The Reach of Manifold Learning:Uniqueness bounds for latent representations
Bibliographic record
Abstract
In manifold learning the aim is to represent data sets in lower dimension by fitting a manifold to the data. Here the word manifold refers to the more general idea of a subsets which can be described in fewer variables than the ambient space, as opposed to the stringent mathematical definition of a manifold. Lower dimensional representations can be found by projecting a data point to the (unique) nearest point on the manifold. However, for any non-linear manifold, such a unique projection does not exist everywhere. One way to study where a unique projection does exist is through reach. Reach gives a global bound on how far one can move from the manifold while ensuring that a unique nearest point exists. However, this bound is too restrictive in a practical setting, as many points further away will still have a unique nearest point. Instead we introduce a new uniqueness bound called the pointwise normal reach, which gives a bound on how far one can move in a normal direction while ensuring unique projections. In addition to the pointwise normal reach, we introduce a related uniqueness bound on immersed manifolds. This bound is beneficial as it utilises standard analytical tools to compute the bound. Finally, we use these bounds in practice, as we aim to understand the uniqueness of representation in the autoencoder algorithm. We create a test which asks if a datapoint is within reach of the assigned latent representation. A datapoint which fails this test is not ensured to have a unique best choice of latent representation. We employ Monte Carlo Sampling to estimate the pointwise normal reach of three autoencoders. We find that the algorithms fail this test for most data points. Thus, almost no data points are certain to have a unique best choice of representation. To improve this, we introduce a regularising term into the training algorithm. This significantly improves the amount of points which passes the reach test. However, the experimental setup is currently too expensive to employ for non-toy datasets.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.012 | 0.087 |
| Meta-epidemiology (narrow) | 0.003 | 0.001 |
| Meta-epidemiology (broad) | 0.002 | 0.003 |
| Bibliometrics | 0.003 | 0.002 |
| Science and technology studies | 0.003 | 0.010 |
| Scholarly communication | 0.004 | 0.018 |
| Open science | 0.004 | 0.012 |
| Research integrity | 0.004 | 0.011 |
| Insufficient payload (model declined to judge) | 0.006 | 0.002 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".