Welcome to Perceptron, TechCrunch’s weekly roundup of AI news and research from around the world. Machine learning is a key technology in virtually every industry right now, and there’s far too much going on for anyone to keep track of. This column aims to bring together some of the most interesting recent discoveries and papers in the field of artificial intelligence – and explain why they matter.

(Formerly known as Deep Science; see previous editions here.)

This week’s roundup begins with a pair of groundbreaking studies from Facebook/Meta. The first is a collaboration with the University of Illinois at Urbana-Champaign that aims to reduce the amount of emissions from concrete production. Concrete accounts for around 8% of carbon emissions, so even a small improvement could help us meet climate goals.

What the Meta/UIUC team did was train a model on over a thousand concrete mixes, which differed in the proportions of sand, slag, crushed glass and other materials (you can see a more photogenic concrete sample at the top). By finding the subtle trends in this data set, he was able to produce a number of new formulas that optimize for both strength and low emissions. The winning formula was found to have 40% fewer emissions than the regional standard, and satisfied… well, some strength requirements. This is extremely promising, and follow-up studies in the field should get the ball rolling again soon.

The second Meta study concerns the modification of the functioning of language models. The company wants to work with neural imaging experts and other researchers to compare how language patterns compare to real brain activity during similar tasks.

In particular, they are interested in the human ability to anticipate words well in advance of the current one while speaking or listening – such as knowing that a sentence will end in a certain way, or that there is a “but” to come. AI models are getting really good, but they still mostly work by adding words one by one like Lego bricks, sometimes looking back to see if it makes sense. They’re just getting started but they already have some interesting results.

Back on the cutting edge of materials, Oak Ridge National Lab researchers are getting into the fun of formulating AI. Using a dataset of whatever quantum chemical calculations, the team created a neural network that could predict material properties, but then flipped it to be able to input properties and have it suggest materials.

“Instead of taking a material and predicting its given properties, we wanted to choose the ideal properties for our purpose and work backwards to design those properties quickly and efficiently with a high degree of confidence. This is called reverse design,” said ORNL’s Victor Fung. It seems to have worked – but you can check for yourself by running the code on Github.

Picture credits: ETHZ

Concerned with physical predictions on a whole other scale, this ETH project estimates treetop heights around the globe using data from ESA’s Copernicus Sentinel-2 satellites (for optical imagery) and NASA’s GEDI (orbital laser telemetry). Combining the two in a convolutional neural network yields an accurate global map of tree heights up to 55 meters high.

Being able to do this type of regular biomass survey on a global scale is important for climate monitoring, as NASA’s Ralph Dubayah explains: “We just don’t know how tall the trees are at the global scale. We need good world maps of tree locations. Because every time we cut down trees, we release carbon into the atmosphere, and we don’t know how much carbon we release.

You can easily browse the data as a map here.

This DARPA project is also about landscapes and involves creating very large-scale simulated environments for virtual autonomous vehicles to drive through. They awarded the contract to Intelalthough they may have saved some money by contacting the creators of the game snow runnerwhich basically does what DARPA wants for $30.

Picture credits: Intel

RACER-Sim’s goal is to develop off-road AVs that already know what it’s like to rumble over rocky desert and other challenging terrain. The 4-year program will first focus on creating the environments, building models in the simulator, and later on transferring skills to physical robotic systems.

In the field of pharmaceutical AI, which currently has around 500 different companies, MIT has a sensible approach in a model that only suggests molecules that can actually be made. “Models often suggest new molecular structures that are difficult or impossible to produce in the lab. If a chemist can’t actually make the molecule, its disease-fighting properties can’t be tested. »

Sounds cool, but can you do it without unicorn horn powder?

The MIT model “ensures that molecules are made up of materials that can be purchased and that the chemical reactions that occur between these materials follow the laws of chemistry.” It’s a bit like what Molecule.one does, but integrated into the discovery process. It would certainly be nice to know that the miracle drug your AI offers doesn’t require any pixie dust or other exotic materials.

Another job by MIT, the University of Washington, and others is teaching robots to interact with everyday objects — something we all hope will become mainstream in the next two decades, because some of us don’t don’t have a dishwasher. The problem is that it’s very hard to tell exactly how people interact with objects, because we can’t relay our data in high fidelity to train a model. So there is a lot of data annotation and manual labeling involved.

The new technique focuses on observing and inferring 3D geometry up close, so it only takes a few instances of a person grabbing an object for the system to learn to do it itself. Normally this could take hundreds of examples or thousands of repetitions in a simulator, but this one required only 10 human demonstrations per object in order to effectively manipulate that object.

Picture credits: MIT

It achieved an 85% success rate with this minimal training, much better than the base model. It’s currently limited to a handful of categories, but researchers hope it can be generalized.

The last place this week is a bit Promising work from Deepmind on a multimodal “visual language model” that combines visual knowledge with linguistic knowledge so that ideas like “three cats sitting on a fence” have a sort of cross-representation between grammar and imagery. That’s how our own minds work, after all.

Flamingo, their new “general purpose” model, can do visual identification but also engage in dialogue, not because it’s two models in one but because it combines language and visual understanding. As we have seen in other research organisations, this type of multimodal approach produces good results but is still very experimental and computationally intensive.