Representation space is a concept for systematically exploring ways to describe problems that reduce the cost of solving them, preserve relevant information, and allow results to be verified.
Author: Roman Shaban
Status: Conceptual article and research programme
Version: 0.1
Some difficulties in science, education, and decision-making may arise from an unsuitable representation of the problem.
The concepts, variables, data formats, and levels of description we choose influence which patterns are easy to detect, which operations are available, and how much effort a solution requires.
Changing the representation may make a problem easier. It may also conceal important distinctions, move complexity into preprocessing, or produce an answer that no longer serves the original purpose.
The research question is therefore: Which representation provides a verifiable advantage for a particular problem when all relevant costs and losses are considered?
A representation is a way of encoding entities, properties, relationships, and constraints for a particular task.
The same system may be described through text, tables, graphs, equations, probabilistic models, or models of its dynamics. Each description enables certain operations and has its own limitations.
A model uses a representation to describe, explain, or predict a phenomenon. The representation establishes the form of the description; the model adds specific assumptions and relationships.
A representation transformation is a rule for moving between descriptions. It may be reversible, partially reversible, or involve information loss.
A representation space is a set of available descriptions together with admissible transformations between them. This is a working concept in the proposed programme. It should not be confused with the mathematical representation theory of groups.
A network of relationships can be described as a long list or represented as a graph. A graph makes it easier to examine routes, dependencies, and nodes on which the system depends.
A frequency representation of a signal may reveal periodic components. An appropriate change of variables can simplify an equation. A table with consistent fields allows comparisons that are difficult to make in unstructured text.
However, the length of a description and the cost of computation are different properties. Writing a number as 10¹² is more compact than writing it in full decimal notation, but this does not establish that every operation involving it becomes cheaper.
A better representation is defined by the task and the results of evaluation.
A valid comparison should distinguish three cases.
Representational equivalence: The descriptions preserve specified information or results, and the transformation between them is known.
Structural analogy: The descriptions share an organizational pattern that may suggest a hypothesis or method. Transferring an analogy between domains requires separate validation.
Compression: The new description preserves selected properties while discarding some detail. Its suitability depends on whether that detail matters for the task.
Similar equations in biology and economics do not, by themselves, establish identical mechanisms. A shared mathematical form may support the transfer of analytical tools, while causal explanations require evidence within each domain.
An invariant is a property that remains unchanged under a specified class of transformations.
A claim that an invariant has been identified should state:
What is preserved: A quantity, relationship, order, structure, or result.
Under which transformations: Changes of coordinates, units, notation, or another explicitly defined transformation.
Within what scope: The relevant entities, conditions, and levels of precision.
How it was verified: Through proof, computation, or testing against data.
Preservation under changes of description makes a property a valuable object of study. Establishing that it is fundamental requires further justification.
The philosophical hypothesis that reality consists of invariants across representations remains an open interpretation. Practical work can proceed by testing specific transformations and their consequences.
Evaluation should consider several distinct characteristics.
Accuracy: Does the representation preserve the quality of solutions and predictions?
Information loss: What is discarded, and does that loss affect the conclusions?
Description complexity: How many rules, parameters, and assumptions are needed under an agreed measurement method?
Total cost: What are the costs of searching for the representation, transforming the data, solving the problem, verifying the result, and interpreting it in the original context?
Generalization: Does the approach work on new cases that were not used to select it?
Robustness: How sensitive are the results to noise, missing information, and changing conditions?
Traceability: Can the path from the original data to the conclusion be reconstructed?
Parameter count alone is insufficient for comparison. Complexity may be hidden in preprocessing, encoding rules, or external resources.
Representation Gain is a proposed label for the measurable advantage of a new representation over a baseline approach in a defined task.
An initial cost measure can be expressed as:
Cost gain = total cost of the baseline approach / total cost of the new approach
This measure is meaningful only under comparable conditions and when predefined quality requirements are satisfied. A value greater than one indicates lower costs in that comparison.
Accuracy, information loss, and conditions of use should be reported separately. This is a working measure for the proposed programme; this publication provides no empirical validation of it.
Cross-domain representation gain means an advantage demonstrated in multiple independently evaluated domains. Success on one task does not establish that the method applies universally.
Specify the task. Define the input data, required output, constraints, and acceptable error.
Describe the baseline. Establish how the problem is currently solved and what costs are involved.
Generate alternatives. Change variables, descriptive structures, scales, or formats while recording the transformation rules.
Check preservation of meaning. State what is preserved, what is lost, and how the result should be interpreted in the original problem.
Separate search from evaluation. Reserve new cases for final testing; do not select the representation using their outcomes.
Compare total costs and quality. Include search and verification costs and apply the same evaluation criteria.
Record limitations and failures. Show where the approach helps, where it provides no advantage, and when it should be revised.
The protocol should produce a validated proposal for representing a task, with documented transformations, measured advantages, and explicit limitations.
Suppose a course introduces many terms, but students struggle to distinguish their functions and relationships.
Baseline representation: A sequence of definitions and examples.
Alternative representation: A concept map showing relationship types, conditions of application, and practical situations.
Hypothesis: The map will help students distinguish concepts and select appropriate ways to apply them.
Evaluation: Compare learning outcomes on unfamiliar tasks under comparable time and teaching conditions. Assess accuracy, transfer of knowledge, and the cost of preparing the materials separately.
Rejection condition: If the map improves memorization but reduces understanding of important distinctions, its advantage for the stated goal remains unconfirmed.
This is an illustrative scenario, not a report of completed research.
Representation Discovery Engine is a working name for a possible software tool that generates and compares alternative representations of problems.
Its proposed cycle is:
Task → alternative representations → transformation checks → result comparison → selection → independent evaluation
Such a system could propose feature sets, dependency graphs, methods of aggregating data, and alternative problem formulations. Each proposal should include its assumptions, information losses, verification method, and comparative results.
Developing new formal languages and discovering broad invariants across disciplines are long-term research directions. Their feasibility is not established in this publication.
The concept connects with representation learning, model selection, the Minimum Description Length principle, dimensionality reduction, and the methodology of scientific modelling.
The general idea that the form of a description matters has scientific predecessors. The potential contribution of this programme should be sought in specific implementations: rules for comparing representations, accounting for total costs, checking transformations, and conducting reproducible evaluations.
Two starting points for comparison are:
Bengio, Courville, and Vincent. Representation Learning: A Review and New Perspectives. Read the paper.
Grünwald. A Tutorial Introduction to the Minimum Description Length Principle. Read the paper.
For science: Exploring alternative variables and testing the transfer of methods.
For education: Comparing ways of organizing knowledge.
For business: Finding descriptions of processes that reveal dependencies and support better decisions.
For AI development: Evaluating representations against explicit quality criteria.
An initial prototype can be small: one task, one baseline approach, several alternatives, and a set of independent tests.
Commercial potential depends on concrete benefits: reduced costs, improved accuracy, or greater verifiability of results. These benefits require separate validation.
The programme aims to determine when changing a representation genuinely helps solve a problem, and to make that advantage measurable and reproducible.
Title: Representation Space and the Search for Better Knowledge Models
Author: Roman Shaban
Type: Conceptual article and research programme
Language: English
Version: 0.1
Subject: The influence of representation on solution complexity and model quality
Central hypothesis: Systematic representation search can improve results at an acceptable total cost
Proposed method: Alternative generation, transformation checks, comparison with a baseline, and independent evaluation
Key concepts: Knowledge representation; representation learning; model selection; invariants; information compression; Minimum Description Length; artificial intelligence; scientific methodology
Verification status: No empirical results are reported in this publication
Development criterion: A reproducible advantage over a defined baseline without unacceptable loss of quality