TLDR: Similarity Field Theory introduces a mathematical framework where dynamic similarity relations form the structural basis of intelligence and cognition. It defines concepts like similarity fields, entities, and fibres, and formalizes intelligence as the capacity of a generative operator to preserve conceptual structure. The theory includes an Incompatibility Theorem for asymmetric relations and a Stability Theorem for system evolution. It also provides a new perspective on machine learning interpretability, viewing neural networks and large language models as compositions of similarity fields. Crucially, it demonstrates how LLMs can be used as scientific instruments to conduct virtual experiments, probing collective human cognition and revealing insights into ‘mind share’ versus ‘market share’ through pairwise typicality judgments and a novel ‘lock-filter’ denoising method.
A groundbreaking new mathematical framework, dubbed Similarity Field Theory, is set to redefine our understanding of intelligence and cognition. Proposed by Kei-Sing Ng, this theory posits that the fundamental building blocks of any comprehensible dynamic system are the ever-changing and persisting relationships of similarity between entities. It offers a fresh perspective on how intelligent systems preserve, modify, and generate these similarity structures, moving beyond traditional views that often rely on rigid axioms.
The Core Idea: Similarity as the Foundation
At its heart, Similarity Field Theory introduces a ‘similarity field’ – a mathematical mapping that quantifies how similar any two entities are, ranging from 0 (no similarity) to 1 (perfect similarity). Crucially, this field is designed to be flexible, allowing for asymmetry (where entity A might be very similar to B, but B is less similar to A) and non-transitivity (where A is similar to B, and B is similar to C, but A is not necessarily similar to C). This flexibility is vital for modeling the complexities of human cognition, where such nuances are common.
Within this framework, a ‘concept’ is defined as an entity that groups other entities based on their similarity to it, forming what the theory calls ‘fibres’ – essentially, sets of entities that meet a certain similarity threshold to the concept. The theory then provides a formal, generative definition of intelligence: an operator is intelligent if, when given entities belonging to a specific concept’s fibre, it can generate new entities that also belong to that same fibre.
Key Principles: Asymmetry and Stability
The paper proves two significant theorems that highlight the theory’s implications:
-
The Incompatibility Theorem: This theorem reveals a fundamental limitation arising from asymmetric similarity. If two entities perceive each other’s positions differently (i.e., their similarity values are not equal), they cannot simultaneously belong to each other’s conceptual fibres if the threshold for inclusion is set by the other’s standard. This offers a mathematical explanation for real-world phenomena like negotiation deadlocks, where differing perspectives can prevent mutual agreement.
-
The Stability Theorem: This principle states that any coherent, stable cognitive system must rely on at least one long-term, foundational belief or concept that acts as a source of stability. Without such an ‘anchor’ or eventual confinement within a stable conceptual level set, the system faces inevitable collapse.
Intelligence, Learning, and Creativity
The theory extends to define learning as the process by which a system’s generative operator or its similarity field is modified over time to improve its intelligence with respect to a concept. Creativity, too, finds a formal definition: it’s not creation from nothing, but rather the ‘re-contextualization’ of an existing entity. This happens when an entity, already part of one concept, is recognized as having high similarity to a different, emerging concept, thereby populating that new conceptual fibre for the first time. The example of WD-40, initially an aerospace anti-corrosion agent that found new life as a general-purpose lubricant, perfectly illustrates this idea.
A New Lens for Machine Learning and AI
Similarity Field Theory offers a powerful new way to interpret machine learning models, particularly neural networks and large language models (LLMs). Instead of viewing neural networks purely as statistical learning machines, the theory suggests they can be seen as complex systems composed of nested similarity fields. Each neuron, for instance, can be understood as computing a similarity value with respect to a latent concept. This perspective opens a principled route to model interpretability, allowing researchers to deconstruct models into their constituent conceptual fibres and understand how they are composed.
For LLMs, the theory proposes that a token is a concept, and a prompt is a higher-order concept. The LLM itself acts as both a vast meta-concept of language and a generative operator. Its prediction head, after calibration, can be treated as a surrogate for a similarity field, where selecting a high-probability next token means instantiating a high similarity to the prompt’s concept.
Probing Societal Cognition with LLMs
Perhaps one of the most exciting applications of Similarity Field Theory is its potential to transform LLMs into novel scientific instruments for probing societal cognition. By treating a trained LLM as a similarity field approximating collective human cognition, researchers can conduct ‘virtual experiments’ to quantify internalized conceptual structures. The paper details an experiment where LLMs were asked to make pairwise judgments about which carbonated soft drink or energy drink brand was “more typical.” The results, aggregated using the Bradley-Terry-Luce model, showed a strong alignment between the LLM’s inferred brand typicality and real-world market shares, even though the models were never given market share data.
Furthermore, the researchers applied a “lock-filter” based on the Incompatibility Theorem. When an LLM judged two brands to be mutually “more typical” of each other (an incompatible similarity), these judgments were down-weighted as likely noise. This correction significantly improved the accuracy of the LLM’s predictions, bringing them even closer to true market shares. The remaining deviations between the LLM’s “mind share” and actual “market share” offer fascinating insights, suggesting that real-world factors like marketing and distribution can elevate a brand’s market share beyond its pure cognitive typicality.
Also Read:
- The Inherent Design Flaw: How Transformers Generate Coherence Without Truth
- Beyond Pre-Training: How Experience Scaling Enables Continuous Learning for Large Language Models
A Foundational Shift
Ultimately, Similarity Field Theory suggests a foundational shift in how we perceive the world – from a collection of discrete objects to a dynamic interplay of similarity relations. It provides a computable language for the very structure of cognition, offering a powerful new framework for characterizing, comparing, and constructing intelligent systems. For more details, you can read the full research paper here.
