Research on machine learning and AI, now a key technology in virtually every industry and business, is far too extensive for anyone to read. This column, Perceptron, aims to collect some of the most relevant recent discoveries and articles – particularly in the field of, but not limited to, artificial intelligence – and explain why they matter.
A “audibleThat uses sonar to read facial expressions was one of the projects that caught our attention in recent weeks. So did ProcTHOR, a framework from the Allen Institute for AI (AI2) that procedurally generates environments that can be used to train real robots. Among the other highlights, Meta created an AI system which can predict the structure of a protein based on a single amino acid sequence. And researchers at MIT developed new hardware which they claim provides faster computations for AI with less energy.
Developed by a team at Cornell, the “earable” looks like a large pair of headphones. Loudspeakers send acoustic signals to the side of a wearer’s face, while a microphone picks up the barely perceptible echoes from the nose, lips, eyes and other facial features. These “echo profiles” allow the earable to capture movements such as eyebrows and eyes that shoot, which an AI algorithm translates into full facial expressions.
The earable has a few limitations. It only lasts three hours on battery power and has to transfer processing to a smartphone, and the echo-translating AI algorithm has to train on 32 minutes of facial data before it can recognize facial expressions. But the researchers argue that it’s a much tighter experience than the recorders traditionally used in animations for movies, TV and video games. For example, for the mystery game LA Noire, Rockstar Games built an installation with 32 cameras aimed at each actor’s face.
Perhaps Cornell’s earpiece will one day be used to create animations for humanoid robots. But those robots will first have to learn to navigate through a room. Fortunately, AI2’s ProcTHOR takes a step (no pun intended) in this direction, creating thousands of custom scenes, including classrooms, libraries, and offices in which simulated robots must complete tasks such as picking up objects and moving furniture.
The idea behind the scenes, which have simulated lighting and include a subset of a huge range of surface materials (e.g. wood, tile, etc.) and household objects, is to expose the simulated robots to as much variety as possible. It is an established theory in AI that performance in simulated environments can improve the performance of real systems; Autonomous car companies like Alphabet’s Waymo simulate entire neighborhoods to fine-tune the behavior of their real cars.
As for ProcTHOR, AI2 claims in a paper that scaling the number of training environments consistently improves performance. That bodes well for robots on their way to homes, workplaces and elsewhere.
Training these types of systems naturally requires a lot of computing power. But that may not be the case forever. Researchers at MIT say they’ve created an “analog” processor that can be used to create super-fast networks of “neurons” and “synapses,” which in turn can be used to perform tasks such as recognizing images, translate languages and more.
The researchers’ processor uses “protonic programmable resistors” arranged in an array to “learn” skills. Increasing and decreasing the electrical conductivity of the resistors mimics the strengthening and weakening of synapses between neurons in the brain, part of the learning process.
Conduction is controlled by an electrolyte that controls the movement of protons. When more protons in a channel are pushed into the resistor, the conductivity increases. When protons are removed, conduction decreases.
An inorganic material, phosphosilicate glass, makes the MIT team’s processor extremely fast because it contains nanometer-sized pores whose surfaces provide the perfect pathways for protein diffusion. As an added benefit, the glass can run at room temperature and is not damaged by the proteins as they move through the pores.
“Once you have an analog processor, you’re no longer training networks that everyone else is working on,” said lead author and MIT postdoc Murat Onen in a press release. “You’re going to train networks with unprecedented complexity that no one else can afford, and therefore outperform them all. In other words, this is not a faster car, this is a spacecraft.”
Speaking of acceleration, machine learning is now being deployed managing particle accelerators, at least in experimental form. At the Lawrence Berkeley National Lab, two teams showed that ML-based simulation of the full machine and beam gives them a highly accurate prediction that is as much as 10 times better than ordinary statistical analysis.
“If you can predict the properties of the beam with an accuracy that exceeds their fluctuations, you can use the prediction to improve accelerator performance,” says Daniele Filippetto of the lab. It’s no small feat to simulate all the physics and equipment involved, but surprisingly the early efforts of the various teams to do so yielded promising results.
And at Oak Ridge National Lab, an AI-powered platform lets them do hyperspectral computed tomography using neutron scattering, to find optimal… maybe we should just let them explain.
In the medical field, there is a new application of machine learning-based image analysis in the field of neurology, where researchers from University College London have trained a model to detect early signs of epilepsy-causing brain lesions.
A common cause of drug-resistant epilepsy is what is known as focal cortical dysplasia, an area of the brain that has developed abnormally but for some reason doesn’t appear clearly abnormal on MRI. Early detection can be very helpful, so the UCL team trained an MRI inspection model called Multicenter Epilepsy Lesion Detection on thousands of samples of healthy and FCD-affected brain regions.
The model was able to detect two-thirds of the FCDs that were shown, which is actually quite good because the signs are very subtle. In fact, it found 178 cases where doctors couldn’t pinpoint FCD, but they could. Of course, the last word is up to the specialists, but a computer that suggests something is wrong can sometimes be enough to take a closer look and get a reliable diagnosis.
“We focused on creating an AI algorithm that was interpretable and could help doctors make decisions. Showing doctors how the MELD algorithm made its predictions was an essential part of that process,” said Mathilde Ripart of UCL.