Quantum Leaps and Machine Minds
Addressing AI Emergent Properties and Advancing Explainability Through Interdisciplinary Research
We find ourselves in an era of breathtaking transformation, where it seems like Artificial Intelligence (AI) is etching itself into the very fabric of our daily lives. While there’s some hype today about the promises and perils of AI, the systems we’re building are remarkably impressive and, in the case of foundational models like large language models (LLMs), they’re beginning to exhibit a striking phenomenon known as emergent properties: surprising behaviors that spring forth not from explicit programming but the complex interplay of countless algorithmic components. In other words, large generative AI systems are giving us the most compelling empirical evidence of unexplainable emergent characteristics, manifesting themselves in both advantageous and misleading ways. For example, when a model like ChatGPT effectively navigates its “latent space” to correctly answer a question it wasn’t explicitly trained on, the utility of such emergent behavior is undeniable. On the other hand, having a persuasive dialogue with ChatGPT that’s later found to be hallucinatory rants can be a waste of time at best. In both instances, these behaviors diverge from the initial intentions of the human architects of these systems, therefore, fine-tuning them in the hopes of producing more predictable outcomes with fewer distortions without compromising their performance is easier said than done. As we navigate such complexities, it’s becoming clear that rather than controlling these systems in the classical sense, increasing our ability to orchestrate more precise behaviors may require the development of new mathematical models that can draw not only from the evolving body of research in machine learning but also the familiar grounds of quantum physics.
Quantum-Inspired Paradigms for Addressing Emergent AI Behaviors
Emergent properties in quantum systems are fascinating and have been a subject of extensive research. Like the emergent properties of large AI models, these are behaviors that arise from the collective dynamics of particles, which are not predictable from the simple sum of individual particle properties. Like the emergent “insights” of LLMs, such particle properties are new phenomena that "emerge" when considering the system as a whole. Examples of emergent phenomena in quantum systems include superconductivity (where materials conduct electricity with zero resistance), superfluidity (where liquids flow without viscosity), and quantum entanglement (where particles become interconnected in such a way that the state of one instantaneously influences the state of the other, regardless of distance).
One of the earliest influential theories of superconductivity is called Bardeen–Cooper–Schrieffer (BCS) theory. According to BCS, at temperatures below a critical temperature, electrons in a superconductor can form pairs, known as Cooper pairs, which can move through the crystal lattice of a superconductor without scattering off impurities and lattice vibrations that are the main causes of electrical resistance in ordinary conductors. Similarly, superfluidity is a state of matter in which a fluid, under certain conditions, can flow without friction. This means a superfluid can flow up and over the walls of a container and remain in motion indefinitely without any external force. This can actually be observed in liquid helium and certain ultra-cold atomic gases.
Understanding quantum emergent properties has had practical real-world implications. For example, the discovery of superconductivity and BCS theory were instrumental to the development of superconducting magnets used in Magnetic Resonance Imaging (MRI) scanners. Superconductors are also crucial in the construction of particle accelerators and are being explored for use in power transmission lines and next-generation quantum computers, where quantum entanglement is also a relevant feature. The study of superfluids has also led to advances in low-temperature technologies and has been crucial for investigating “exotic” states of matter like Bose-Einstein condensates.
Shifting Perspectives on AI Explainability
In both quantum mechanics and machine learning, we often have to deal with systems that are composed of many interacting parts. In quantum mechanics, these might be particles in a quantum field. In machine learning, these could be the numerous interconnected nodes and layers in a neural network. In quantum mechanics, advanced mathematical models like Quantum Field Theory and Many-Body Theory can be used to describe the complex interactions among quantum particles, helping predict system behaviors that underlie emergent properties. The development of similar formal mechanisms for predicting the emergent properties of large AI models can not only enable us to harness the full potential of these models but also serves as a vital safeguard against unintended consequences arising from undesirable emergent behaviors in real-world applications, especially in life sciences and healthcare. Therefore, predicting and interpreting emergent patterns within AI systems effectively calls for a shift in our conventional expectations of explainability. Although transparency and interpretability continue to be critical goals in AI development, a transition may be required from reliance on descriptive explanations. A shift towards formal mathematical model-based analysis can aid in deciphering the sophisticated dynamics of AI systems based on frameworks dedicated to the quantitative examination and portrayal of emergent AI behaviors.
Academic institutions play a pivotal function in nurturing these frameworks. Given their influential role in shaping novel research pathways, they can serve as a fertile ground for interdisciplinary research. For example, introduction of curricula combining mathematical modeling, quantum mechanics, computer science, and AI ethics could cultivate a body of experts equipped to decipher and predict complex emergent AI behaviors. Furthermore, academic-led research could lay a solid theoretical foundation to enrich the optimization of emergent AI behaviors for commercially viable use cases.
The Important Role of Regulators
In healthcare and life sciences, regulatory bodies such as the US Food and Drug Administration (FDA) have a crucial role in helping advance the field of AI in a responsible manner. Their core mission, to ensure medical product safety and efficacy, brings with it the ability to deploy reliable and replicable frameworks for scrutinizing the emergent behaviors of AI systems. Frameworks which lean heavily on mathematical modeling can offer objective toolsets to achieve this mandate, empowering regulators to develop a nuanced understanding of AI system behaviors for implementing appropriate regulatory review and monitoring procedures. Regulators can not only utilize these quantitative frameworks in the review process but actively participate in their development and refinement via intramural and extramural research support and collaborative research agreements, which have been effective mechanisms of advancing the use of innovations such as real-world evidence in drug development.
Regulatory standards should best evolve in tandem with the development of new mathematical frameworks. Standards that reflect the insights derived from the objective understanding of AI emergent behaviors can influence the entire lifecycle of AI systems, from design and development to deployment and assessment. Harmonization between regulatory standards and the rapidly evolving field of generative AI would ensure that the systems used in healthcare and life sciences are not only effective but also safe for deployment at scale. The FDA's Drug Development Tools (DDT) Qualification Programs, which aim to qualify and make publicly available tools that expedite drug development and regulatory review processes, could be potentially extended to accommodate the development and validation of AI evaluation toolkits.
Leveraging interdisciplinary mathematical models that incorporate principles from quantum physics and adjacent disciplines presents an opportunity to accelerate the evolution of safe and effective generative AI systems. The obstacles we face are as much technical as they are organizational, necessitating flexibility in reimagining our 20th-century interdisciplinary boundaries to foster a more fluid and collaborative space where innovative ideas can thrive as we push the frontiers of AI development.