Abstract
There are many cases where scientists have fossils for an extinct species but do not have a complete skeleton. This article proposes a machine learning framework that could take incomplete fossils and make informed guesses about what the rest of the skeleton might have looked at. These guesses would be informed by a generative AI system trained to take fragments of skeletons from existing animals where the full skeleton is known, and use these fragments to guess what the whole skeleton would have looked like. This could be useful in reconstructing all kinds of extinct species from their fragmentary fossil remains.
Fossil2Skeleton is a proposed machine learning framework for generating plausible complete skeletons from preserved bones and fossil fragments. It is based on the idea that the shape and dimensions of one part of an animal can contain information about other parts of its skeleton. The system would learn these relationships from three-dimensional scans and measurements of documented skeletons from living and extinct animals. During training, selected bones or regions would be hidden, and the model would predict their measured anatomy from the remaining material. Comparing these predictions with the known structures would allow it to learn relationships among bone shapes, body proportions, and anatomical connections. Including fossil specimens would expose the model to additional anatomical variation and evolutionary combinations that are poorly represented among living animals. When given an unfamiliar fragment, a generative model would produce several coordinated skeletal reconstructions, preserve the observed anatomy, and estimate uncertainty in the missing regions. Tests on withheld specimens and species would assess whether detailed fragment morphology improves prediction beyond body size and evolutionary relatedness. Potential applications include reconstructing incomplete fossils, estimating body proportions, examining the functional consequences of alternative anatomies, and identifying additional measurements that would most reduce uncertainty. Fossil2Skeleton would provide a quantitative way to investigate how much of an animal’s anatomy can be inferred from the parts that survive.
Fossils could expand the anatomical variation represented in the model and supply direct examples of relationships among bones in extinct lineages. I describe a staged implementation using geometric representation learning, prediction of withheld anatomy, and conditional generation, together with experiments separating the contribution of detailed fragment morphology from size and evolutionary context. I further propose connecting Fossil2Skeleton with my Skeleton2DNA and Endocast2Brain proposals to investigate agreement among successive biological inferences. In the motivating scenario, a Gigantopithecus jaw conditions alternative skeletons and skulls; their predicted endocranial geometry supports one set of brain predictions, while genomic hypotheses inferred from the candidate skeletons support another. Agreement would become informative if it predicts accuracy in species with independently measured brains after accounting for shared inputs and correlated errors. The available Gigantopithecus fossils provide no actual braincase, so the entire cranial branch would retain its dependence on inferred anatomy. Initial neurological targets would concern measurable aspects of brain size and organization, with behavioral interpretation requiring additional validation. This article develops the research program, its fossil training strategy, its relationship to existing methods, and the limited synthetic prototype released on GitHub.
Keywords: fossil reconstruction; skeletal integration; machine learning; comparative anatomy; Gigantopithecus; paleoneurology; uncertainty propagation
What a fragment can predict
The same jaw can be compatible with several body configurations. Its size may partly reflect overall growth, while its proportions and tooth geometry may also reflect feeding specialization. A useful model must learn which relationships persist across individuals and species, how they vary with development, and where they become unreliable. It should produce coordinated alternatives whose differences can be inspected and measured.
Let S denote a complete skeletal configuration, F the observed fragment geometry, M the observation record, and C independently available context. S includes bone shapes, dimensions, identities, and standardized anatomical relationships. M records which structures were measured, their preservation, and whether a missing element is unobserved or anatomically absent. The target is a conditional distribution over skeletal configurations:
The parameter set θ is estimated from reference specimens. This formulation requires predictions for different bones to remain related. Separate estimates of pelvic breadth, femur length, and joint size could each appear reasonable while forming an inconsistent animal when combined. Joint generation allows these measurements to covary. It also retains alternative configurations that a single average reconstruction could obscure.
The observation process needs its own representation. Crushing, erosion, incomplete scanning, and measurement error can alter a preserved structure. Candidate reconstructions should be checked against the fossil under documented ranges of those effects. An exact surface match to a distorted fragment can produce a poor reconstruction of the original anatomy, so preservation evidence must constrain the permitted corrections.
Training on living and extinct animals
The training corpus should contain measured anatomy from living and extinct animals. Living specimens can supply associated skeletons, developmental information, and, in some cases, genomic or imaging data. Fossils add observed anatomical configurations from lineages and body plans that living animals may represent incompletely. Their inclusion could expand the relationships the model has an opportunity to learn, including relationships useful for predicting both extinct and living animals.
A sufficiently documented fossil skeleton provides a direct training example. Researchers can conceal its jaw, limbs, or other preserved regions and train the model to predict those measurements from the retained anatomy. A complete skeleton is unnecessary for every example. An associated fossil preserving a jaw, pelvis, and femur can teach the model about those measured relationships, while its unknown vertebrae remain unlabelled.
The contribution depends on association. Isolated fossil jaws expand knowledge of jaw geometry; associated jaws and femora provide information about how those structures covary. Thousands of unrelated isolated bones cannot establish an individual relationship that the collection never records. Specimen identities, anatomical associations, stratigraphic context, and uncertainty about attribution therefore belong in the dataset itself.
Fossil specimens can contribute at several stages. Isolated bones and preserved surfaces can support geometric representation learning. Associated material can support supervised prediction across bones. Specimens with documented damage can help assess preservation effects when a credible target is available. Original material, mirrored portions, comparative substitutions, and sculpted additions in museum mounts must retain separate provenance. The measured portions can supply targets without treating the entire display as observed anatomy.
There is methodological support for expanding reference anatomy in this way. Torres-Tamayo and colleagues (2025), in their study of fossil pelvic reconstruction, recommended inclusion of fossil reference specimens to represent morphology more fully. Their results also show that the composition of a reference sample affects reconstruction. For Fossil2Skeleton, I propose testing the larger hypothesis that fossil training data improve prediction across separate skeletal regions and potentially benefit living targets as well.
The relevant experiment would compare living-only training with training that includes fossils, while controlling sample count, model capacity, taxonomic composition, and measurement quality. Further comparisons should distinguish fossils used only for geometric pretraining from fossils supplying associated anatomical targets. Tests on withheld living and extinct specimens would reveal where the additional variation helps, where it has little effect, and where it requires more selective conditioning or sampling.
Representing anatomy and learning skeletal relationships
A practical system would begin with a documented inventory of homologous skeletal elements within one comparative group. Surface points, mesh patches, landmarks, and dimensions would provide complementary descriptions of each bone. Physical scale should remain available even when shape is normalized for selected computations. Deriving an input normalization from the hidden complete skeleton would leak target information into the prediction.
The first learning stage would train encoders and decoders to represent measured bone geometry. Collections of isolated bones, including fossils, could contribute extensively. Masked reconstruction of surface patches would provide a self-supervised objective of the kind demonstrated for general point clouds by Point-MAE (Pang et al., 2022). Orientation handling would also require care: the placement of disarticulated fragments on a scanner bed has little anatomical meaning. Equivariant geometric networks offer one possible computational approach (Satorras et al., 2021).
The second stage would train prediction across associated skeletal elements. The input would vary from multiple bones to a single jaw or a small fragment. Observation masks should include correlated loss of neighboring regions, as well as random deletion, so the training task reflects plausible preservation patterns. Target availability would be recorded separately from the input mask, allowing supervised loss only where a measured target exists.
A proposed full implementation would combine fragment tokens with candidate-bone tokens. Cross-attention would let candidate bones consult the preserved evidence; attention among candidate bones would help coordinate dimensions and articulation. Transformer-based point-cloud completion provides a relevant computational precedent in PoinTr (Yu et al., 2021). The skeletal application adds anatomical identities, relationships across separate bones, and a requirement to test transfer between biological groups.
Conditional diffusion could generate multiple coordinated solutions. During training with documented target skeletons, noise would be added to skeletal representations and the model would learn to predict that corruption while receiving the fragments as context. At inference, fixed model weights would guide successive denoising updates from random candidate states. Point-voxel diffusion demonstrates probabilistic shape completion, and latent diffusion provides a general strategy for conducting generation in compressed representations (Zhou et al., 2021; Rombach et al., 2022).
Geometry-space checks would evaluate whether decoded candidates match preserved surfaces and respect joint relationships. Bone count, fusion, and anatomical absence would require explicit states; an initial fixed inventory could later be expanded to handle variation in skeletal organization. For fully latent generation, incomplete targets require a specified missing-data treatment or appropriately bounded training tasks. Filling unknown target anatomy with a generated completion and relabelling it as measured would create an artificial learning signal.
Testing the anatomical information in a fragment
The first biological study should use a curated group of associated specimens and a limited set of reproducible measurements. Jaw geometry could predict withheld long-bone lengths, joint dimensions, pelvic measurements, or skeletal landmarks. Errors should be reported separately for completion within a bone, prediction across nearby structures, and prediction across distant regions. Those tasks test different levels of anatomical dependence.
The main comparisons would include size-based regression, a nearest-relative reconstruction, a statistical shape model, and a predictor receiving evolutionary context without detailed fragment geometry. A control that shuffles fragment–target pairings within suitable size and taxonomic strata would test whether meaningful anatomical associations drive performance. Sensitivity analyses should specify which size information each comparator receives, since a jaw provides an imperfect estimate of the animal’s overall scale.
Evaluation should exclude individuals, whole species, and larger evolutionary groups in separate experiments. Repeated scans, casts, alternative reconstructions, and every mask of the same specimen must remain together. A species exclusion claim also requires control over geometric pretraining, template construction, hyperparameter selection, and any external data that reveal the withheld anatomy. These conditions apply equally to fossil and living specimens.
Uncertainty should be measured against actual outcomes. Across comparable test cases, intervals labelled as 90 percent predictive intervals should contain the measured target about 90 percent of the time, with their widths reported alongside coverage. Regional calibration and performance on unusual proportions are especially relevant to Gigantopithecus. Sensitivity to reference collections and to model assumptions would reveal forms of uncertainty that repeated sampling from a single trained model might miss.
Controlled removal and restoration of observations could produce an anatomical information map. Such a map would record which fragment regions improve prediction of which missing structures on withheld specimens. It could also identify combinations of fragments whose joint predictive value exceeds their separate contributions. Expected information gain could then rank additional measurements by their anticipated reduction of uncertainty, with the ranking evaluated by actually restoring those measurements in held-out cases.
Gigantopithecus and inference through multiple pipelines
My proposed scenario begins with a Gigantopithecus jawbone. Fossil2Skeleton uses its geometry and permitted comparative information to generate plausible skeletons, including candidate skulls. These candidates preserve alternatives in body proportions and cranial structure. The next step follows two paths toward a common set of questions about the brain.
Along the anatomical path, each candidate skull supplies a predicted internal cranial cavity. An endocast derived from that cavity becomes the input to my proposed Endocast2Brain system, which would infer aspects of brain anatomy from endocranial geometry. For Gigantopithecus, that endocast would be synthetic because the known material supplies no actual braincase. The mapping from a jaw to the skull interior therefore needs its own uncertainty and validation, including variation in cranial thickness and internal geometry.
Along the genomic path, the candidate skeletons enter my proposed Skeleton2DNA system, extending the skeleton-to-genome program introduced in my earlier dinosaur reconstruction work (Reser, 2025). This system would generate genomic hypotheses or bounded sets of genomic features consistent with skeletal morphology and other permitted evidence. A further mapping from those genomic hypotheses to brain traits would yield a second set of neurological predictions. Enamel proteins could add measured molecular constraints when their specimen and population relationships are represented correctly.
I would then compare what the genomic path predicts about the brain with what the endocranial path predicts. Do they support overlapping ranges of brain volume? Do they predict compatible proportions or other shared anatomical measures? Does their agreement identify conclusions that remain stable across many candidate skeletons? The longer-term question is whether validated convergence could constrain aspects of Gigantopithecus neurology and, with further evidence, behavior.
Each link requires relevant paired data. Endocast2Brain would need cranial and brain measurements from known animals, with regional targets defined explicitly. Dumoncel and colleagues (2021) directly compared brain and endocast geometry in the same living human individuals, establishing an empirical way to investigate that relationship. A comparative application to extinct apes would require suitable nonhuman data and tests of transfer.
The genomic branch requires both anatomy–genome and genome–brain relationships. García-Marín and colleagues (2024) demonstrated associations between genomic variation and intracranial or subcortical volumes in humans. This provides evidence for particular genotype–brain relationships within the studied populations; its use in an extinct ape would require a separate comparative model. Initial genomic targets should be features for which predictive relationships can be tested, with any extension to whole genomes evaluated on its own terms.
The inference should propagate distributions through both paths. Choosing one attractive skeleton, one genome, and one brain would discard much of the unresolved information. An ensemble can instead track which neurological predictions persist across skeletal, genomic, and model alternatives. Some chains may remain broad throughout, while others may concentrate on a small number of measurable outcomes.
When agreement between predictions becomes informative
Building inferences on inferences is a legitimate modelling strategy when the dependencies are represented and the results are tested. Biological intermediates could connect datasets that are difficult to combine directly, allowing learned anatomical and developmental relationships to constrain a shared outcome. The additional knowledge comes from comparative observations used to train the models. Repeatedly transforming one jaw creates further predictions whose dependence on that jaw must remain explicit.
A joint model can express this requirement without assuming independent prediction paths. Let B represent specified brain traits, S skeletal anatomy, E endocranial geometry, G genomic variables, and P any measured ancient molecular evidence. Brain prediction would integrate over the uncertain intermediate variables:
The expression states that skeletal, cranial, and genomic alternatives must be considered together under the observed evidence and model assumptions, retaining the observation record M. The integral is schematic, with sums replacing integration for discrete alternatives. When molecular and anatomical material come from different individuals, the model needs separate individual variables linked through a population or species model. Their association should never be invented to simplify the calculation.
Agreement could occur because both paths are accurate, because both share an erroneous anatomical candidate, or because both largely reproduce the same evolutionary prior. It could also be selected artificially by retaining only mutually compatible outputs from inaccurate generators. A combination rule must therefore account for correlated errors and be evaluated after any agreement-based selection. Multiplying the two output distributions as if they were independent would generally require an assumption that this design has not established.
The decisive experiment would conceal the corresponding evidence from a living species whose skeleton, genome, and brain are independently documented. Starting from a jaw and the allowed contextual information, researchers would run both paths and compare their predictions with the actual brain. Evolutionary-context and body-size predictors, a direct jaw-to-brain model, each separate path, and the combined system would form the principal comparisons. The gain from using intermediate anatomy and genomic hypotheses would then be measurable.
Agreement becomes scientifically useful if it predicts smaller errors, improved probability scores, or informative uncertainty intervals beyond those comparisons. This must hold on species excluded from the relevant training and calibration stages, including genomic pretraining where the claim concerns unseen sequences. Chance overlap should be assessed with null models, and shared training labels or reference templates should be audited. Disagreement is also a useful output when it identifies an intermediate representation or learned relationship that fails.
Neurological and behavioral interpretation
The first common targets should be measurable brain traits that both paths can actually predict. Total brain volume and coarse regional dimensions are more tractable starting points than detailed connectivity or cellular organization. Comparison requires a common outcome scale: an inferred developmental pathway and an endocranial shape estimate do not constitute agreement until their implications for the same measured brain trait have been established.
A genomic hypothesis could be informative about a neural feature that has little visible expression in an endocast. Conversely, endocranial geometry might constrain a gross anatomical feature more directly than the selected genomic variables do. The framework should retain such differences. It can test overlap where comparable outputs exist while reporting the additional predictions unique to each path.
Behavioral interpretation requires another validated relationship between anatomy, development, and observed behavior. A prediction concerning brain size or regional proportions does not uniquely determine an animal’s social organization, temperament, learning capacities, or subjective experience. Relevant comparative studies could establish bounded associations with specified behaviors, after accounting for factors such as body size and ecological context. The resulting Gigantopithecus claims would inherit the uncertainty of the entire chain.
Independent paleoecological evidence could test the coherence of some functional hypotheses. The Gigantopithecus extinction study used environmental records and dental analyses to investigate changing food availability and responses to environmental change (Zhang et al., 2024). Such observations can constrain proposed feeding or ecological scenarios. They should be reserved for independent evaluation when the experiment claims a new prediction, and their use as conditioning evidence should be recorded when they enter the model.
Related methods and the scope of the contribution
Virtual fossil reconstruction and geometric morphometrics already provide methods for estimating missing and distorted anatomy (Gunz et al., 2009). Statistical shape modelling has been applied to fossil pelvic reconstruction, while Wolcott’s current project describes generative modelling of lizard vertebrae and fossil reconstruction tasks (Torres-Tamayo et al., 2025; Wolcott, 2026). These works address relevant components and anatomical settings.
The computational ingredients also have established precedents: masked geometric learning, transformer completion, equivariant processing, and probabilistic shape generation. My proposal concerns their use within an integrated research program spanning variable fossil fragments, coordinated whole-skeleton inference, training across living and extinct animals, and comparison of downstream biological predictions. I formulated the central Fossil2Skeleton concept before identifying the related reconstruction literature during development of the first article. The references locate the relevant methods and overlapping tasks at their demonstrated scope.
The combined dataset remains the principal practical requirement. Digital collections such as openVertebrate provide an infrastructure for obtaining three-dimensional comparative specimens (Blackburn et al., 2024). Building the proposed training corpus would still require curation of association, taxonomic identity, maturity, physical scale, preservation, and access conditions. Brain and genomic tasks would add their own pairing requirements; broad geometric coverage alone cannot supply those missing links.
The public prototype and companion articles
A Python prototype was developed with ChatGPT under my direction and released in the Fossil2Skeleton GitHub repository. The source is explicitly labelled as a reconstructed implementation because an earlier prototype archive was unavailable. It supplies a narrow software starting point for masked three-dimensional landmark completion and preserves the distinction between this demonstration and the full proposed framework (Reser, 2026c).
The implementation includes a synthetic shape generator, observation masks, mean-shape and nearest-reference baselines, and a PyTorch landmark transformer. The transformer combines coordinate, landmark-identity, and observation-state embeddings to predict coordinates. Its completion routine preserves observed points, while repeated predictions with dropout provide an exploratory spread estimate. This spread has not been calibrated as biological uncertainty.
The baseline example and nine core utility tests were run successfully. The tests exercise synthetic generation, masks, baseline completion, and scoring; they do not test the biological accuracy of the transformer. Transformer training was not executed in the development environment because PyTorch was unavailable there. The repository contains no trained checkpoint or specimen dataset, and the synthetic generator builds the cross-landmark dependencies it asks the demonstration to recover. It therefore checks software mechanics while leaving the central anatomical hypothesis for empirical study.
The full model would require the associated specimen corpus, richer geometry, conditional generation, and biological evaluations described above. The repository also contains the earlier manuscript and the proposed pipeline illustration. The following resources document the program and its current implementation:
Fossil2Skeleton source code
https://github.com/jaredreser/Fossil2Skeleton
Initial Fossil2Skeleton article on Iterated Insights
https://iteratedinsights.com/2026/10/21/fossil2skeleton/
Gigantopithecus companion article on Iterated Insights
https://iteratedinsights.com/2026/10/21/reconstructing-gigantopithecus-from-fossil-jaws-to-skeletal-and-genomic-hypotheses/
Applications and an experimental path forward
The first biological implementation should concentrate on a manageable comparative group, associated skeletons, and measured anatomical targets. It should establish the contribution of detailed fragment geometry before extending generation to complete surfaces. Training on documented fossils can then be introduced through controlled comparisons, with separate tests of benefits for living and extinct targets. Model complexity should follow the information and sample sizes available for each task.
Historical discovery sequences offer useful reconstruction tests. Additional Deinocheirus specimens revealed anatomy that could not be observed when the animal was known primarily from its forelimbs (Lee et al., 2014). A retrospective experiment could supply only the earlier evidence and test predictions against the later anatomy, provided training material and descriptions are audited to prevent leakage. Publishing quantitative predictions before future discoveries would provide a stronger prospective test.
Other applications include incomplete dinosaur skeletons, extinct mammals, archaeological animal remains, and museum reconstructions whose inferred portions remain visible. Anatomical ensembles could support sensitivity analyses of mass, posture, or locomotion, analogous to the importance of reconstruction assumptions in dinosaur mass modelling (Bates et al., 2009). A poor skeletal fit could also identify an assemblage requiring closer investigation, with mixed individuals, deformation, and unusual anatomy treated as separate explanations.
Gigantopithecus would remain a demanding extrapolation even after progress on living species. Its value as a motivating case is that it makes the information deficit concrete and directs attention to measurable constraints. The research should establish which predictions become narrower when the actual jaw is supplied, which remain dominated by comparative expectations, and which new observations would discriminate among the remaining alternatives.
Conclusion
Fossil2Skeleton proposes learning anatomical relationships across the combined record of living and extinct animals. Preserved fossils can contribute directly to that learning wherever their geometry and associations are documented. The resulting model would estimate coordinated alternatives for missing anatomy and identify the observations that make those alternatives more precise.
The Gigantopithecus scenario extends this strategy from skeletal reconstruction toward neurological inference. Candidate skull interiors and candidate genomic features could support different predictions of the same brain traits. If their agreement reliably identifies greater accuracy on animals with known brains, that convergence could become a useful tool for investigating an extinct animal whose soft tissues are unavailable. The scientific aim is to determine how far the surviving evidence constrains anatomy, neurology, and eventually behavior, while preserving the dependencies through which every conclusion was reached.
Origin of the proposal
After recently watching Marq Evans’s 2026 documentary Capturing Bigfoot, I began thinking about Gigantopithecus and how little of its anatomy survives. The documented remains of this enormous extinct ape consist of jawbones and teeth. What could a sufficiently well trained model learn about the rest of the animal from those structures? That question led me to formulate Fossil2Skeleton on October 21, 2026. The film supplied the occasion for the idea; the reconstruction problem arose from the surviving Gigantopithecus material and the possibility of learning anatomical relationships from comparative specimens.
The 2024 investigation by Zhang and colleagues described four mandibles and almost 2,000 isolated teeth, with no postcranial remains. Ancient enamel proteins provide an additional source of information: Welker and colleagues (2019) placed Gigantopithecus within the orangutan lineage using proteins recovered from a fossil molar. These observations establish a restricted but substantial anatomical and molecular starting point. They leave the skull, limbs, vertebral column, and brain largely dependent on inference.
Fossil2Skeleton would learn how surviving fragments relate to the rest of a skeleton. A jaw and a femur develop within the same organism, under shared genetic, developmental, and functional influences. Their relationships may therefore contain predictive information that can be measured across specimens. The central experiment is to determine whether detailed fragment geometry predicts missing anatomy beyond the information already supplied by body size, evolutionary relatedness, and reference averages.
The present article extends two companion papers on Iterated Insights: the initial Fossil2Skeleton framework and its application to Gigantopithecus (Reser, 2026a, 2026b). The extensions developed here concern explicit inclusion of fossil specimens in training and the comparison of neurological predictions reached through different inferential paths. I introduce these as testable components of my broader reconstruction program. Their value will depend on measured predictive performance and on how clearly the models preserve the distinction between observed material and generated anatomy.
References
Bates, K. T., Manning, P. L., Hodgetts, D., & Sellers, W. I. (2009). Estimating mass properties of dinosaurs using laser imaging and 3D computer modelling. PLOS ONE, 4(2), e4532. https://doi.org/10.1371/journal.pone.0004532
Blackburn, D. C., et al. (2024). Increasing the impact of vertebrate scientific collections through 3D imaging: The openVertebrate (oVert) Thematic Collections Network. BioScience, 74(3), 169–186. https://doi.org/10.1093/biosci/biad120
Dumoncel, J., Subsol, G., Durrleman, S., Bertrand, A., de Jager, E., Oettlé, A. C., Lockhat, Z., Suleman, F. E., & Beaudet, A. (2021). Are endocasts reliable proxies for brains? A 3D quantitative comparison of the extant human brain and endocast. Journal of Anatomy, 238(2), 480–488. https://doi.org/10.1111/joa.13318
Evans, M. (Director). (2026). Capturing Bigfoot [Documentary film]. https://www.capturingbigfoot.com/about
García-Marín, L. M., et al. (2024). Genomic analysis of intracranial and subcortical brain volumes yields polygenic scores accounting for variation across ancestries. Nature Genetics, 56, 2333–2344. https://doi.org/10.1038/s41588-024-01951-z
Gunz, P., Mitteroecker, P., Neubauer, S., Weber, G. W., & Bookstein, F. L. (2009). Principles for the virtual reconstruction of hominin crania. Journal of Human Evolution, 57(1), 48–62. https://doi.org/10.1016/j.jhevol.2009.04.004
Lee, Y.-N., et al. (2014). Resolving the long-standing enigmas of a giant ornithomimosaur Deinocheirus mirificus. Nature, 515, 257–260. https://doi.org/10.1038/nature13874
Pang, Y., Wang, W., Tay, F. E. H., Liu, W., Tian, Y., & Yuan, L. (2022). Masked autoencoders for point cloud self-supervised learning. European Conference on Computer Vision. https://arxiv.org/abs/2203.06604
Reser, J. E. (2025, July 18). How to use AI to reconstruct dinosaur genomes from bird DNA and skeletons. Observed Impulse. https://www.observedimpulse.com/2025/07/ai-mediated-reconstruction-of-dinosaur.html
Reser, J. E. (2026a, October 21). Fossil2Skeleton: Probabilistic reconstruction of complete skeletal anatomy from fossil fragments. Iterated Insights. https://iteratedinsights.com/2026/10/21/fossil2skeleton/
Reser, J. E. (2026b, October 21). Reconstructing Gigantopithecus: From fossil jaws to skeletal and genomic hypotheses. Iterated Insights. https://iteratedinsights.com/2026/10/21/reconstructing-gigantopithecus-from-fossil-jaws-to-skeletal-and-genomic-hypotheses/
Reser, J. E. (2026c). Fossil2Skeleton (Version 0.1.0 reconstruction) [Computer software]. GitHub. Repository state inspected October 22, 2026, commit 0a4aa2247a63561abbaf90ffd53dcd3a435fda59. https://github.com/jaredreser/Fossil2Skeleton
Rombach, R., Blattmann, A., Lorenz, D., Esser, P., & Ommer, B. (2022). High-resolution image synthesis with latent diffusion models. IEEE/CVF Conference on Computer Vision and Pattern Recognition. https://arxiv.org/abs/2112.10752
Satorras, V. G., Hoogeboom, E., & Welling, M. (2021). E(n) equivariant graph neural networks. Proceedings of Machine Learning Research, 139, 9323–9332. https://proceedings.mlr.press/v139/satorras21a.html
Torres-Tamayo, N., VanSickle, C., Woodert, M. A., Haeusler, M., & Schlager, S. (2025). Statistical shape modelling as a novel reconstruction tool in palaeoanthropology: A case study on fossil pelves. Methods in Ecology and Evolution, 16, 2569–2584. https://doi.org/10.1111/2041-210X.70154
Welker, F., et al. (2019). Enamel proteome shows that Gigantopithecus was an early diverging pongine. Nature, 576, 262–265. https://doi.org/10.1038/s41586-019-1728-8
Wolcott, K. A. (2026). 3D generative AI for lizard vertebrae [Research project description]. Accessed October 22, 2026. https://katherineawolcott.com/3d-gen-ai
Yu, X., Rao, Y., Wang, Z., Liu, Z., Lu, J., & Zhou, J. (2021). PoinTr: Diverse point cloud completion with geometry-aware transformers. IEEE/CVF International Conference on Computer Vision. https://arxiv.org/abs/2108.08839
Zhang, Y., et al. (2024). The demise of the giant ape Gigantopithecus blacki. Nature, 625, 535–539. https://doi.org/10.1038/s41586-023-06900-0
Zhou, L., Du, Y., & Wu, J. (2021). 3D shape generation and completion through point-voxel diffusion. IEEE/CVF International Conference on Computer Vision. https://arxiv.org/abs/2104.03670
No comments:
Post a Comment