AI that sees with sound, learns to walk and predicts seismic physics

Research into machine learning and artificial intelligence, now key technologies in almost every industry and business, is far too extensive for anyone to read. The purpose of this column, Perceptron, is to collect some of the most important recent discoveries and articles – especially, but not limited to, in the field of artificial intelligence – and explain why they are important.

This month, engineers at Meta detailed two recent innovations from the depths of the company’s research labs: an artificial intelligence system that compresses audio files and an algorithm that can speed up protein folding by 60 times. Elsewhere, MIT scientists revealed that they are using spatial acoustic information to help machines better imagine their environment, simulating how a listener would hear sound from any point in space.

Meta’s compression work doesn’t exactly reach uncharted territory. Last year, Google announced Lyra, a neural audio codec trained to compress low-bitrate speech. But Meta claims its system is the first to work for CD-quality stereo audio, making it useful for commercial applications such as voice calls.

Meta audio compression

Meta audio compression

Architectural drawing of Meta’s AI audio compression model. Image Credits: Mint

Using artificial intelligence, Meta’s compression system, called Encodec, can compress and decompress audio in real time on a single CPU core at speeds ranging from about 1.5 kbps to 12 kbps. Compared to MP3, Encodec can achieve about 10 times the compression rate at 64 kbps without noticeable quality loss.

The researchers behind Encodecom say that human evaluators preferred the quality of Encodec-processed audio over Lyra-processed audio, suggesting that Encodec could eventually be used to provide better quality audio in situations where where bandwidth is limited or too high.

As for Meta’s protein-folding work, it has less immediate commercial potential. But it could lay the foundation for important scientific research in the field of biology.

Folding of meta proteins

Folding of meta proteins

Protein structures predicted by Meta’s system. Image Credits: Mint

Also Read :  The weekend’s best deals: A bunch of Apple devices, Sony headphones, 4K TVs, and more

Meta says its AI system, ESMFold, has predicted the structures of about 600 million proteins from bacteria, viruses and other microbes that have yet to be identified. That’s more than three times the 220 million structures that Alphabet-backed DeepMind was able to predict earlier this year, covering almost all proteins from known organisms in DNA databases.

Meta’s system is not as accurate as DeepMind’s. Of the roughly 600 million proteins he made, only a third were of “high quality.” However, it is 60 times faster at structure prediction, which allows it to extend structure prediction to much larger protein databases.

In order not to draw too much attention to Meta, the company’s artificial intelligence department also this month detailed a system designed for mathematical reasoning. Researchers at the company say their “neural problem solver” has learned to generalize to new, different kinds of problems from a dataset of successful mathematical proofs.

Meta is not the first to build such a system. OpenAI has developed its own, called Lean, which it announced in February. In addition, DeepMind has experimented with systems that can solve challenging mathematical problems in the study of symmetries and knots. But Meta claims its neural problem solver has been able to solve five times more international math Olympiads than any previous AI system and has beaten other systems on widely used math benchmarks.

Meta notes that AI for solving math could benefit the fields of software verification, cryptography, and even space.

Focusing on the MIT work, researchers there have developed a machine learning model that can capture how sounds in a room travel through space. By modeling the acoustics, the system can learn the geometry of the room from audio recordings, which can then be used to produce visual representations of the room.

The researchers say the technology could be used for virtual and augmented reality software or robots that need to navigate complex environments. In the future, they plan to improve the system so that it can be generalized to new and larger scenes, such as entire buildings or even entire cities.

Also Read :  Time to get spooky: An alien experience

At Berkeley’s robotics department, two separate teams are accelerating the speed at which a four-legged robot can learn to walk and perform other tricks. One team sought to combine the best work from several other advances in reinforcement learning to enable a robot to go from a blank slate to robustly walking on uncertain terrain in just 20 minutes in real time.

“Perhaps surprisingly, we find that with several careful design decisions in terms of task setup and algorithm implementation, it is possible for a quadruped robot to learn to walk from scratch with deep RL in less than 20 minutes, in a variety of environments and “Crucially, it does not require new algorithmic components or any other unexpected innovation,” the researchers write.

Instead, they choose and combine some cutting-edge approaches and get amazing results. You can read the contribution here.

Introducing a robot dog from EECS professor Pieter Abbeel’s lab in Berkeley, California in 2022. (Photo courtesy of Philipp Wu/Berkeley Engineering)

Another movement learning project from the lab of (TechCrunch friend) Pieter Abbeel has been described as “imagination training”. The robot is set up so that it can try to predict how its actions will play out, and while it’s pretty helpless at first, it quickly gains more knowledge about the world and how it works. This leads to a better forecasting process, which leads to better knowledge, and so on in the feedback until he walks in less than an hour. He learns just as quickly to recover from being pushed or otherwise “excited,” as the jargon says. Their work is documented here.

Work with a potentially more immediate application came earlier this month from Los Alamos National Laboratory, where researchers developed a machine learning technique to predict the friction that occurs during earthquakes — making it possible to predict earthquakes. Using a language model, the team says they were able to analyze the statistical characteristics of seismic signals emitted by a fault in a laboratory earthquake machine to predict the timing of the next earthquake.

Also Read :  AI users regard it as a co-worker

“The model is not limited by the physics, but it predicts the physics, the actual behavior of the system,” said Chris Johnson, one of the research leaders on the project. “We are now making a future prediction from past data that goes beyond describing the current state of the system.”

Dreamtime

Dreamtime

Image Credits: Dreamtime

The researchers say that applying the technique to the real world is challenging because it is unclear whether there is enough data to train a prediction system. Still, they are optimistic about applications that could include predicting damage to bridges and other structures.

Last week, MIT researchers warned that neural networks used to simulate real neural networks should be scrutinized for training biases.

Neural networks, of course, are based on the way our own brains process and signal information, reinforcing certain connections and combinations of nodes. But this does not mean that synthetic and real work the same. Indeed, the MIT team found that neural network-based simulations of reticular cells (part of the nervous system) produced similar activity only when carefully constrained by their creators. If they were allowed to manage themselves, just like real cells, they did not produce the desired behavior.

This is not to say that deep learning models are useless in this field – far from it, they are very valuable. But as Professor Ila Fiete said in a school news release, “they can be powerful tools, but one must be very careful in interpreting them and determining whether they really make new predictions or even shed light on what it is that the brain is optimizing for.” “

Source

Leave a Reply

Your email address will not be published.

Related Articles

Back to top button