AI and machine learning research are currently hot subjects in nearly every business industry. Engineers at Meta unveiled two recent discoveries from the company’s research laboratories this month: an AI system that compresses audio files and a method that can speed up AI performance for protein folding by 60x.
To offer machines a better feel of their surroundings, MIT researchers announced that they are modeling how a listener might hear a sound from anywhere in a room. Last year, Google launched Lyra, a neural audio codec designed to compress speech at low bit rates. However, Meta claims that its technology is the first to deliver CD-quality stereo audio, making it appropriate for business applications such as voice conversations.
Encodec Technology
Meta’s Encodec compression technology employs artificial intelligence to provide real-time audio compression and decompression at rates ranging from 1.5 kbps to 12 kbps on a single CPU core. In comparison to MP3, Encodec can achieve a compression rate of roughly 10x at 64 kbps without reducing quality appreciably.
According to the researchers who created Encodec, human listeners preferred the Encodec-processed audio quality over the Lyra-processed audio, implying that Encodec might be used to deliver higher-quality audio in situations when bandwidth is limited or expensive.
Meta’s protein folding research has limited direct commercial relevance. However, it may lay the groundwork for substantial biological research in the future. ESMFold, Meta’s AI system, predicted the structures of around 600 million uncharacterized proteins from bacteria, viruses, and other microbes, according to Meta.
The Meta system is less accurate. Only one-third of the 600 million proteins manufactured were of “high quality.” However, it may be used on much bigger protein databases because it predicts structures 60 times faster. To avoid drawing too much attention to Meta, the company’s AI section introduced this month a system that employs mathematics to reason.
According to business researchers, their “neural issue solver” learned to generalize to new, distinct forms of problems from a dataset of successful mathematical proofs. Meta did not invent such a system. OpenAI developed its own, dubbed Lean, and released it in February. Separately, DeepMind has experimented with computers that can solve tough mathematical challenges in its study of symmetry and knots.
MIT research
In terms of MIT research, researchers developed a machine-learning model that can show how noises in space will ricochet throughout the room. The technology may use sound recordings to simulate a room’s acoustics and create visual representations.
The researchers suggest using the method for robots on rough terrain and virtual and augmented reality applications. They hope to expand the technique to include entire buildings and cities.
Two distinct teams at Berkeley’s robotics department are accelerating the rate at which a quadrupedal robot can learn to walk and perform other tricks. One team attempted to merge the best-of-breed work from numerous previous improvements in reinforcement learning to enable a robot to develop from a completely blank slate to robust walking on uncertain terrain in under 20 minutes in real time.
Surprisingly, we discover that a quadrupedal robot can learn to walk with deep RL in under 20 minutes, across a variety of locales and surface types, with a few smart design considerations in terms of task setup and algorithm implementation. The researchers underline that this does not demand new algorithmic parts or any other unexpected innovation. They, on the other hand, select and mix a number of cutting-edge tactics to achieve amazing results.
According to the researchers, it is challenging to apply the technique in practice because it is uncertain whether there is enough data to train the forecasting system. They remain optimistic about the applications, some of which may include anticipating damage to bridges and other structures.
What is AI?
Unlike animals and humans, machines exhibit artificial intelligence (AI). Several key sub-fields of AI research (as opposed to AI itself) have adopted working definitions of the intelligent agents field of study, which refers to any system that senses its surroundings using an AI component and performs actions that maximize its chances of attaining its goals.
While intelligent agents as systems that use artificial intelligence are an essential use of AI, many AI systems do not execute any procedural (hard-coded) steps with the AI’s outputs at all, such as computer vision, speech recognition, or recommender systems.
Previously, the phrase “artificial intelligence” was used to describe robots that mimic and demonstrate “human” cognitive skills associated with the human mind, such as “learning” and “problem-solving.” Key AI researchers now characterize AI as rationality and acting rationally, which doesn’t limit intelligence.
Also Read : Boult Audio releases X30, X50 TWS Earbuds at Rs 999
178 total views, 1 views today