Deep Learning AI Agents Master Real-World Gameplay
▼ Summary
– Axiom, a new machine learning system by Verse AI, mimics human brain learning to master simple video games efficiently, offering an alternative to traditional neural networks.
– Axiom uses active inference, modeling game interactions with prior knowledge and updating expectations based on observations, inspired by the free energy principle.
– The free energy principle, developed by neuroscientist Karl Friston, underpins Axiom and emphasizes learning through action in the world, not just passive data absorption.
– Unlike deep reinforcement learning, Axiom achieves game mastery with fewer examples and less computational power, making it more efficient for certain tasks.
– Axiom’s architecture, described as a “digital brain,” is being tested in finance for real-time learning and could advance progress toward artificial general intelligence (AGI).
A groundbreaking machine learning system called Axiom is demonstrating remarkable efficiency in mastering simple video games by mimicking how the human brain processes and interacts with the world. Developed by Verse AI, this innovative approach challenges traditional artificial neural networks by incorporating prior knowledge about physical object interactions and refining its understanding through real-time observation—a method known as active inference.
Unlike conventional deep reinforcement learning, which relies on extensive trial-and-error training, Axiom achieves proficiency in games like drive, bounce, hunt, and jump with significantly fewer examples and less computational power. The system is rooted in the free energy principle, a neuroscientific theory pioneered by Karl Friston, Verses’ chief scientist. Friston emphasizes that true intelligence requires not just learning but also adapting actions within an environment, a capability Axiom aims to replicate.
François Chollet, creator of the ARC 3 benchmark for evaluating AI reasoning, praises the system’s originality. “This work aligns with crucial challenges in achieving artificial general intelligence (AGI),” he notes, highlighting the need for alternatives to mainstream approaches like large language models.
While modern AI has thrived on deep learning, powering everything from speech recognition to image generation, Axiom presents a fundamentally different architecture. Gabe René, CEO of Verses, describes it as a “digital brain” capable of real-time learning with greater accuracy and efficiency. Early adopters, including a financial firm, are already testing its potential for market modeling.
Interestingly, the free energy principle traces its origins to Geoffrey Hinton, a deep learning pioneer and Turing Award winner who once collaborated with Friston. This connection underscores how diverse perspectives continue to shape AI’s evolution.
For those intrigued by Friston’s theories, further exploration reveals their influence on cutting-edge consciousness research, offering fresh insights into both artificial and biological intelligence.
(Source: Wired)
