Can convolutional neural networks approximate the phylogenetic tree of human populations with face images? 1

Abstract

Inferring phylogenetic trees in human populations is a challenging task that has traditionally relied on genetic, linguistic, and geographic data. In this study, we explore the application of Deep Learning and facial embeddings for phylogenetic tree inference based solely on facial features. We use pre-trained ConvNets as image encoders to extract facial embeddings and apply hierarchical clustering algorithms to construct phylogenetic trees. Our methodology differs from previous approaches in that it does not rely on preconstructed phylogenetic trees, allowing for an independent assessment of the potential of facial embeddings to capture relationships between populations. We have evaluated our method with a dataset of 30 ethnic classes, obtained by web scraping and manual curation. Our results indicate that facial embeddings can capture phenotypic similarities between closely related populations; however, problems arise in cases of convergent evolution, leading to misclassifications of certain ethnic groups. We compare the performance of different models and algorithms, finding that using the model with ResNet50 backbone and the face recognition module yields the best overall results. Our results show the limitations of using only facial features to accurately infer a phylogenetic tree and highlight the need to integrate additional sources of information to improve the robustness of population classification.

Keywords

Convolutional neural networks deep learning hierarchical clustering phylogenetic tree

1 Introduction

Phylogenetic analysis of human populations and other species traditionally relies on genetic sequences, including chromosomal DNA [5, 17], mitochondrial DNA (mDNA) [19, 28], and other sequences such as amino-acid chains [39]. However, in cases where genome data is unavailable, phenotypic features are still utilized to derive Phylogenetic trees. These features are typically determined by experts, who rely on measurements and the quality of the samples.

Convolutional Neural Networks (ConvNets) have emerged as state-of-the-art approaches in various vision-related tasks, including image classification [6, 10, 20, 23], image segmentation [22], inpainting [42], and more. These models have demonstrated exceptional performance and have been successfully applied in medical imaging, assisting experts in different tasks [20].

In this study, we explore the possibility of using ConvNets to approximate the phylogenetic tree of human populations from facial images. By taking advantage of the ability of ConvNets to extract meaningful features from images, we aim to create a model that can capture important information about population structure from facial features.

The motivation for this research is to investigate an alternative approach to infer genetic relationships and ancestry when genetic data are limited or unavailable. Facial images are readily available and, in the case of historical photographs, may be the only source of ethnicity information.

In this paper we present a methodology, experimental setup, and results. We compare the predicted phylogenetic tree generated by our ConvNet model with known genetic ancestry information obtained from the literature. We discuss possible applications and limitations of this approach.

Overall, our study aims to contribute to the field of population genetics by exploring the capabilities of Convolutional Neural Networks in approximating the phylogenetic tree of human populations based on facial images. By merging the domains of computer vision and phylogenetics, we hope to provide valuable insights for various fields, including forensic anthropology and human migration studies.

2 Related work

One of the primary applications of Deep Learning in the biosciences is phylogeny inference [2, 3, 13, 16, 31, 37, 43]. However, research conducted by Zaharias et al. [41] showed that ConvNets do not outperform well-established algorithms in comparative results.

Rather than employing deep architectures for tree inference, our approach focuses on utilizing Deep Learning to extract essential features that can be utilized by standard tree inference methods. However, there is limited literature available on this specific approach.

Kiel [14] conducted a study on tree generation by employing ConvNets for feature extraction, which were then used to construct a distance matrix. Subsequently, a hierarchical clustering process was applied to construct the phylogenetic tree. This study specifically focused on analyzing images of bivalves.

In another study, Hunt and Pedersen [12] introduced the hierarchically structured dataset Rove-Tree-11, comprising images of pinned insects obtained from museums. The authors followed a similar procedure by extracting features using ConvNets and subsequently applying Bayesian algorithms for tree construction.

To our knowledge, there is no existing work specifically addressing the application of ConvNets for phylogeny inference using human faces. However, it is worth noting that the related problem of ethnic classification has been explored using ConvNets. For instance, Masood et al. [21] utilized a ConvNet to categorize three primary ethnic groups, achieving superior results compared to traditional methods. Vo et al. [38] reported marginally improved outcomes in ethnic classification using a pre-trained VGG-16. In their study, the authors distinguished one ethnic group from others. Similarly, Wu et al. [40] employed VGG-16 to classify three principal ethnic categories. Das et al. [7] proposed a ConvNet for the simultaneous classification of gender, age, and ethnicity, aiming to reduce associated biases. Srinivas et al. [36] also undertook joint classification of age, gender, and ethnicity, with a specific focus on Asian populations.

3 Methodology

In our study, we adopted the methodology introduced by Kiel et al. [14] and Hunt and Pedersen [12]. We employed pre-trained ConvNets or image encoders to extract the essential features from the images. Subsequently, we applied hierarchical clustering to infer the phylogenetic tree. Additionally, we compared our results with established phylogenetic trees of human populations.

3.1 Ground truth

The phylogenetic trees of human populations have been extensively discussed; however, a clear consensus regarding the positioning of certain ethnic groups is yet to be established. In the work of Duda and Zrzavý, three potential phylogenetic tree models, labeled as Model A, Model B, and Model C (see Fig. 2 in [8]), were presented. We utilized these models as a basis for comparison in our study.

3.2 Data acquisition

Data acquisition was carried out through web scraping of images corresponding to the 55 human populations studied by Duda and Zrzavý. We downloaded 30 images for each population using the Python module simple_image_download. This approach was chosen to reduce possible biases introduced by manual selection, although we manually excluded five categories: Druze, She, Daur, Naxi, Tu and Jul’hoah.

We employed the face recognition module [9] to detect, select, and crop the faces from the downloaded images. Images that did not yield a recognizable face were removed. This process helped to eliminate noisy data points. To enhance the reliability of our results, we only retained classes with a minimum of 10 images. In total, this process resulted in 30ethnic classes with 419 images.

This dataset is rather small. However, we implemeted pre-trained visual encoders (see Subsection 3.3). Thus, a larger corpus is not needed for this study, although it might be recommendable to capture different variations between populations from the same ethnic group.

3.3 Visual encoders

Once the faces were selected, we employed three different image encoders:

1.
VGG-16 [33], pre-trained on the VGGFaces dataset [26].
2.
ResNet50 [10], pre-trained on the VGGFaces dataset [26].
3.
Face Recognition module [9, 15].

We performed two subtasks: 1.
Subtask A: Ethnic Classification. We allocated 33% of the dataset for testing purposes.
2.
Subtask B: Tree Inference. We computed the mean embeddings per class and applied traditional tree inference algorithms, including UPGMA (unweighted pair group method with arithmetic mean) [35], Neighbor Joining (NJ) [30], and Hierarchical Agglomerative Clustering (HAC) provided by Scikit-Learn [27]. In all cases, we used Euclidean distance.

3.4 Metrics

For subtask B, where we generated trees using the embeddings, we employed the following metrics to compare the generated trees with the theoretical trees:

1.
Smith’s Tree Distance (TD) [34].
2.
Robinson-Foulds Distance (RF) [29].
3.
Matching Split Distance (MSD) [4, 18].
4.
Clustering Information Distance [34].
5.
Shared Phylogenetic Information (SPI) [24].
6.
Nye Similarity (NS) [25].
The first 4 computes the distance between phylogenetic trees, while the Shared Phylogenetic Information and Nye Similarity measure how closely related are two trees.
4 Results

In the following section, we present the results of our study. These findings provide insights into the performance of various pre-trained models in the complex task of ethnic classification. The results are divided into the two mentioned subtasks, each addressing a unique aspect of the problem. The detailed outcomes and their implications are discussed in the subsequent sections.

4.1 Subtask A

Ethnic classification with 30 classes proved to be challenging for the pre-trained models, with the highest achieved F1-macro score of 0.29043 obtained using the kNN algorithm with k = 3. The main results are summarized in Table 1. The difficulty of the task can be attributed to the presence of similar classes within the dataset.

Table 1
Measured metrics for the subtask A

Model F1-macro F1-micro F1-weighted Accuracy

VGG-16 0.20282 0.24460 0.21250 0.24460

ResNet50 0.19400 0.25899 0.21840 0.25899

Face Recognition 0.29043 0.35971 0.32073 0.35971

Model	F1-macro	F1-micro	F1-weighted	Accuracy
VGG-16	0.20282	0.24460	0.21250	0.24460
ResNet50	0.19400	0.25899	0.21840	0.25899
Face Recognition	0.29043	0.35971	0.32073	0.35971

4.2 Subtask B

The main results for Subtask B are presented in Tables 2–4. The results demonstrate that ResNet50 and the Face Recognition module performed better compared to VGG-16. Specifically, ResNet50 with UPGMA (see Figs. 1–3) achieved superior results in most metrics, except for Robinson-Foulds and Clustering Information Distance. On the other hand, the Face Recognition module produced better results in Robinson-Foulds (with UPGMA, see Fig. 4) and Clustering Information Distance (with Hierarchical Agglomerative Clustering).

Table 2
Results on VGG-16

Algorithm TD RF MSD CID MCI NS

Model A

UPGMA 0.688575 54 121 22.542632 100.015146 10.870847

NJ 0.566358 50 82 19.033985 159.771058 13.928667

HAC 0.640707 53 107 20.145751 109.136906 13.007681

Model B

UPGMA 0.710208 53 136 23.750598 93.962567 11.010331

NJ 0.548491 49 89 18.819446 172.615237 14.584915

HAC 0.655686 54 118 21.078091 105.432789 12.318192

Model C

UPGMA 0.731870 54 145 24.415879 87.875272 9.734230

NJ 0.577056 48 103 19.752918 164.912569 13.535632

HAC 0.680633 53 123 21.825063 100.937681 11.944767

Algorithm	TD	RF	MSD	CID	MCI	NS
Model A
UPGMA	0.688575	54	121	22.542632	100.015146	10.870847
NJ	0.566358	50	82	19.033985	159.771058	13.928667
HAC	0.640707	53	107	20.145751	109.136906	13.007681
Model B
UPGMA	0.710208	53	136	23.750598	93.962567	11.010331
NJ	0.548491	49	89	18.819446	172.615237	14.584915
HAC	0.655686	54	118	21.078091	105.432789	12.318192
Model C
UPGMA	0.731870	54	145	24.415879	87.875272	9.734230
NJ	0.577056	48	103	19.752918	164.912569	13.535632
HAC	0.680633	53	123	21.825063	100.937681	11.944767

Table 3

Results on ResNet50

Algorithm	TD	RF	MSD	CID	MCI	NS
Model A
UPGMA	0.490349	50	65	16.169379	186.665055	15.703860
NJ	0.534697	52	69	17.000377	148.246081	14.135766
HAC	0.529579	50	73	16.604242	158.694323	14.708359
Model B
UPGMA	0.517162	49	76	17.417423	179.372811	15.455678
NJ	0.559508	49	81	18.182932	148.880003	14.105193
HAC	0.573918	51	93	18.398273	139.900618	14.013012
Model C
UPGMA	0.515838	50	86	17.331158	182.498941	14.904587
NJ	0.611447	50	104	19.821444	134.149158	12.425904
HAC	0.579466	52	107	18.529285	140.351354	13.198341

Table 4

Results on Face Recognition module

Algorithm	TD	RF	MSD	CID	MCI	NS
Model A
UPGMA	0.531971	46	71	17.05631	157.234	15.2347
NJ	0.5171284	50	71	17.15543	178.6264	14.22633
HAC	0.5153341	46	71	15.92858	160.4522	15.19407
Model B
UPGMA	0.554995	45	81	18.185035	155.459966	14.821004
NJ	0.538441	49	78	18.241332	172.544834	14.444354
HAC	0.554376	47	89	17.525435	146.513155	14.546637
Model C
UPGMA	0.520160	44	93	17.001595	171.200408	14.900155
NJ	0.565848	50	95	19.124111	164.632439	13.602013
HAC	0.571622	48	107	18.024423	143.422043	13.388849

Fig. 1

Phylogenetic tree generated with UPGMA and ResNet50 embeddings.

Fig. 2

Distance matrix generated with UPGMA and ResNet50 embbeddings.

Fig. 3

Visualization of the Shared Phylogenetic Information of the tree generated with UPGMA and ResNet50 embeddings (b) compared to the ground truth (a).

Fig. 4

Phylogenetic tree generated with UPGMA and Face Recognition embeddings.

5 Discussion

Phylogenetic trees are primarily constructed using DNA samples from various individuals to capture differences and similarities among populations. These trees identify characteristics that can determine whether an individual belongs to a specific ethnic group and their relatedness to others. Biologists and experts often rely on DNA due to its complexity and reliability. However, obtaining DNA samples involves considerable time and effort as it requires human participation.

This research aims to develop a Machine Learning approach to estimate phylogenetic trees from an alternative data source: facial images. This could have a positive impact on both biology and genetics, as Deep Learning algorithms can help identify similarities among humans without actual genetic data. This research may also offer new insights into the relevance of phenotypical facial features and characteristics for analyzing related populations.

This work is based on the hypothesis that different genetic populations produce different phenotypic populations. While this is a general fact in biology, environmental factors can also influence phenotypic expression [1]. Recent statistical or Machine Learning-based studies provide evidence that it is possible to differentiate face shapes/images. For example, Hopman et al. [11] found that it is possible to differentiate two phylogenetically related populations (Dutch and English populations in this case) based solely on facial traits. If this holds true for all genetically-related populations, it suggests that Machine/Deep Learning approaches could be used to distinguish even similar populations based solely on images.

Other works aim to predict facial structure using genetic information (DNA-to-face approaches) and the inverse problem (face-to-DNA approaches) [1]. Mathematically, let $X$ represent the set of possible genetic codes and $Y$ the set of all possible facial images. For simplicity, we will only consider sequences that encode facial traits. For each DNA sequence, we can encode a face, implying a function $f : X \to Y$ where f (x) is the encoded face for a given genetic sequence. There is an approximate function $\hat{f} \approx f$ deduced from data. Although this function f is not invertible (since two or more sequences might encode the same face), we consider a pseudo-inverse function g. Thus, face-to-DNA approaches attempt to estimate $\hat{g} \approx g$ .

Is it possible to estimate the Phylogenetic Tree using only facial images? To answer this question, let us first make some assumptions:

A
It is possible to effectively estimate the Phylogenetic Tree in $X$ .
B
It is possible to effectively estimate g with $\hat{g}$ .

Given assumptions A and B, it is reasonable to conclude that it is possible to approximate the Phylogenetic Tree using only the set of images $Y$ . It is important to note that we are also assuming that it is possible to effectively estimate the function $h : Y \to Y^{'}$ , where $Y^{'}$ is the set of face morphologies. In this context, $g : Y^{'} \to X$ . Therefore, by using the function $\hat{h} \circ \hat{g}$ , we can convert a given image to a genetic sequence, which can then be used to construct the Phylogenetic Tree. This provides a positive answer to one of the main theoretical questions. However, it does not address whether Convolutional Neural Networks are capable of performing such a task.

Empirically, this work suggests that ConvNets are somewhat capable of deriving Phylogenetic Trees, but they do present some significant limitations. Face embeddings serve as phenotypic features that preserve similarity between related groups in our study. The hierarchical clustering method performed well for closely related human populations, but like other phenotype-based phylogenetic tree induction approaches, it is not robust in cases of convergent evolution. While Kiel et al. (2021) [14] addressed this issue by considering a supertree approach using well-known phylogenetic trees, our objective was to compare the potential use of embeddings without relying on published trees.

Another approach proposed by the same Kiel et al. is the use of model ensembles, where predictions from different taxonomic models are combined. This approach requires re-training the networks to classify larger clusters of human populations, making it supervised and reliant on the previous construction of the tree.

In our case, we do not depend on pre-built trees as we applied transfer learning with the VGGFaces dataset, which addresses individual face classification. Therefore, the ConvNets were trained to classify non-hierarchically clustered labeled classes to produce the necessary embeddings. An unsupervised approach, such as using an autoencoder architecture, could also be employed for this purpose [32].

However, while this approach is independent of pre-built phylogenetic trees, it sacrifices the ability to differentiate similar features arising from convergent evolution. Without additional information, this problem may be difficult to overcome. Although it might be useful to add other sources of information, the objective of this work was to rely solely on images.

In our research, the Australian and Makrani ethnic groups were consistently misclassified by most of the proposed approaches. Makrani, a Central Asian population, was often clustered with sub-Saharan African populations in the generated trees. A similar issue was observed with the Australian population.

Apart from these mentioned groups, the methodology generally produced consistent results. In the best model (Fig. 2), we observe three primary clusters: African, Euroasiatic, and Eastern Asian-Oceanic-American populations. However, although the African and Euroasiatic clusters form a larger cluster, the Eastern Asian-Oceanic-American cluster appears to be significantly different, which might not be consistent with the models. In fact, the Uyghur-Hazara populations, related to other Central Asian populations, are clustered with the Yakut, another Turkic-speaking population, in this particular case.

Additionally, despite Model A generally being the closest model to the generated trees, our best tree supports the idea that the Cambodian-Lahu population is clustered with Polynesian groups (Model C), and it indicates the proximity between the Uyghur-Hazara peoples and the Yakut ethnic group, as shown in Models B and C.
6 Conclusions

Our study explored the application of Deep Learning and face embeddings for phylogenetic tree inference in human populations. We utilized pre-trained ConvNets and image encoders to extract facial features, and then applied hierarchical clustering algorithms to construct phylogenetic trees. Our research aimed to compare the potential use of embeddings without relying on pre-built phylogenetic trees.

We found that while face embeddings captured phenotypic similarities between closely related human populations, the methodology encountered challenges in cases of convergent evolution. The hierarchical clustering approach performed well overall, but misclassified certain ethnic groups, such as the Australian and Makrani populations. These misclassifications highlight the limitations of solely relying on facial features for accurate classification and tree inference.

Comparing the results of different models and algorithms, we observed that ResNet50 and the Face Recognition module produced better results in most metrics, demonstrating their effectiveness in capturing facial similarities.

Our study demonstrated that face embeddings, although capturing phenotypic similarities, have limitations in accurately representing complex evolutionary relationships. Future research could explore the integration of additional genetic, linguistic, and geographic information to improve the accuracy and robustness of phylogenetic tree inference in human populations.

References

Alshehhi

Almarzooqi

Alhammadi

Werghi

Tay

G.K.

Alsafar

, Advancement in human face prediction using DNA, Genes14(1) (2023), 136.

Azer

E.S.

Ebrahimabadi

M.H.

Malikić

Khardon

Sahinalp

S.C.

, Tumor phylogeny topology inference via deep learning, iScience23(11) (2020), 101655.

Bhattacharjee

Bayzid

M.S.

, Machine learning based imputationtechniques for estimating phylogenetic trees from incompletedistance matrices, BMC genomics21 (2020), 1–14.

Bogdanowicz

Giaro

, Matching split distance for unrootedbinary phylogenetic trees, IEEE/ACM Transactions onComputational Biology and Bioinformatics9(1) (2012), 150–160. doi: 10.1109/TCBB.2011.48.

Chen

, Y-LineageTracker: ahigh-throughput analysis framework for Y-chromosomal next-generationsequencing data, BMC Bioinformatics22(1) (2021), 1–15.

Chollet

, Xception: Deep Learning with Depthwise Separable Convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1251–1258, 2017.

Das

Dantcheva

Bremond

, Mitigating Bias in Gender, Age and Ethnicity Classification: a Multi-Task Convolution Neural Network Approach. In Proceedings of the European conference on computer vision (eccv) workshops, pages 0–0, 2018.

Duda

Zrzavy

, Human population history revealed by asupertree approach, Scientific Reports6(1) (2016), 1–10.

Geitgey

, Face Recognition, https://github.com/ageitgey/face_recognition, 2022.

10.

Zhang

Ren

Sun

, Deep Residual Learning for Image Recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016.

11.

Hopman

S.M.

Merks

J.H.

Suttie

Hennekam

Hammond

, Faceshape differs in phylogenetically related populations, EuropeanJournal of Human Genetics22(11) (2014), 1268–1271.

12.

Hunt

Pedersen

K.S.

, Rove-Tree-11: The not-so-Wild Rover, A hierarchically structured image dataset for deep metric learning research. In Proceedings of the Asian Conference on Computer Vision, pages 2967–2983, 2022.

13.

Jiang

Balaban

Zhu

Mirarab

, DEPP: Deep LearningEnables Extending Species Trees using Single Genes, SystematicBiology72(1) (2023), 17–34.

14.

Kiel

, Assessing bivalve phylogeny using Deep Learning and Computer Vision approaches, bioRxiv, pages 2021–04, 2021.

15.

King

D.E.

, Dlib-ml: A machine learning toolkit,-, The Journal ofMachine Learning Research10 (1758), 1755–1758, (2009).

16.

Kumar

Sharma

, Evolutionary sparse learning forphylogenomics, Molecular Biology and Evolution38(11) (2021), 4674–4682.

17.

Kushniarevich

Utevska

Chuhryaeva

Agdzhoyan

Dibirova

Uktveryte

Möls

Mulahasanovic

Pshenichnov

Frolova

, et al. Genetic heritage of thebalto-slavic speaking populations: a synthesis of autosomal,mitochondrial and y-chromosomal data, PloS One10(9) (2015), e0135820.

18.

Lin

Rajan

Moret

B.M.E.

, A metric for phylogenetic treesbased on matching, IEEE/ACM Transactions on ComputationalBiology and Bioinformatics4(9) (2012), 1014–1022. doi: 10.1109/TCBB.2011.157.

19.

Llamas

Fehren-Schmitz

Valverde

Soubrier

Mallick

Rohland

Nordenfelt

Valdiosera,

Richards

S.M.

Rohrlach

, et al. Ancient mitochondrial DNA provideshigh-resolution time scale of the peopling of the Americas, Science Advances2(4) (2016), e1501385.

20.

Luján-García

J.E.

Yáñnez-Márquez

Villuendas-Rey

Camacho-Nieto

, A transfer learning methodfor pneumonia classification and visualization, AppliedSciences10(8) (2020), 2908.

21.

Masood

Gupta

Wajid

Gupta

Ahmed

, Prediction of Human Ethnicity from Facial Images Using Neural Networks. In Data Engineering and Intelligent Computing: Proceedings of IC3T 2016, pages 217–226. Springer, 2018.

22.

Mohammed

, ResAttUNet: Detecting Marine Debris using an Attention activated Residual UNet, arXiv preprint arXiv:2210.08506, 2022.

23.

Naranjo-Torres

Mora

Hernández-García

Barrientos

R.J.

Fredes

Valenzuela

, A review of convolutionalneural network applied to fruit image processing, AppliedSciences10(10) (2020), 3443.

24.

Nelson

, Cladistic analysis and synthesis: Principles anddefinitions, with a historical note on adanson’s familles desplantes -, Systematic Biology28(1) (1979), 1–21. doi: 10.1093/sysbio/28.1.1.

25.

Nye

T.M.W.

Liò

Gilks

W.R.

, A novel algorithm andweb-based tool for comparing two alternative phylogenetic trees, Bioinformatics22(1) (2006), 117–119. doi: 10.1093/bioinformatics/bti720.

26.

Parkhi

O.M.

Vedaldi

Zisserman

, Deep face recognition, 2015.

27.

Pedregosa

Varoquaux

Gramfort

Michel

Thirion

Grisel

Blondel

Prettenhofer

Weiss

Dubourg

Vanderplas

Passos

Cournapeau

Brucher

Perrot

Duchesnay

, Scikit-learn: Machine learning in Python, Journal of Machine Learning Research12 (2011), 2825–2830.

28.

Reich

Green

R.E.

Kircher

Krause

Patterson

Durand

E.Y.

Viola

Briggs

A.W.

Stenzel

Johnson

P.L.

, et al. Genetic history of an archaic hominin group from Denisova Cave inSiberia, Nature468(7327) (2010), 1053–1060.

29.

Robinson

D.F.

Foulds

L.R.

, Comparison of phylogenetic trees, Mathematical Biosciences53(1-2) (1981), 131–147.

30.

Saitou

Nei

, The neighbor-joining method: a new method forreconstructing phylogenetic trees, Molecular Biology andEvolution4(4) (1987), 406–425.

31.

Sapoval

Aghazadeh

Nute

M.G.

Antunes

D.A.

Balaji.

Baraniuk

Barberan

Dannenfelser

Dun

Edrisi

, et al. Current progress and open challenges for applying deeplearning across the biosciences, Nature Communications13(1) (2022), 1728.

32.

Schonfeld

Ebrahimi

Sinha

Darrell

Akata

, Generalized Zero- and Few-Shot Learning via AlignedVariational Autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8247–8255, 2019.

33.

Simonyan

Zisserman

, Very Deep Convolutional Networks for Large-Scale Image Recognition, arXiv preprint arXiv:1409.1556, 2014.

34.

Smith

, Information theoretic generalized robinson-foulds metricsfor comparing phylogenetic trees, Bioinformatics36(2020), 5007–5013.

35.

Sokal

R.R.

Michener

C.D.

, A statistical method for evaluating systematic relationships, Multivariate Statistical Methods, Among-groups Covariation (1975), 269.

36.

Srinivas

Atwal

Rose

D.C.

Mahalingam

Ricanek

Bolme

D.S.

, Age, Gender, and Fine-Grained Ethnicity Prediction using Convolutional Neural Networks for the East Asian Face Dataset. In 2017 12th IEEE International Conference on Automatic Face % Gesture Recognition (FG 2017), pages 953–960. IEEE, 2017.

37.

Suvorov

Hochuli

Schrider

D.R.

, Accurate inference of treetopologies from multiple sequence alignments using deep learning, Systematic Biology69(2) (2020), 221–233.

38.

Nguyen

, Race recognition using deepconvolutional neural networks, Symmetry10(11) (2018), 564.

39.

Wolf

Müller

Dandekar

Pollack

J.D.

, Phylogeny ofFirmicutes with special reference to Mycoplasma (Mollicutes) asinferred from phosphoglycerate kinase amino acid sequence data, International Journal of Systematic and Evolutionary Microbiology54(3) (2004), 871–875.

40.

Yuan

Wang

Gao

Cai

, Race Classification from Face using Deep Convolutional Neural Networks. In 2018 3rd International Conference on Advanced Robotics and Mechatronics (ICARM), pages 1–6. IEEE, 2018.

41.

Zaharias

Grosshauser

Warnow

, Re-evaluating deep neural networks for phylogeny estimation: the issue of taxon sampling, Journal of Computational Biology29(1) (2022), 74–89.

42.

Zheng

Lin

Cohen

Shechtman

Barnes

Zhang

Amirghodsi

Luo

, CM-GAN: Image Inpainting with Cascaded Modulation GAN and Object-Aware Training, arXiv preprint arXiv:2203.11947, 2022.

43.

Zou

Zhang

Guan

Zhang

, Deep residual neuralnetworks resolve quartet molecular phylogenies, MolecularBiology and Evolution37(5) (2020), 1495–1507.

Can convolutional neural networks approximate the phylogenetic tree of human populations with face images? 1

Abstract

Keywords

1 Introduction

2 Related work

3 Methodology

3.1 Ground truth

3.2 Data acquisition

3.3 Visual encoders

4.1 Subtask A

Table 1 Measured metrics for the subtask A Model F1-macro F1-micro F1-weighted Accuracy VGG-16 0.20282 0.24460 0.21250 0.24460 ResNet50 0.19400 0.25899 0.21840 0.25899 Face Recognition 0.29043 0.35971 0.32073 0.35971

References

Table 1
Measured metrics for the subtask A

Model F1-macro F1-micro F1-weighted Accuracy

VGG-16 0.20282 0.24460 0.21250 0.24460

ResNet50 0.19400 0.25899 0.21840 0.25899

Face Recognition 0.29043 0.35971 0.32073 0.35971