A new artificial intelligence model developed in Russia is said to be capable of learning new actions all by itself
A group of Russian scientists has developed a new artificial intelligence model capable of self-adaptation to new tasks and context without additional interference by humans.
The model has allowed the developers to overcome one of the key limitations in contextual machine learning, a team from the T-Bank (formerly, Tinkoff Bank) AI Research Lab and the Moscow-based Artificial Intelligence Research Institute (AIRI) said in a paper published online.
The previously existing models, while capable of learning to perform new tasks when fed enough data, were still limited by a pre-defined and fixed set of actions, the researchers explained. The introduction of a new “action space” would then require a new set of data, often quite extensive, as well as model re-learning. This limitation would make re-adaptation a costly endeavor for some applications, they said.
The team took a specific machine learning model called Algorithm Distillation (AD) and further modified it to meet the set goal. The AD method trains AI to perform tasks by autoregressively predicting actions while using its learning history dataset as context.
The Russian model was dubbed ‘Headless-AD’ and presented at the International Conference on Machine Learning in Vienna this week.
The Headless-AD approach meant the model acquired the ability to learn and deploy new actions in response to new tasks without the need for additional input or re-learning by a human.
According to the team, their AI was capable of performing five times more actions than it was initially taught. The researchers said this could have broad applications from space technologies to smart home assistants.
Such a model could be taught some basic actions on generalized data and then adapt to the specific conditions of a particular context, according to the team’s report. Some Russian media then suggested that the new AI model might be just smart enough to pass the so-called ‘coffee test’ reportedly failed by the now famous ChatGPT.
First put forward by Apple co-founder Steve Wozniak, the test requires an AI machine to “go into an average American household and figure out how to make coffee, including identifying the coffee machine, figuring out what the buttons do, finding the coffee cabinet, etc.”
The problem for most AI is that although average households have much in common, they still are all slightly different, which would normally require an AI machine to be taught on a particular dataset related to a particular household to be able to perform the task there.
Doing the same task in a new household would require re-learning on a new dataset. However, the self-adaptable Russian AI could potentially be up to the task, the reports suggested.
