A recent study led by Michael Timothy Bennett from the Australian National University reveals that biological systems are more efficient than current artificial intelligence (AI) models. Published in the Journal of The Royal Society Interface, the research questions whether biological systems possess intrinsic intelligence superior to contemporary AI.
The study defines intelligence as the ability to adapt efficiently, using minimal resources to identify causes and effects in complex environments. Biological systems excel in adapting with significantly less data and energy compared to AI.
Bennett highlights that an organoid—a small collection of lab-cultivated cells—outperforms reinforcement learning systems in the classic video game Pong. Despite being designed under human-imposed conditions, the organoid's adaptability exemplifies biological efficiency.
In contrast, current machine learning algorithms require vast amounts of data and energy for tasks easily managed by humans and other organisms with minimal experience. This disparity raises questions about the foundational architecture of both systems.
A key feature of biological systems is their multiscale competitive architecture (MCA), where adaptation occurs at all levels, from cells to ecosystems. Each cell operates as an autonomous agent with its own 'policy' for interaction, forming organs that contribute to the organism's functional identity.
This decentralization allows biological systems to adapt more efficiently than conventional computational systems, which often rely on centralized decision-making. Bennett illustrates this with an analogy of drones: if a leading drone is damaged in a hierarchical swarm, the structure collapses. However, in a decentralized swarm, each drone can adapt and maintain group coherence.
The study also points out the limitations of current AI systems, which depend on static layers of abstraction that restrict adaptability. Once an architecture is designed, it remains fixed, hindering responses to new circumstances.
Biological systems can adjust their interaction 'window' with the environment, redefining rules at lower abstraction levels for better adaptation at higher levels. This dynamic optimization mechanism contributes to their superior efficiency.
Although biological systems are not flawless, as seen in cancer, where cells lose their connection to the organism's identity and pursue individual goals, the study notes that excessive restrictions can lead to fragmentation and systemic failures akin to biological cancer.
The principles of delegation and adaptability highlighted in this research could inform the design of cyber-physical systems, such as specialized hardware for more efficient learning. Bennett warns that rigorously regulated AI systems could face similar failures if restrictions inhibit adaptation.
The study references the UK's 'Safeguarded AI' program, which aims to develop safe and adaptive general AI systems. Bennett cautions that imposing too many restrictions could lead to the very failures regulations seek to prevent.
In conclusion, the intelligence of biological systems is intricately linked to their capacity for delegation and adaptation across various abstraction levels. This organizational model offers valuable insights for creating more robust and efficient artificial systems, laying the groundwork for future research that could transform our understanding and development of intelligent systems.