Do Manifold Untangling Results Survive a Change of Analysis Settings?

Deep networks are often described as "untangling" their inputs. Images of the same object seen from different angles start out as a curved, overlapping cloud of points, and layer by layer the network reshapes these clouds into compact, well-separated groups that a single straight cut can tell apart. A family of geometric measurements makes this story quantitative. Chief among them is mean-field manifold analysis, a statistical-physics method that turns a cloud of activations into three numbers: how many categories a linear classifier could separate per neuron (capacity), how spread out each category's cloud is (radius), and how many directions it extends along (dimension). These measurements underpin claims across vision, speech, and language. Yet every study computes them with its own lab scripts, with its own choices of how many examples per category to sample, how to subsample neurons, how to center the data, and which solver to use. Nobody has checked whether the published conclusions depend on those choices.

We propose to answer that question directly and to leave behind a measurement tool that stays correct over time. The work has two parts. The first is a "multiverse" audit: re-derive the flagship untangling claims under a systematic grid of defensible analysis settings, and report how often each claim still holds. The second is a containerized reference implementation. Its test suite is pinned to published geometric values, so that when a library upgrade silently changes a capacity or radius value, the change is caught automatically. The audit tells the community which knobs actually matter. The regression-tested implementation keeps those knobs fixed after the original authors have moved on.

Core Claim

Our claim is that representational-geometry results, like the benchmark results audited by NerfBaselines and A Sober Look at the Unsupervised Learning of Disentangled Representations and their Evaluation, carry hidden analytic degrees of freedom. Some reported conclusions will prove robust. Others, especially fine-grained rankings of models or layers and the relation between capacity and out-of-distribution robustness, may flip under reasonable alternative settings. The contribution is a quantified answer, per claim, rather than a new geometry metric.

The specific claims under audit are the following. Radius and dimension shrink with depth in ImageNet convolutional networks: AlexNet dimension "drops from over 80 in early layers to approximately 20 in the final feature layer," and radius shows a "steady decline from roughly 1.4 to 0.8" (Separability and geometry of object manifolds in deep neural networks). Masked language-model word manifolds lose radius, which "typically drops by 10-20% from the first to the last layer" (Emergence of Separable Manifolds in Deep Language Representations). Speech networks shed speaker information in deep layers (Untangling in Invariant Speech Recognition). Finally, "ultra-rich" training expands manifolds and hurts corruption robustness despite high in-distribution capacity (Feature Learning beyond the Lazy-Rich Dichotomy).

How This Relates to Prior Work

Two strands of prior work supply the method. The first comes from the evaluation literature. Locatello et al. trained more than 14,000 disentanglement models under a locked protocol and found that seeds and hyperparameters mattered more than method, and that the metrics disagreed with one another. NerfBaselines showed that trivial protocol choices, such as float versus integer image evaluation or the version of an image-similarity library, flip leaderboards. It also showed that 12 of 19 official codebases no longer installed, and it restored reproduction to within under 1% by wrapping them in containers. We borrow Locatello's variance-decomposition logic and NerfBaselines' container-plus-thin-wrapper design. Unlike Locatello, we audit discriminative separability metrics rather than generative factor scores. Unlike NerfBaselines, what we standardize is a measurement of internal representations, not an output-quality score.

The untangling papers themselves are the objects of the audit and motivate it. Each reports geometry from a single analysis configuration. The theory in Classification and Geometry of General Perceptual Manifolds is exact only in a large-dimension limit, which makes finite-sample choices such as the number of points per manifold and neuron subsampling natural suspects for instability. Chou et al. is the paper most likely to contradict our worry, because they already compare regimes systematically. Their analysis, however, is largely confined to final layers and uses bespoke scripts. We treat their capacity-versus-robustness finding as a hypothesis to stress-test rather than as settled.

Two further works constrain the scope of the cross-metric comparison. Neural Feature Geometry Evolves as Discrete Ricci Flow measures curvature on nearest-neighbor graphs of activations, so its results depend on the neighborhood size, one more setting to vary. Unveiling and Mitigating Generalized Biases of DNNs through the Intrinsic Dimensions of Perceptual Manifolds uses per-class intrinsic dimension. Whether these metrics agree has never been tested in one testbed, a gap On Meta-Evaluation frames as the missing meta-evaluation step.

The sustainability literature justifies the second half of the project. Core neuroscience tools have a "truck factor," the number of maintainers whose departure would stall the project, of only 1 to 3 (Cycling on the Freeway). Grant-cycle funding leaves maintenance unfunded (Computational and Informatic Advances for Reproducible Data Analysis in Neuroimaging). The research file found no curated manifold-analysis software on Lacuna at all. The mechanism we adopt resembles the automated validator that drove adoption of the Brain Imaging Data Structure, the neuroimaging data-format standard: an automatic check, not a policy document. The Need for Open Source Software in Machine Learning argued for open, citable ML software. We add the specific safeguard that the software must keep reproducing published numbers.

Research Program

We will first build the reference harness. It contains mean-field capacity, radius, dimension, and center correlations; per-class intrinsic dimension; and Ollivier and Forman Ricci curvature on neighbor graphs. All of these run behind one interface that takes a frozen model, a dataset with fixed category and sample definitions, and an explicit analysis configuration. Every configuration value is recorded with each result. The harness ships as Docker and Apptainer images with pinned dependencies, following the NerfBaselines design.

Its test suite includes "golden-value" tests. These re-run the reference networks and assert that the outputs fall within tolerance of the values printed in the original papers, and they run in continuous integration on every dependency update. We will also report the health of the harness itself using the in-process indicators of Towards Sustainable Digital Humanities Software, namely defect density and issue spoilage, so that its maintainability is measured rather than asserted.

With the harness in place, we will run the multiverse. Its axes are the choices a careful analyst could defensibly make: the number of categories sampled, the points per category, the neuron subsampling or random-projection dimension, global versus per-manifold centering, the classification margin, the solver and its tolerance, numerical precision, the neighbor count for curvature, and the estimator for intrinsic dimension. For each published claim, we compute the fraction of configurations under which it holds. For comparative claims, we compute the rank correlation of model and layer orderings across configurations. Following Locatello, we split the total variance of each metric into parts attributable to analysis settings, training seed, and architecture.

Baselines

The naive reference point is the published single-configuration result. Wherever the original code can still be run, we run it in a container on the same checkpoints and data samples; otherwise we reproduce it from the paper's stated settings. Every multiverse result is reported relative to this point. The audit is informative if analysis settings account for a variance share comparable to that of architecture. It is equally informative, in the reassuring direction, if they do not.

The strongest existing method for the scientific question of whether geometry predicts robustness is plain linear probing: training a linear classifier on each layer's activations and using its accuracy. It is cheap and directly tied to the task. We give the probe the same activations, the same class samples, and a tuning budget matched to the geometry pipeline's setting search. The benchmark is CIFAR-10 and CIFAR-10C, using the lazy-to-rich training sweep of Chou et al. To beat the probe, geometric features must add explained variance in corruption accuracy across models, measured by cross-validated R², beyond what probe accuracy and clean accuracy already explain, and this advantage must hold across most multiverse configurations rather than only the published one. We found no reported result on Lacuna for probe accuracy predicting CIFAR-10C robustness. The number is therefore produced here, not taken from prior work.

The ablation that isolates our key sustainability ingredient compares containers with golden-value tests against containers alone. We replay a sequence of historical versions of the numerical stack (deep-learning framework, array and scientific libraries, optimization solver) and count how many version changes shift a metric beyond tolerance without raising any error. Containers alone freeze a single snapshot. The golden tests are what detect silent drift when that snapshot must eventually be updated. If no version change ever shifts the values, the golden tests are unnecessary, and we will say so.

Why not just…?

Why not just use linear probes or accuracy and skip the geometry? This is the most serious threat, and the proposal is designed partly to settle it. Chou et al. report that capacity tracks feature-learning richness better than weight-change or kernel-alignment markers, and the intrinsic-dimension work finds class biases hidden behind equal accuracy. Neither result was tested against a well-tuned probe under varied analysis settings. The incremental-R² experiment above answers the question. If geometry adds nothing once probe accuracy is known, we will report that geometry is descriptive but not predictive of robustness, a useful negative result.

Why not just fix one canonical configuration by convention and move on? Picking defaults without knowing their sensitivity hides the degree of freedom instead of removing it, which is exactly how the rendering-protocol problems in NerfBaselines went unnoticed. The multiverse tells us which settings can be fixed harmlessly and which must be reported. If every claim proves stable, the right outcome is precisely a canonical configuration, now justified by evidence.

Why not just release code with a pinned environment file, or compare representations with centered kernel alignment? Pinned environment files are what failed in 12 of 19 NerfBaselines methods. Our drift-replay experiment tests directly whether that failure mode also produces silent numerical changes, not only install failures, in geometry code. Centered kernel alignment measures how similar two representations are, not whether categories are separable, so it answers a different question. We nonetheless include it as an extra axis in the cross-metric agreement analysis, to check whether it simply duplicates what the untangling metrics report.

Proposed Experiments

The first experiment is reproduction. Inside the harness, we reproduce the AlexNet, VGG-16, and ResNet-50 ImageNet trajectories and check them against the published dimension and radius values. We reproduce BERT-style masked radius reduction on Penn Treebank and OntoNotes, speaker and word capacity in speech networks on LibriSpeech, the lazy-to-rich sweep on CIFAR-10, CIFAR-100, and CIFAR-10C, and Ricci coefficients on MNIST, Fashion-MNIST, and CIFAR-10. Failures to reproduce will be diagnosed and reported as findings.

The second experiment is the multiverse audit described above, reporting for each claim the fraction of configurations in which it holds, the rank correlation of orderings across configurations, and the variance decomposition. The third tests cross-metric agreement. On shared checkpoints, we ask whether capacity gains, intrinsic-dimension compression, and curvature contraction rank layers and models consistently, which addresses the disagreement question that Locatello raised for disentanglement metrics. The fourth is the probe comparison for corruption robustness. The fifth is the dependency-drift replay.

Three outcomes will decide how the project is read. If the depth-wise compression claims hold in nearly all configurations while the capacity-robustness relation holds only in a minority, the field's headline story is sound but its most recent and consequential claim is fragile. If analysis settings explain more variance than architecture does, then comparisons across papers are currently unreliable and the harness defaults become necessary rather than optional. If golden tests catch silent drift that containers alone would have let through, the sustainability mechanism earns its place on evidence rather than advocacy.