Pith. sign in

REVIEW 5 major objections 5 minor 2 references

Quantifying Avian Morphological Evolution through Deep Representation Learning

T0 review · 5 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The visual traits learned by a bird-recognition network show that avian form exploded soon after the K-Pg extinction, then decelerated — evidence for the niche-filling model of adaptive radiation.

desk verdict The DTT early-burst claim rests on an unvalidated spherical ASR whose weighting is backwards; the framework is promising but the headline result is not yet supported. read the letter →

arxiv 2602.03824 v2 pith:OUOHRCQK submitted 2026-02-03 q-bio.PE cs.CVq-bio.QM

Jiao Sun This is my paper · ORCID
classification q-bio.PEcs.CVq-bio.QM
keywords morphologicalevolutiondeeplearningrepresentationdisparitythroughtimeadaptiveradiationbirdsvisualmorphospaceancestralstatereconstruction
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that the high-dimensional feature vectors of a deep neural network trained only to classify bird species can serve as a valid, landmark-free morphospace for studying morphological evolution. By treating the network's final-layer weights for each species as morphological trait vectors, the author shows that this space recovers taxonomic hierarchy and captures convergent evolution without any a priori taxonomic input. The central macroevolutionary finding is a disparity-through-time curve that rises steeply after the K-Pg mass extinction, reaching roughly half of modern avian visual disparity shortly afterward, then decelerates — a pattern that deviates strongly from neutral Brownian-motion expectations. If correct, this supports the niche-filling hypothesis: surviving bird lineages rapidly diversified into vacated ecological niches. The work also argues that the network encodes whole-body shape rather than local texture, challenging a common assumption about convolutional networks.

What carries the argument

The load-bearing mechanism is the final-layer weight vector of a fine-grained image classifier, projected onto a unit hypersphere, as a high-dimensional morphological phenotype. These vectors define a cosine-similarity morphospace in which taxonomic and convergent structure emerges. The disparity-through-time analysis runs on a spherical Brownian-motion model: ancestral states are inferred by interpolating along the great-circle arc between sister nodes, with a correction factor that scales Euclidean contrast variance by the squared arc-to-chord ratio, and null simulations are generated by projecting Gaussian noise onto the tangent space of the sphere and moving along geodesics.

What would settle it

A concrete check: run a landmark-based geometric-morphometric disparity-through-time analysis on the same bird clades using the same timetree and comparable sampling; if that curve shows gradual disparity accumulation rather than an early burst, the classifier-embedding result would be contradicted. Alternatively, a simulation study demonstrating that the spherical ASR systematically inflates early disparity under a Brownian-motion null would undermine the method.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a convolutional neural network trained to recognize over 10,000 bird species implicitly learns a biologically meaningful morphospace. The 512-dimensional weight vectors extracted from the final fully connected layer, when L2-normalized and treated as points on a unit hypersphere, cluster by taxonomic group and place convergent taxa close together. Disparity-through-time analysis using a novel spherical ancestral-state-reconstruction algorithm reveals an empirical early-burst trajectory: after the K-Pg extinction, relative morphological disparity jumps to about 50% of the present-day value far faster than a Brownian-motion null model pre

Load-bearing premise

The whole argument rests on the premise that the classifier's final-layer vector for each species is a faithful, unbiased measurement of its morphology, and that the spherical ancestral-state-reconstruction algorithm correctly models trait evolution on a hypersphere — if either fails, the early-burst pattern could be an artifact.

Editorial extensions

If this is right

  • If the embedding morphospace is valid, species richness is the primary driver of morphospace expansion, with strong rank correlations at both order and family levels.
  • The recovered clustering reinforces that deep-network embeddings can rediscover established taxonomy and expose convergent phenotypes, such as ground-dwelling birds grouping together or penguins, auks, and procellariiforms sharing a visual region.
  • The disparity-through-time curve implies birds occupied about half of their current visual form space shortly after the K-Pg extinction, supporting an early-burst tempo for avian morphological diversification.
  • The model's apparent focus on whole-body plan rather than texture suggests that fine-grained visual classification may engage shape representations more strongly than generic object recognition.
  • The approach provides a high-throughput, anatomy-independent morphometric tool that could be extended to other imageable organismal groups.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The early-burst result should be tested against classical landmark-based morphometrics on a subset of the same bird clades; if landmark-based DTT shows a different tempo, the classifier-embedding geometry may be capturing something other than morphology.
  • The same classifier-embedding pipeline could be applied to other imageable taxa (e.g., insects, plants, fungi) to build a common visual morphospace, enabling cross-taxon disparity comparisons that landmark data cannot easily provide.
  • The shape-over-texture claim could be strengthened by generating a broader suite of texture-shape conflict images across many orders and measuring whether classification remains shape-driven; the current single adversarial example is suggestive but not exhaustive.
  • The spherical ancestral-state-reconstruction algorithm is central to the DTT result; a formal validation against known evolutionary simulations would clarify whether its arc-versus-chord correction introduces any bias in early disparity estimates.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper presents a deep-learning-based morphometric framework for birds. A ResNet34 is trained on the DIB-10K dataset to classify over 10,000 species; the weights of the final fully connected layer are treated as 512-dimensional morphological trait vectors. The authors show that hierarchical clustering of these vectors recovers much of the traditional taxonomy, that taxonomic similarity correlates with morphological disparity, and that an analysis of disparity through time (DTT) reveals an early-burst pattern after the K-Pg extinction. The early-burst result is obtained with a newly proposed 'spherical ancestral state reconstruction' (ASR) algorithm and compared with a Brownian-motion null simulation.

Significance. If the results were robust, this would be a useful high-throughput tool for macroevolutionary studies and a striking demonstration that visual morphospace expanded rapidly after the K-Pg boundary. The paper makes its code and model weights publicly available and includes an adversarial-example analysis arguing that the network captures shape rather than texture. However, the central DTT conclusion is not adequately supported: the spherical ASR is not derived or validated, the null simulation lacks confidence intervals, and the removal of 249 species is not transparently reported. These issues are fixable in principle but require substantial additional work.

major comments (5)
  1. [§2.5, Eq. (2)] The spherical ASR is stated without derivation. Eq. (2) sets v_p = slerp(v_a, v_b; t) with t = l_b/(l_a+l_b), so the parent is pulled toward the lineage with the longer branch. Under Brownian motion, the conditional expectation of the parent is precision-weighted by inverse branch lengths: (v_a/l_a+v_b/l_b)/(1/l_a+1/l_b), so the shorter branch should dominate. Thus Eq. (2) is backwards for BM. Since the empirical DTT curve (Fig. 7) is reconstructed with this ASR, the early-burst magnitude/timing may be an artifact. A simulation study with known ancestral states is needed to validate or correct this.
  2. [§2.5, Eq. (3) and 'equivalent branch lengths'] The correction factor Contrast2 = Euclidean variance × θ²/||va-vb||² is ad hoc; the 'equivalent branch length' for internal nodes is undefined. Without precise definitions, the ASR is not reproducible. Please provide a full derivation and pseudocode for all steps, including how internal branch lengths are assigned.
  3. [§2.5, Fig. 7] The null simulation is shown only as the mean of 100 runs, without confidence intervals or the number of species used. The claim that the empirical curve 'deviates extremely' cannot be assessed without a distribution. Report the full null distribution, the number of tips after pruning the Stiller et al. (2024) tree, and how missing species were handled.
  4. [§2.4 / Discussion] 249 species were removed after 'manual review' because they 'lacked biological significance.' The criteria are subjective and the list is not given. This removal could bias both the disparity and DTT analyses. Provide the list, explicit a priori criteria, and a sensitivity analysis including/excluding these species.
  5. [§2.2] The 'morphological trait' vector is the row of the final fc-layer weight matrix (a per-class template), not an embedding of an input image. The paper should justify why this is a valid phenotypic measurement; the Grad-CAM evidence does not establish that the weight vectors capture morphology. Compare with image-derived features or use an unsupervised embedding to address the circularity of recovering taxonomy from a classifier trained on those very labels.
minor comments (5)
  1. [§2.2] Typo: 'aenerative' should be 'generative'.
  2. [§4.3] Typo: 'futherly' should be 'further'; 'disaprity' should be 'disparity' in the residual discussion.
  3. [Fig. 4 caption] The caption says 'red dash line' — should be 'red dashed line'.
  4. [Fig. 5] Many labels render as '??', making the figure unreadable. Provide a vector figure or higher-resolution image.
  5. [Title] The title in the PDF body differs from the arXiv metadata title; unify them.

Circularity Check

1 steps flagged · score 2.0 of 10

Minor circularity in morphospace validation (taxonomy recovery is inherited from species-label training); the DTT early-burst claim is not circular.

  1. self definitional [§2.2 Model training; §3.2 Similarity clustering]
    "The task of recognising over 10,000 bird species is a fine-grained visual categorisation (FGVC) problem... all species are regarded as equal categories, without any a priori taxonomic knowledge being introduced. ... The weights are extracted from the final fully connected layer (fc) of the ResNet34 model. Weights of each species were regarded as a 512-dimensional vector representing the morphological traits. ... The clustering process... yielded a dendrogram that illustrates the relationships between the different avian species based on their morphological features learned by the ResNet34 mode"

    The species labels used as training targets are themselves taxonomic units, and the fc weight vectors are the rows of the final classification layer optimized to separate those labels. Clustering these vectors will therefore separate species and, because species are nested in a taxonomy, will tend to recapitulate that taxonomy. The 'emergent hierarchy' is inherited from the training labels rather than discovered from a taxonomy-free morphological measurement. This makes the morphospace-validation claim partly circular, although the DTT early-burst analysis is a separate downstream result.

full rationale

The central DTT claim is not circular: the empirical disparity-through-time curve is compared to a parametric Brownian-motion null that is generated using sigma^2 estimated from the same contrasts and then passed through the same spherical ASR. This is a standard parametric bootstrap; the early-burst deviation is an observed property of the embeddings, not a fitted constant. The unvalidated spherical ASR (Eq. 2-4) is a serious correctness risk, and the branch-length weighting appears inconsistent with Brownian-motion precision weighting, but that is an estimator-bias concern, not a circularity. The one genuine circular element is the validation of the morphospace via taxonomy recovery: the 'morphological traits' are the final-layer weights of a classifier trained on species labels, so recovering taxonomy from them is partly by construction. This validation is supporting rather than the core derivation, and the paper offers independent evidence (Grad-CAM, adversarial shape-vs-texture test) for the biological relevance of the embeddings. Hence a modest score of 2 is appropriate.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new particles or forces. Its axioms are domain assumptions about the embedding geometry and the evolutionary model, plus an ad hoc reconstruction algorithm. Free parameters include dimension-reduction and clustering thresholds, model-fit constants, and the time-bin size; the most consequential free choices are the post-hoc removal of 249 species and the unvalidated spherical ASR.

free parameters (6)
  • PCA variance threshold (80%) = 80%
    The 512-d vectors are reduced to the lowest dimension explaining 80% of variance; this choice affects all downstream analyses (§2.2).
  • Taxonomic purity threshold (85%) = 85%
    Used to define 'taxonomically consistent' nodes in the clustering analysis (§2.3).
  • Stretched exponential parameters (order-level) = lambda=0.1420, beta=0.3081
    Fitted in the model comparison for disparity vs diversity (§3.3).
  • Hill equation parameters (order-level) = n=0.3973, k=7.4957
    Fitted in the model comparison for disparity vs diversity (§3.3).
  • Power law exponent (family-level) = -0.1429
    Fitted in the model comparison for disparity vs diversity (§3.3).
  • DTT time bin size = 1 ma
    Spherical variance is calculated per 1-million-year slice, affecting the resolution of the DTT curve (§2.5).
assumptions (5)
  • domain assumption Phenotypic evolution follows Brownian motion on a unit hypersphere.
    Stated in §2.5: 'we assume that the feature vectors randomly walk on a unit hypersphere.' No empirical justification or test is provided.
  • domain assumption The fc-layer embedding encodes semantic information primarily in angular direction.
    Cites Wang et al. (2017), which is about face verification, and applies it to bird morphology without validation.
  • domain assumption The timetree from Stiller et al. (2024) accurately reflects avian divergence times and species sampling.
    Used as the backbone for all DTT and ASR analyses; no assessment of how the tree's taxonomic sampling relates to the 10,922 species in the embedding space.
  • ad hoc to paper The spherical ASR algorithm correctly estimates ancestral states under Brownian motion on a sphere.
    Presented as a geometric algorithm in §2.5 with no derivation, no simulation test, and no comparison to alternative estimators.
  • domain assumption The DIB-10K dataset and IOC taxonomy are accurate after the described cleaning.
    The cleaning pipeline (pHash, Faster R-CNN, manual review) is described but the final error rate is not quantified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Quantifying Avian Morphological Evolution through Deep Representation Learning." pith.science (2026). https://pith.science/paper/OUOHRCQK

@misc{pith2026260203824,
  author       = {Pith},
  title        = {Pith review of: Quantifying Avian Morphological Evolution through Deep Representation Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OUOHRCQK}},
  note         = {Machine review of arXiv:2602.03824}
}
read the original abstract

The evolution of biological morphology is fundamentally linked to ecological adaptation and species survival, yet traditional morphological evolution relies on landmark-based geometric morphometrics, a process constrained by subjective manual annotation, strict requirements for anatomical homology, and an inability to easily quantify complex, non-rigid traits such as plumage and texture. To overcome these limitations, we propose a scalable, landmark-free morphometric framework driven by deep learning. By extracting high-dimensional feature vectors from a Convolutional Neural Network (ResNet34) trained on images of over 10,000 bird species, we project raw visual semantics into a high-dimensional morphospace. Even without a priori taxonomic knowledge, this visual morphospace naturally recovers classical hierarchical taxonomy and effectively captures both homology and convergence. Analyses reveal a highly significant phylogenetic signal within the network's embeddings, with principal components correlating strongly with established ecological and morphological traits. Furthermore, by implementing a novel spherical Ancestral State Reconstruction algorithm, we uncover a pronounced "early-burst" pattern of disparity following the K-Pg mass extinction, supporting the niche-filling hypothesis of adaptive radiation.

Figures

Figures reproduced from arXiv: 2602.03824 by the authors.

Figure 1
Figure 1. Geometric diagram of the ancestor state reconstruction algorithm. [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Geometric diagram of the brownian motion simulation algorithm. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Grad-CAM reveals that the network’s attention is consistently focused on birds, [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: (a-f) The relationship between taxa size and mean pairwise angle, taxa size and [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: The unrooted clustering result with high-purity branches collapsed. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Comparison of models of the phenotypic disparity of taxa (orders or famalies) as a [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Relative disparity-through-time (DTT) plot for avian visual morphospace. The [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: The adversarial example of a chicken-plumaged penguin generated with Google Nano [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

2 extracted references · 2 canonical work pages

  1. [162]

    https://doi.org/10.1186/s12862-021-01888-5 Zhang, Y., Tang, H., & Jia, K. (2018). Fine-grained visual categorization using meta-learning optimization with sample selection of auxiliary data.European Conference on Com- puter Vision, 241–256. https://doi.org/10.1007/978-3-030-01237-3_15 17

    work page doi:10.1186/s12862-021-01888-5 2018

  2. [778]

    The DongNiao International Birds 10000 Dataset

    https://doi.org/10.1109/CVPR.2016.90 Hedges, S. B., & Sibley, C. G. (1994). Molecules vs. morphology in avian evolution: The case of the" pelecaniform" birds.Proceedings of the National Academy of Sciences,91(21), 9861–9865. https://doi.org/10.1073/pnas.91.21.9861 Huerta-Cepas, J., Serra, F., & Bork, P. (2016). Ete 3: Reconstruction, analysis, and visu- a...

    work page Pith review arXiv doi:10.48550/arxiv.2010.06454 2016

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.