Automated identification and deep classification of cut marks on bones and its paleoanthropological implications
Introduction
The emergence of meat-eating is one of the most highly debated issues in human evolution. It has been linked to encephalization, stone tool use and other major changes in the structural behavior of early hominins (see review in [1]). Recent discoveries argue that both stone tools and meat-eating could have potentially occurred more than one million years before the earliest evidence of encephalization [2], [3]. If true, this will constitute a paradigm shift on our current interpretations on how these behaviors emerged during human evolution. This emphasizes the importance of correctly identifying the bone surface modifications that support such an early meat-eating behavior. The correct interpretation of these modifications, and more specifically of cut marks, is also of great relevance for understanding other important early hominin behavioral features, including early archaeological site function, meat used as a regular or a fallback food, and hunting or scavenging as methods of carcass acquisition. Inference of other social components could be achieved through the analysis of the butchering process (e.g., carcass disarticulation patterns). In fact, the use of cut marks in the fossil record has enabled the identification of specific butchering behaviors by early Pleistocene hominins and their carcass acquisition strategies in the earliest stages of evolution of Homo (see review in [4]).
The uncertainty surrounding the spatial association of stone tools and faunal remains during palimpsest formation, dictates that only anthropogenic bone surface modifications can be used effectively to link functionally lithics and bones. Hence, there is great need to accurately identify cut marks and to distinguish them from modifications created by other non-hominin agents, such as trampling or sediment abrasion. Claims have been made about the Pliocene antiquity of carcass processing using purported cut-marked bones in Dikika (Ethiopia) [2] or in the Plio-Pleistocene boundary in Quranwala (India) [5]. The claims made based on these discoveries remain controversial [6], as the purported cut-marked fossils from both sites also bear clear evidence of trampling or sedimentary abrasion [7].
The importance of a correct identification of bone surface modifications is not restricted to the earliest periods in human evolution. Recent claims have also been made for a 30,000-year-old presence of humans in the American continent based on the presence of cut-marked bones in old fossiliferous deposits (e.g., [8], [9], [10], [11]). This contradicts current evidence of the earliest presence of humans in the continent at a much later age. Similar to the findings in Dikika and Quranwala, the Arroyo del Vizcaino site (Uruguay) also contains abundant conspicuous evidence of trampling and/or sedimentary abrasion [9], [10], [11].
Taphonomists broadly disagree on how cut marks could be properly identified. A series of microscopic criteria were experimentally defined [12], [13], [14], but there is substantial disagreement on how these criteria should be interpreted by individual researchers [13], [15]. Recently, it has been documented that the interpretation of some of these variables remains subjective, hampering an objective scientific approach to cut mark identification and replication [44]. The lack of objective methods underscores the fact that the present identification of BSM (namely, cut marks) depends strongly on the subjective assessment and knowledge of each researcher [44]. Cut mark identification is, thus, commonly carried out outside replicable hypothesis-testing frameworks.
In order to introduce objectivity in cut mark identification, we introduce automatic image classification by machine learning algorithms. Algorithms such as Convolutional Neural Networks (CNN) [16], [17], [18], [19], [20], [21] and Support Vector Machine (SVM) [22], [23], [24], have been shown to match and even exceed human performance in image and pattern recognition. Their use for image processing and classification is greatly facilitated by access to open source software packages such as Neuroph of TensorFlow [25], [26], [27].
In the present study the sample size (79 marks) was kept intentionally short in order to maximize human expert scores (the larger the sample the higher the identification failure rates by humans) and to minimize the computer's accuracy (larger sample sizes lead to better training and higher identification rates) [22], [23], [24], [25], [26], [27]. Despite this initial advantage for human experts, we find that computer algorithms are more successful in the task of mark identification. The proposed CNN and SVM methods enable taphonomists to perform BSM identification in a more objective way than has ever been possible.
We present results that demonstrate the superiority of machine learning algorithms in identifying BSM over “subjective” assessment by several human experts. The present study, using modern bones, suggests that cut marks, like other taphonomic entities, are subject to morphological evolution, through a palimpsestic multiple-agent processes and an interplay between stasis and change [28], [29], [30]. This approach may enable the objective resolution of many cut mark-related controversies, including whether hominin butchering behaviors are identifiable in Pliocene fossils and whether the Americas were occupied by humans more than 30,000 years ago, given the presence of bones bearing BSM which could be interpreted as purported cut marks.
Section snippets
Methodological description of the sample
A selection of 79 experimental marks was used as the training set (SI). These were composed of 42 trampling marks and 37 cut marks. Trampling marks were created by using four types of sediments: fine-grained (0.06–0.2 mm), medium-grained (0.2–0.6 mm) and coarse-grained (0.6–2.0 mm) sand, as well as a combination of the previous sand types over a clay substratum, and granular gravel (>2.0 mm). These marks were selected from the trampling experiment reported by Domínguez-Rodrigo et al. [13] and
Results
For evaluation, 20 experimental marks are tested to compare the performance of CNNs, SVMs, and human experts.
The best CNNs architecture identified marks correctly with a 91% mean accuracy (sd = 5.3). Note that cut marks and trampling marks have balanced classification accuracy unlike SVM and human experts: 90% mean accuracy (sd = 10.6) for cut marks and 90% mean accuracy (sd = 8.0) for trampling marks.
The best accuracy obtained by SVMs is 81.5% (sd = 7.5) using the grid method to select points
Discussion
A few years ago, a multivariate analysis of trampling and cut marks using a set of microscopic variables was argued to yield an accuracy of >80% in the identification of marks [13]. However, this high-accuracy classification rate was automatically obtained by a discriminant analysis. It did not test the performance of the human analyst. The study assumed that humans could objectively interpret the same categories in each variable and the discriminant analysis would automatically classify each
Conclusions
We have implemented convolutional deep neural networks to identify cut and trampling marks on bones. The algorithms exhibit an accuracy that is almost 50% better than those produced by experienced taphonomists trained on BSM. The data and analysis presented here are introductory to the potential of machine learning algorithms in taphonomic research. The present methods are readily extensible to hundreds or thousands of images suggesting that taphonomic research can be dramatically improved by
Acknowledgements
This collaborative work was carried out with support from a Research Salvador Madariaga grant to MDR (Ministry of Education, Culture and Sport, Spain. Ref PRX16/00010). WB, GA and PK acknowledge support by the ERC Advanced Investigator award No 341117 (FMCoBe). MDR thanks D. Lieberman and the Human Evolutionary Biology Department at Harvard and the Royal Complutense College at Harvard, where this research was conducted. We also thank Lucía Cobo and Julia Aramendi for their help during the
Wonmin Byeon is a researcher at NVIDIA Research in Santa Clara, US. Before joining NVIDIA, in 2017, she was a post-doctoral researcher at ETH Zurich and IDSIA, Switzerland, working with Juergen Schmidhuber and Petros Koumoutsakos. She received her PhD in Computer Science from Technical University Kaiserslautern, Germany in 2016. Her research interests are in the fields of deep learning and computer vision, especially with Multi-dimensional Long Short-Term Memory recurrent neural network
References (44)
- et al.
Intentional cut marks on bovid from the Quranwala zone, 2.6 Ma, Siwalik Frontal Range, northwestern India
C. R. Palevol
(2016) - et al.
A new protocol to differentiate trampling marks from butchery cut marks
J. Archaeol. Sci.
(2009) - et al.
Taphonomic identification of cut marks made with lithic handaxes: an experimental study
J. Archaeol. Sci.
(2010) - et al.
Size matters. An evaluation of descriptive criteria for identifying cut marks made by unmodified rocks during butchery
J. Archaeol. Sci.
(2014) - et al.
Image processing with neural networks—a review
Pattern Recognit.
(2002) - et al.
Experimental study of cut marks made with rocks unmodified by human flaking and its bearing on claims of 3.4-million-year-old butchery evidence from Dikika, Ethiopia
J. Archaeol. Sci.
(2012) - et al.
TensorFlow: biology's gateway to deep learning?
Cell Syst.
(2016) - et al.
Speeded-up robust features (SURF)
Comput. Vis. Image Underst.
(2008) - et al.
Use and abuse of cut mark analyses: the Rorschach effect
J. Archaeol Sci.
(2017) - et al.
The meat of the matter: an evolutionary perspective on human carnivory
Azania: Archaeol. Res. Afr.
(2017)
Evidence for stone-tool-assisted consumption of animal tissues before 3.39 million years ago at Dikika, Ethiopia
Nature
3.3-million-year-old stone tools from Lomekwi 3, West Turkana, Kenya
Nature
Deconstructing Olduvai: A Taphonomic Study of the Bed I Sites
Configurational approach to identifying the earliest hominin butchers
Proc. Natl. Acad. Sci. USA
3.3-Million-year-old stone tools and butchery traces? More evidence needed
PaleoAnthropology
Megafauna. Giant Beasts of Pleistocene South America
Among others, cut-marks are archaeological evidence: Reply to “Archaeological evidences are still missing: Comment on Fariña et al. Arroyo del Vizcaíno Site, Uruguay” by Suárez et al
Proc. R. Soc. B
Arroyo del Vizcaíno, Uruguay: a fossil-rich 30-ka-old megafaunal locality with cut-marked bones
Proc. R. Soc. B
Bone surface modifications, reasonable certainty and human antiquity in the Americas: the case of the arroyo del Vizcaíno site
Am. Antiq.
Bone surface modifications in zooarchaeology
J Archaeol Method Theory
Facial expression analysis using shape and motion information extracted by convolutional neural networks
A face detection system using shunting inhibitory convolutional neural networks
Cited by (39)
High-accuracy in the classification of butchery cut marks and crocodile tooth marks using machine learning methods and computer vision algorithms
2022, GeobiosCitation Excerpt :This supervised computer vision (CV) method is even more effective than human experts. A pioneer application of this method to a limited set of BSM showed that the machine could classify >50% better than human experts (Byeon et al., 2019). Subsequent applications to various taphonomic problems showed the enormous potential of DL methods for implementing objective approaches to BSM identification in modern experiments and in the archaeological record.
A deep variational convolutional Autoencoder for unsupervised features extraction of ceramic profiles. A case study from central Italy
2022, Journal of Archaeological ScienceCitation Excerpt :Currently, the use of ML in archaeology is almost entirely limited to a supervised approach for classification. It involves geospatial analysis (for a comprehensive overview see Bickler 2021, 188), text recognition (Sanders 2018; Barucci et al., 2021; Srivatsan et al., 2021), the analysis of bone cutmarks (Byeon et al., 2019), and especially the classification of artefacts including buildings (Caspari and Crespo, 2019; Pavan Kumar et al., 2022), decorations (Chetouani et al., 2020; Felicetti et al., 2021; Romanengo et al., 2020), pottery (Lucena et al., 2016; Ostertag and Beurton-Aimar, 2020; Gualandi et al., 2021), and lithics (Nash and Prewitt, 2016; Elliot et al., 2021; Navarro et al., 2021). Even if clustering and dimensionality reduction algorithms are widely used in archaeology, unsupervised approaches based on tunable models, are rarely used (Hermoza and Sipiran, 2018; Bogacz et al., 2020; Hu et al., 2020; Parisotto et al. 2020, 2022).
Wonmin Byeon is a researcher at NVIDIA Research in Santa Clara, US. Before joining NVIDIA, in 2017, she was a post-doctoral researcher at ETH Zurich and IDSIA, Switzerland, working with Juergen Schmidhuber and Petros Koumoutsakos. She received her PhD in Computer Science from Technical University Kaiserslautern, Germany in 2016. Her research interests are in the fields of deep learning and computer vision, especially with Multi-dimensional Long Short-Term Memory recurrent neural network (MD-LSTM) and Convolutional Neural Network (CNN) for high-dimensional data understanding. Email: [email protected]
Manuel Domínguez-Rodrigo is co-director of the IDEA (Institute of Evolution in Africa) and professor of the Department of Prehistory, Ancient History and Archeology of the Complutense University. He has been co-director of the paleoanthropological projects of Peninj (Lago Natron) (1995–2005), Eyasi (2002–2006) and, currently, of the Olduvai Gorge (2006–present). He has published 8 books and more than 200 impact articles. He has been guest professor and researcher at the Universities of Harvard, Rutgers and St. Louis and the Royal Complutense College in Harvard (USA). His specialties are taphonomy and paleoanthropology. He is pioneering the application of high computing tools, such as algorithms of “machine learning” and “deep learning” or “computer vision” to the world of paleoanthropology. He is currently co-director of TOPPP (www.olduvaiproject.org). Email: [email protected].
Georgios Arampatzis works since January 2015 as a postdoctoral researcher at the Computational Science & Engineering Laboratory at ETH Zurich and since 2018 at Collegium Helveticum. He received his Bachelor, Master and PhD degrees from the Mathematics and Applied Mathematics Department at the University of Crete in 2006, 2011 and 2014, respectively. During 2014–2015 he worked as a postdoctoral researcher at the Mathematics and Statistics Department at the University of Massachusetts, Amherst. His research interests include numerical computations, Monte Carlo methods, spatio-temporal kinetic Monte Carlo, sensitivity analysis for stochastic processes, uncertainty quantification algorithms with applications in molecular dynamics, fluid dynamics and pharmacodynamics and differential privacy.
Enrique Baquedano is co-director of the IDEA (Institute of Evolution in Africa) and director of the Regional Archaeological Museum. He is also co-director of TOPPP (www.olduvaiproject.org) and the excavations of the Neanderthal site of Pinilla del Valle (Madrid). His doctoral research work focused on taphonomic and historiographic topics. He is a visiting professor at the University of Alcalá de Henares. He has published several impact articles and is an expert in disseminating heritage through a large number of exhibitions. He is the architect of the permanent exhibitions of the Regional Archaeological Museum of Madrid and the museums of Olduvai and the National Museum of Tanzania in Dar es Salaam (Tanzania).
José Yravedra is a professor at the UCM and a member of TOPPP since 2008. He is also the director of several research projects, such as: “The evolution of human behavior a million years ago in East Africa, reviewing the evidence of beds III and IV of the Olduvai Gorge. The exploitation of megafaunas in the African Lower Paleolithic, new perspectives from BK (Olduvai Gorge)”. He has also directed other projects in the Iberian Peninsula and has been involved in more than 40 research projects in African and Iberian sites. On the other hand, he has more than 300 publications distributed in books such as “Taphonomy applied to Zooarchaeology”, monographs and scientific articles published in international journals such as Nature Communications, Quaternary Science Reviews, Boreas, Journal of Human Evolution, Journal of Archaeological Science, Quaterary International etc. His lines of research are mainly Paleolithic, Taphonomy and Zooarchaeology. He is currently working very actively in taphonomy of carnivores with specialization on canids and felines, and in the creation of new taphonomic documentation techniques through the development of microtaphonomy, being a pioneer in the application of photogrammetry and geometric morphometry.
Miguel Ángel Maté-González is a PhD in Geotechnologies applied to Construction, Energy and Industry (USAL-2017). He is currently a teaching and research staff at the University of Salamanca. He has solved difficulties related to the use of geomatics sensors, representation techniques, 3D display, sensor calibration and has generated a new methodology for the analysis of 3D cut marks on bones. That methodology has been used in different archaeological and palaeontological projects and numerous scientific papers and has been presented at various international congresses.
Petros Koumoutsakos holds the Chair for Computational Science at ETH Zurich and serves as Fellow of the Collegium Helveticum. Petros is elected Fellow of the American Society of Mechanical Engineers (ASME), the American Physical Society (APS), the Society of Industrial and Applied Mathematics (SIAM). He has held visiting fellow positions at Caltech, the University of Tokyo, MIT and the Radcliffe Institute of Advanced Study at Harvard University. He is recipient of the Advanced Investigator Award by the European Research Council and the ACM Gordon Bell prize in Supercomputing. He is elected Foreign Member to the US National Academy of Engineering (NAE). His research interests are on the fundamentals and applications of computing and data science to understand, predict and optimize fluid flows in engineering, nanotechnology, and medicine.
- 1
These authors contributed equally.