Vladimir Vapnik: Predicates, Intelligence & Plato
The Science of Intelligence
Vladimir Vapnik argues that the engineering of intelligence (building systems that act like humans) is a distinct and separate pursuit from the science of understanding intelligence (exploring the fundamental nature of cognitive systems). While modern AI often focuses on building useful tools, like self-driving cars, Vapnik emphasizes the need to uncover the core predicates that define intelligence.
The Role of Predicates and Invariance
Drawing a parallel between the structural units of Russian folklore and the mechanics of learning, Vapnik stresses that:
• Intelligence relies on a universal set of predicates—fundamental statements or functions that describe the world.
• These predicates allow the construction of invariants, capturing the essence of an object regardless of specific noise in the data.
• A small set of "good" predicates can dramatically reduce the set of admissible functions, thereby reducing the amount of training data required to reach accurate predictions.
Plato, Wigner, and Modern Learning
Vapnik frames this endeavor through a philosophical lens, connecting Plato’s Theory of Forms to modern statistical learning:
"I strongly believe in these Plato ideas that there exists a world of predicate and a world of reality, and predicate and reality is somehow connected."
He posits that:
• The world of ideas is small, while the world of things is arbitrarily vast.
• True intelligence involves discovering the limited, powerful predicates that project onto the world of things to create understanding.
• The goal of a machine should be to move beyond simple heuristic-driven architectures toward approaches grounded in theoretical principles, specifically utilizing both strong and weak convergence.
The Challenge: Redefining Machine Learning
Vapnik issues a call to rethink the current trajectory of AI, particularly in fields like computer vision:
• He criticizes the reliance on massive datasets (e.g., needing 60,000 images for simple digit recognition), arguing that true intelligence should solve these tasks with significantly less data.
• He proposes that the path to general-purpose visual understanding lies in identifying these universal predicates—such as degree of symmetry—rather than just building larger neural networks.
• He discusses the potential of learning using privileged information, where a model receives additional structural or descriptive information (e.g., poetic descriptions of digits) to guide the training process.
Ultimately, Vapnik remains a staunch advocate for mathematical rigor and philosophical honesty, suggesting that the most powerful solutions to complex problems are often elegant, closed-form, and deeply rooted in the structural symmetries of our world.