular devices into a flower-shaped instrument for indicating the polarization of light. Each of 12 petals pointed in a different direction around the center of the flower, and within in each petal about 250 moons were aligned to the direction of the petal." caltech.edu/about/news/enginee

Human brain has over 400 miles of total vasculature. The researchers "focused on a cell type called a pericyte. Pericytes cover the capillary surface and are broadly categorized as a vascular mural cell." "When we turned on pericytes, we observed a direct effect of decreased capillary blood flow. When we removed pericytes by ablating them with focused light, we observed an increase in capillary blood flow." pulse.seattlechildrens.org/unt

Ice is nice. But you might not know there are lots of different forms of ice, which you wouldn't know because they don't normally occur here, on Earth, with the temperature and pressure ranges typical here. "In snowflakes or ice cubes, the oxygen atoms are arranged hexagonally. This ice form is called ice one (ice I). "18 crystalline forms of ice were known so far, which differ in the arrangement of their atoms." uibk.ac.at/newsroom/new-crysta

"Gravity might play a bigger role in the formation of elementary particles than scientists used to believe." "Due to their small size, the gravitational interaction between elementary particles (electrons, protons, and neutrons) is weak compared to Coulomb forces -- attraction and repulsion determined by charge. For example, negatively charged electrons move around the atomic nucleus that contains positively charged protons." eng.rudn.ru/media/news/science

"This robot doesn't need electronics to be able to walk. In fact it doesn't need electronics at all. The robot is controlled and powered by pressurized air. The movement of its four legs is driven by a series of valves that let pressurized air in and out in specific sequences. A custom valve allows the robot to change directions. The system is inspired by neural circuits found in nature that generate rhythmic patterns." youtube.com/watch?v=X5caSAb4kz

Experiments with an artificial model that mimics real neurons. Neurons have a precise action threshold. Action pulse that travels through the axon also looks like a digitally generated square wave. Once a neuron fires, it normally cannot fire again for approx. 10 msec. "Time-based behavior occurs when multiple action pulses pass through a single synapse in rapid succession, close enough to overlap." linkedin.com/pulse/time-ai-lea

3D objects from 2D photos. This is a neural network system that takes a photograph of an object (and actually requires an additional "object mask" to separate the desired object from the background) and outputs a 3D triangle mesh with texture and the corresponding camera pose. The trick employed here is to do an "adversarial" approach, analogous to generative adversarial networks (GANs). In fact the training system is the same as GANs. judyye.github.io/ShSMesh/

Automatic pool stick. Or: what happens when a nerd is married to a pool shark. The robot pool stick is controlled with 6 rods, reminiscent of a pick-and-place machine. Compressed air is used to actually hit the ball. A camera on the ceiling looks at the pool table to figure out where the balls are. A projector tells you where to put the robotic pool stick to make the shot. Even with all that, it doesn't work. To get it to work, he had to ... youtube.com/watch?v=vsTTXYxydO

Most memory studies rely on recalling text exactly, but these researchers decided to flip that around and study how humans remember the "gist" of an experience instead. They went through episodes of an episode of the TV show Sherlock moment-by-moment and labeled them by topic. They ran these topics through a word embedding system. news.dartmouth.edu/news/2021/0

Superintelligence cannot be contained. Mathematically proven! Ok, so the proof starts with Turing's Halting Problem, which is "undecidable". They extended this to the realm of AI and the question of harming humans. Imagine you have code for a robot that can harm humans. But before it runs the instructions to harm humans, it runs a program that could be endless. Since determining if the program terminates is undecidable, determining if the AI harms humans is undecidable. mpib-berlin.mpg.de/computer-sc

"Applying worm blob dynamics to swarm robotics." Researcher "applied the principles observed in the worms to small robotic blobs composed of 'smart active particles,' six 3D-printed robots with two arms and two sensors allowing them to sense light. She added a mesh enclosure and pins to arms that allowed these 'smarticles' to be entangled like the worms and tested a variety of gaits and movements that could be programmed into them. youtube.com/watch?v=615LmMNBFJ

Simulating spider web construction with a virtual spider. The spider being simulated here is the garden cross spider. The virtual spider is "evolved" to make virtual spider webs using genetic algorithms. The "genes" encode "rules" and the "rules" are executed using an "interpreter" that works analogously to a programming language. Inspired by how real spider can lose a leg and regrow it and still make a web. theconversation.com/spider-leg

"Machine learning accelerated computational fluid dynamics". In this machine-learning-boosted simulation, in each time-step, the neural network generates a vector at each grid location based on the current velocity field of the fluid being simulated there. The neural network is a convolutional neural network, by the way, to ensure that the fluid is simulated the same way everywhere ("translation invariance"). arxiv.org/abs/2102.01010

"A statistician teaches deep learning". He says, "Statisticians have different training and instincts than computer scientists." "One contrast between statisticians and deep learning computer scientists is that we generally make predictions and classifications using linear combinations of basis elements -- this is the cornerstone of nearly all regression and much classification, including support vector machines, boosting, bagging, and stacking." arxiv.org/abs/2102.01194

Scaling laws for transfer learning. Transfer learning is when you take a neural network, usually one someone else has already built and that works for whatever they built it for, and train it on your own data. "Scaling laws", on the other hand, is when increasing the size of one thing increases something else by some exponential value, although the exponent can be a fraction. arxiv.org/abs/2102.01293

Show older
Mastodon

Server run by the main developers of the project 🐘 It is not focused on any particular niche interest - everyone is welcome as long as you follow our code of conduct!