When people think of Geoffrey Hinton, they think of backpropagation, Boltzmann Machines, Deep Belief Networks, and the deep learning revolution. But in 1977, nearly a decade before the famous backpropagation paper, Hinton completed a PhD thesis at the University of Edinburgh titled "Relaxation and its Role in Vision." At first glance, it appears to be about computer vision and relaxation methods. It is about much more.
The Core Problem
The thesis opens with a deceptively simple question: how can a visual system choose the correct interpretation when a single image supports many plausible explanations? Real-world scenes are ambiguous and partially hidden. A premature decision introduces errors that propagate through the rest of the reasoning. The challenge is keeping multiple hypotheses alive until enough evidence determines the most consistent one.
Hinton rejected the common approaches of the 1970s. One, the "postponement strategy," delayed decisions by leaving information unspecified but offered no way to compare competing hypotheses. Another assigned fixed meanings to low-level features, which failed when objects were partially hidden or appeared in new contexts.
Perception as Search and Construction
One of the most forward-looking ideas is that perception is not recognition. It is the process of building an internal explanation of the scene. The system must decide which objects are present, how they relate, and which interpretation best fits the evidence. Hinton explicitly rejected template matching, arguing it was too limited for complex or unfamiliar scenes. Instead, perception is constructive: the system builds an interpretation by combining evidence, relationships, and prior knowledge until a coherent explanation emerges.
This sounds remarkably like modern generative models and representation learning, where the goal is not to memorize examples but to learn meaningful internal representations that explain new observations. Hinton was exploring these ideas in 1977.
Uncertainty as a Feature
Another strikingly modern choice: Hinton avoided treating hypotheses as simply true or false. Every hypothesis gets a value between 0 and 1 reflecting current belief strength. As relaxation proceeds, these values update repeatedly until the most consistent interpretation stands out and weaker ones fade. Today we call these probabilities, belief values, confidence scores, activations, or logits. The terminology changed, but the idea that intelligence depends on representing uncertainty rather than making immediate, irreversible decisions remains central.
Distributed, Parallel Computation
Hinton argued that intelligence should not depend on a single central controller. Instead, he described a system of many local hypotheses interacting simultaneously, each contributing a small part of the final solution. This emphasis on networks, interactions, and distributed computation is clearly visible in modern neural networks. At a time when computers executed instructions sequentially, Hinton argued perception was better viewed as many parallel processes influencing one another. Decades before massively parallel hardware became common, he described computation in terms that closely resemble how modern neural networks run today.
Constraints and Global Consistency
No hypothesis should be evaluated in isolation. Each influences others through a network of constraints. When one hypothesis's confidence changes, that change spreads across the network, strengthening compatible explanations and weakening conflicting ones. This later became a theme in graphical models, factor graphs, message passing, and belief propagation. The mathematical frameworks differ, but the conceptual connection is clear.
The Puppet Experiment
To demonstrate these ideas, Hinton used a deliberately simplified domain. A user drew overlapping rectangles on a graphics terminal. Some represented parts of a stick-figure puppet; others were distractors. Every overlap became a candidate joint. The system generated competing hypotheses about which rectangles belonged to the puppet and which overlaps were real connections. The goal was finding the interpretation with the greatest number of mutually consistent joints, robust to missing parts and clutter.
By removing the complexity of natural images, Hinton isolated the combinatorial challenge of visual interpretation while keeping the problem tractable. The puppet task was not a toy example. It was a complete program with hypothesis generation, constraint construction, iterative relaxation, and final selection. The final result emerged from cooperation among local hypotheses rather than central command.
From Local to Global Optimization
A key theme throughout is that solving small local conflicts does not necessarily produce the best overall interpretation. A hypothesis may fit its immediate neighbors while contributing to an incorrect global explanation. The system must evaluate how all hypotheses work together. This shift from local agreement to global optimization reflects a broader direction AI later took.
Relaxation as Philosophy
Reading the thesis, relaxation starts to feel like more than an algorithm. It is a way of approaching difficult problems. Start with tentative beliefs, refine them through repeated interactions, and continue until the solution stabilizes. Many modern methods follow this pattern: gradient descent improves parameters step by step, EM algorithms alternate between refinement stages, belief propagation exchanges information repeatedly, and diffusion models generate samples through gradual updates. The mathematics differ, but the philosophy is the same: good solutions emerge through many small improvements, not one decisive computation.
Schemas and the SETTLE System
In later chapters, Hinton introduced schemas as a way to organize knowledge and connect it to perception. Instead of storing exact examples, knowledge should capture the rules, relationships, and constraints that define a category. This allows the system to interpret new situations by reasoning about underlying structure rather than matching what it has seen. This foreshadows latent representations and internal world models.
The SETTLE system combined schemas, inference rules, and relaxation into a single framework. Inference rules generated candidate conclusions while relaxation continuously evaluated their consistency. Evidence, rules, and competing hypotheses influenced one another until converging on the most coherent interpretation. SETTLE is interesting not because it resembles modern AI in detail, but because it reflects Hinton's early effort to integrate multiple reasoning forms into a unified process.
The Continuity of a Research Philosophy
After finishing the thesis, what impresses most is not a single algorithm but the continuity of Hinton's thinking. Relaxation, competing hypotheses, distributed constraints, optimization, and stable solutions in 1977. Boltzmann Machines, backpropagation, Deep Belief Networks, AlexNet, the Forward-Forward algorithm, and mortal computation decades later. The contributions seem different, yet they connect through a consistent philosophy: intelligence emerges from the interaction of many simple elements, not a central controller; distributed knowledge over isolated symbols; inference over recognition; rich internal representations over final outputs; optimization as the mechanism for intelligent behavior; and uncertainty as something to represent and refine, not ignore.
Hinton's 1977 thesis is less an isolated piece of early research than the beginning of an intellectual journey that shaped nearly five decades of artificial intelligence.