Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
A new machine The learning approach that is inspired by how the human brain seems to model and discover the world has proven to be able to master a certain number of simple video games with impressive efficiency.
The new system, called Axiom, offers an alternative to artificial neural networks that dominate modern AI. Axiom, developed by a software company called AI AI, is equipped with previous knowledge on how objects physically interact with each other in the world of game. It then uses an algorithm to model the way it expects that the game acts in response to the entry, which is updated according to what it observes – a process called in active inference.
The approach is inspired by the principle of free energy, a theory that seeks to explain intelligence by using principles drawn from mathematics, physics and information theory as well as biology. The principle of free energy has been developed by Karl Friston, a renowned neuroscientist who is chief scientist in the verses of the “cognitive computing” company.
Friston told me on videos from his home in London that the approach can be particularly important to build AI agents. “They must support the kind of cognition we see in real brains,” he said. “This requires consideration, not just the ability to learn things, but also to know how you act in the world.”
The conventional approach to learning to play games is to train neural networks through what is known as a deep reinforcement learning, which involves experimenting and refining their parameters in response to positive or negative feedback. The approach can produce superhuman game algorithms, but it requires a lot of experimentation to work. Axiom Masters various simplified versions of popular video games called Drive, Bounce, Hunt and Jump using far fewer examples and less computing power.
“The general objectives of the approach and some of its key features follow what I consider as the most important problems on which to focus on going to Act,” explains François Chollet, an AI researcher who has developed Arc 3, a reference designed to test the capacities of modern AI algorithms. Chollet also explores new approaches to automatic learning and uses its reference to test the capacity of models to learn to solve unknown problems rather than simply imitating previous examples.
“The work seems very original to me, which is great,” he says. “We need more people who try new ideas far from the beaten track of large languages models and models of reasoning language.”
Modern AI is based on artificial neural networks which are roughly inspired by the wiring of the brain but which work in a fundamentally different way. Over the past decade and a little, Deep Learning, an approach that uses neural networks, allowed computers to do all kinds of impressive things, including transcribing speech, recognizing faces and generating images. More recently, of course, in -depth learning has led to important models that are spread and more and more capable chatbots.
Axiom, in theory, promises a more effective approach to building AI from zero. It could be particularly effective for the creation of agents who must learn effectively from experience, explains Gabe René, CEO of verses. René says that a financial company has started to experiment with business technology as a means of modeling the market. “This is a new architecture for AI agents that can learn in real time and more precise, more efficient and much smaller,” explains René. “They are literally conceived as a digital brain.”
A little ironically, since the axiom offers an alternative to modern AI and the deep learning, the principle of free energy was initially influenced by the work of the British Canadian computer scientist Geoffrey Hintonwho received both the Turing price And the Nobel Prize for his pioneering work on in -depth learning. Hinton was a colleague from Frriston at the University College London for years.
To find out more about Friston and the principle of free energy, I highly recommend This 2018 wired function article. Friston’s work has also influenced a New exciting theory of consciousnessDescribed in a wired book revised in 2021.
