A Deep Learning Approach Can Enhance AI Agents’ Interaction with the Real World

An innovative machine learning method inspired by the human brain’s mechanisms for modeling and understanding the world has demonstrated remarkable proficiency in mastering several basic video games.

This new system, named Axiom, presents an alternative to the prevalent artificial neural networks found in today’s AI landscape. Developed by Verse AI, Axiom incorporates prior knowledge of how objects interact within the game environment. It employs an algorithm to predict game responses to inputs, which it refines based on observations—a technique known as active inference.

The method is grounded in the free energy principle, a theory that explains intelligence by integrating concepts from mathematics, physics, information theory, and biology. Karl Friston, a distinguished neuroscientist and chief scientist at the cognitive computing company Verses, was instrumental in developing this principle.

Friston stated during a video conversation from his London home that this approach could be crucial for developing AI agents. “They need to emulate the kind of cognition seen in real brains,” he emphasized. “This necessitates not just the ability to learn but also understanding how to act in the world.”

Traditionally, learning to play games entails training neural networks via deep reinforcement learning, which involves experimenting and adjusting parameters based on positive or negative feedback. This method can yield superhuman gaming algorithms but requires extensive experimentation for success. Axiom, however, excels at simplified adaptations of popular games known as drive, bounce, hunt, and jump, using significantly fewer examples and less computational power.

“The overall goals of this approach and its main features align with what I consider to be critical challenges for achieving AGI,” notes François Chollet, an AI researcher responsible for ARC 3, a benchmark created to assess modern AI capabilities. Chollet is also investigating novel machine learning methods and is utilizing his benchmark to evaluate models’ skills in solving unfamiliar problems rather than merely replicating prior examples.

“I find this work to be very original, which is fantastic,” he adds. “We need more researchers exploring new concepts beyond the traditional realms of large language models and reasoning models.”

Contemporary AI generally relies on artificial neural networks that are loosely inspired by the brain’s architecture yet operate fundamentally differently. In the past decade, deep learning has enabled computers to achieve impressive feats such as transcribing speech, recognizing faces, and generating images. Most recently, it has led to the development of large language models that drive increasingly capable chatbots.

Theoretically, Axiom offers a more efficient method for developing AI from the ground up. It may be particularly adept at constructing agents that learn effectively from experience, suggests Gabe René, CEO of Verses. René mentions that one finance company has started using their technology to model market behaviors. “It represents a new architecture for AI agents that learn in real time, making them more accurate, efficient, and compact,” René states. “They are essentially designed like digital brains.”

Ironically, despite Axiom being an alternative to contemporary AI and deep learning, the free energy principle was influenced by Geoffrey Hinton, a British-Canadian computer scientist acclaimed with both the Turing Award and the Nobel Prize for his pioneering contributions to deep learning. Hinton and Friston were colleagues at University College London for several years.

For additional insights on Friston and the free energy principle, I highly recommend this 2018 WIRED feature article. Friston’s research has also shaped an intriguing new theory of consciousness, which is outlined in a book reviewed by WIRED in 2021.