ADR-010: representation-learning and learning-theory join the topics

Context

On 2026-09-26 the owner asked this record to file a list of works under four of their own headings:

• Knowledge Graphs;
• Relating K-complexity to minimum sufficient statistics;
• Riesz Representation Theorem;
• Representation learning as a spectral approximation.

Most of them are machine-learning theory:

• spectral accounts of contrastive and non-contrastive self-supervised learning;
• the information bottleneck and the conditional entropy bottleneck;
• sheaf-theoretic knowledge-graph embedding;
• compression, rate–distortion and description-length bounds on generalization;
• the Kolmogorov structure function as model selection.

The vocabulary has no word for either strand:

• information-theory is blurbed as information theory “read for [itself]” and “outside machine learning”;
• probabilistic-modeling is Bayesian inference and statistical models;
• mathematics is mathematics read for itself.

Each is the nearest wrong word for a paper on why a contrastive loss recovers a spectral embedding. The record’s rule is to add the word rather than reach for the nearest wrong one (ADR-006, ADR-007, ADR-009).

Filing them here at all follows the owner’s decision of 2026-09-26 that new work stays in nucleation until a transfer is judged appropriate. The works chiefly about ML practice carry anthology-candidate (ADR-005).

Decision

Add two words to topics, each with a label and a blurb:

• representation-learning: how learned systems come to represent their data. That takes in self-supervised and contrastive objectives, spectral embeddings, information bottlenecks, knowledge-graph embeddings, and whether representations converge.
• learning-theory: why learning generalizes. That takes in compression and description length, Kolmogorov complexity and sufficient statistics, rate–distortion and PAC-style bounds, and model selection.

Neither joins the philosophy group.

Alternatives

• One word, machine-learning. It would bind a spectral account of SimCLR to an MDL generalization bound as if they were about the same thing. The owner’s own headings already separate the two strands.
• Stretch information-theory to cover them. Its blurb excludes machine learning deliberately. Several of these works are information-theoretic and about learning, and they carry both words.
• File them in the anthology. That record is for practice and the claims behind it. These are mostly theory, and the owner asked for them here. anthology-candidate marks the ones a transfer should consider.