MATH · IN · MODELS
structures / Manifolds / Persistent-homology / Betti profile across depth

Persistent-homology / Betti profile across depth

PROPERTYmeasurementfunctionaladvancedhow it's classified →

Betti numbers / persistent topological features (independent loops, holes, connected components) of a real model's activation point-cloud, tracked as a profile across layers, training, or model comparison — a computed topological-data-analysis quantity, not a claim that the representation manifold is any specific named shape. Orthogonal to intrinsic dimension and to curvature: a topologically rich point cloud can be low- or high-dimensional, flat or curved.

Replicationcomputed from the corpus — never hand-assigned
3 papers · no shared authors3 architecture classes · across papers3 domains · across papers4 model families · across papers
Filled = two or more values reported by papers that share no author — replication. Outlined = two or more values, but all from a single study — breadth, not replication. Grey = a single value. Derived from paper authorship and each model's architecture class, domain and family; it updates itself when a paper is added.

Definition

Rather than claiming the representation manifold is some named shape, this node covers papers that measure what topological features persist in a real model’s activation point-cloud — a computed quantity via persistent homology (tracking pp-dimensional holes/loops across a filtration of the point cloud, e.g. by distance scale or, for layer-to-layer comparison, a zigzag filtration connecting successive layers’ point clouds) — not a geometric object in its own right. Output is typically a persistence diagram or Betti-number profile: how many topological features of each dimension appear and how long they persist before merging or vanishing.

Relative to intrinsic-dimension-profile and manifold-curvature-profile

Intrinsic-dimension profile across depth tracks how many coordinates the manifold needs; Curvature profile of the representation manifold tracks how bent it is. This node tracks a third, independent question: what topological features (holes, loops, connected components) does the point cloud have, and how do they persist or evolve. All three are orthogonal measurements of the same underlying activation geometry — a manifold can be low-dimensional, flat, and topologically rich (many small holes) all at once, or any other combination; a paper reporting one of these three quantities makes no claim about either of the other two.

Key evidence: layer-wise topological tracking in LLMs

Gardinazzi, Viswanathan, Panerai, Ansuini, Cazzaniga & Biagetti (2024) introduce zigzag persistent homology to track how topological features of the layer-by-layer representation point-cloud persist and evolve across depth in real pretrained Pythia models (up to Pythia-6.9B). The resulting topological-stability signature — which layers’ removal would least disturb the persistent topological structure — is used as a practical layer-pruning criterion, validated against downstream task performance. See gardinazzi-etal-2024-zigzag-persistent-homology-tracks-topological-features-of-the-llm-representation-point-cloud-across-layers-and-flags-prunable-layers.

Key evidence: topology and intrinsic dimension jointly predicting generalization

Magai & Ayzenberg (2022) apply persistent homology and intrinsic dimension estimation jointly to the data manifold formed by penultimate-layer activations of real trained CNNs (CIFAR-10 and related datasets), tracking how topological complexity and intrinsic dimension evolve across network depth and training progress, and showing both descriptors are predictive of generalization performance. They extend the same measurement to a face-recognition system’s embedding manifold under adversarial spoofing attacks, using the topological descriptors as an anomaly signature. See magai-ayzenberg-2022-persistent-homology-and-intrinsic-dimension-of-the-last-layer-data-manifold-track-generalization-in-cnns.

Key evidence: cross-model topological/geometric audit in biology

Kendiukhov (2026) runs an autonomous 141-hypothesis topological/geometric audit of scGPT and Geneformer gene-embedding representations, finding persistent homology statistically significant (p<0.05) in nearly every transformer layer across both independently-trained models. See a-141-hypothesis-automated-audit-of-scgpt-and-geneformer-finds-significant-persistent-homology-in-nearly-every-layer-and-a-cca-canonical-correlation-of-0-80-between-the-two-models.

Key papers

  • Gardinazzi, Y., Viswanathan, K., Panerai, G., Ansuini, A., Cazzaniga, A. & Biagetti, M. (2024). Persistent Topological Features in Large Language Models. arXiv:2410.11042 — origin of the zigzag layer-tracking method and pruning application.
  • Magai, G. & Ayzenberg, A. (2022). Topology and Geometry of Data Manifold in Deep Learning. arXiv:2204.08624 — joint persistent homology / intrinsic dimension tracking predicting generalization.
  • Kendiukhov, I. (2026). What Topological and Geometric Structure Do Biological Foundation Models Learn? Evidence from 141 Hypotheses. arXiv:2602.22289 — large-scale automated audit finding significant persistent homology across two independently-trained genomics foundation models.

Found in (3 observations · 4 families)

scGPT

What Topological and Geometric Structure Do Biological Foundation Models Learn? Evidence from 141 Hypotheses (2026)measured

A 141-hypothesis automated topology/geometry audit of scGPT and Geneformer gene-embedding representations finds statistically significant persistent homology in nearly every transformer layer and a canonical-correlation cross-model alignment of 0.80 between the two independently-trained models

Details

Kendiukhov (2026) runs an autonomous executor-brainstormer loop that proposed, tested, and refined 141 geometric/topological hypotheses across 52 iterations about gene-embedding representations in scGPT and Geneformer, with explicit null controls and disjoint gene-pool splits. Persistent homology is statistically significant (p<0.05) in 11/12 transformer layers in the weakest domain and 12/12 in the other two; manifold-aware distance metrics outperform Euclidean distance for identifying regulatory gene pairs; graph community partitions track known transcription-factor-target relationships; canonical correlation analysis between scGPT and Geneformer's independently- trained gene-embedding spaces yields a canonical correlation of 0.80 and 72% gene-retrieval accuracy (though none of 19 tested methods reliably recover exact gene-level correspondence); robust signal concentrates in immune tissue under stringent nulls.

models: scGPT (whole-human pretrained checkpoint) · method: Persistent homology (Betti number analysis), Singular Vector Canonical Correlation Analysis (SVCCA)

Geneformer

What Topological and Geometric Structure Do Biological Foundation Models Learn? Evidence from 141 Hypotheses (2026)measured

A 141-hypothesis automated topology/geometry audit of scGPT and Geneformer gene-embedding representations finds statistically significant persistent homology in nearly every transformer layer and a canonical-correlation cross-model alignment of 0.80 between the two independently-trained models

Details

Kendiukhov (2026) runs an autonomous executor-brainstormer loop that proposed, tested, and refined 141 geometric/topological hypotheses across 52 iterations about gene-embedding representations in scGPT and Geneformer, with explicit null controls and disjoint gene-pool splits. Persistent homology is statistically significant (p<0.05) in 11/12 transformer layers in the weakest domain and 12/12 in the other two; manifold-aware distance metrics outperform Euclidean distance for identifying regulatory gene pairs; graph community partitions track known transcription-factor-target relationships; canonical correlation analysis between scGPT and Geneformer's independently- trained gene-embedding spaces yields a canonical correlation of 0.80 and 72% gene-retrieval accuracy (though none of 19 tested methods reliably recover exact gene-level correspondence); robust signal concentrates in immune tissue under stringent nulls.

models: Geneformer · method: Persistent homology (Betti number analysis), Singular Vector Canonical Correlation Analysis (SVCCA)

Pythia

Persistent Topological Features in Large Language Models (2024)measured

Zigzag persistent homology tracks how topological features (p-dimensional holes) of a real pretrained LLM's representation point-cloud persist and evolve across layers, and the resulting topological-stability signature identifies layers safe to prune

Details

Gardinazzi, Viswanathan, Panerai, Ansuini, Cazzaniga & Biagetti (2024) introduce zigzag persistent homology as a tool for tracking topological features of the layer-by-layer representation point-cloud in real pretrained language models (Pythia family, up to Pythia-6.9B), a genuine topological-data-analysis measurement applied directly to natural model activations rather than a toy or synthetic setting. By connecting successive layers' point clouds into a zigzag filtration, the persistence of topological features (independent loops/holes) across the network's depth becomes measurable, and layers whose removal would least disturb this persistent topological structure are flagged as safe candidates for pruning — turning a purely descriptive topological measurement into a practical model-compression criterion validated against downstream task performance. Clears scope on quantified geometric-shape measurement (persistent homology of the representation point-cloud) on a real pretrained LLM's natural activations; purely observational-diagnostic (the pruning application is a downstream use of the measurement, not a geometry-tied steering intervention). See [[manifolds]] for the topological-measurement precedent this shares with `a-141-hypothesis-automated-audit-of-scgpt-and-geneformer-finds-significant-persistent-homology-in-nearly-every-layer-and-a-cca-canonical-correlation-of-0-80-between-the-two-models` and `magai-ayzenberg-2022-persistent-homology-and-intrinsic-dimension-of-the-last-layer-data-manifold-track-generalization-in-cnns`.

models: Pythia-6.9B · method: Persistent homology (Betti number analysis)

ResNet

Topology and Geometry of Data Manifold in Deep Learning (2022)measured

Topological descriptors (persistent homology) and intrinsic-dimension estimates of the last-layer data manifold in real trained CNNs change measurably across layers and training, and predict generalization performance

Details

Magai & Ayzenberg (2022) apply persistent homology and intrinsic dimension estimation to the data manifold formed by penultimate-layer activations of real trained convolutional networks on CIFAR-10 and related image datasets, tracking how topological complexity (number and persistence of homological features) and intrinsic dimension evolve across network depth and training progress. They show these topological and geometric descriptors change in a way that is predictive of generalization performance, and extend the analysis to a face-recognition system's embedding manifold under adversarial spoofing attacks, measuring the same topological descriptors as an anomaly signature. Clears scope on quantified geometric-shape measurement (persistent homology, intrinsic dimension) on real trained CNNs' natural penultimate-layer activations; purely observational, no causal steering intervention. See [[manifolds]] for the broader topological-measurement precedent this shares with `a-141-hypothesis-automated-audit-of-scgpt-and-geneformer-finds-significant-persistent-homology-in-nearly-every-layer-and-a-cca-canonical-correlation-of-0-80-between-the-two-models`.

models: ResNet-18 (supervised, ImageNet) · method: Persistent homology (Betti number analysis), Intrinsic dimension estimation (TwoNN)