This article outlines the aims, processes and initial results of the Arch-I-Scan Project. The project endeavours to contribute to improved understandings of Roman foodways by using artificial intelligence to enhance the collection of ceramic artefact data. Through enabling the recording of larger and, crucially, more comprehensively documented datasets, we aim to enable more in-depth and wider ranging studies of past eating and drinking habits and their variations across the Roman world through material-cultural remains. With this aim in mind, this paper begins with an overview of the importance of approaches to material culture, and particularly ceramics, for improving our understandings of cultural patterns in foodways, and particularly food-consumption practices, in the past. We then outline the Arch-I-Scan Project’s rationale and our planned approaches to harnessing the potential of AI for artefact recording, and specifically the recording of Roman ceramic tablewares, specifically terra sigillata. This is followed by an outline of the project’s processes to date and changes we have made to these processes and to the direction of the project’s technical efforts to mitigate the impact of the Covid pandemic. We then present our results to date, contextualised within the overall aims of the project. We also briefly outline how this particular aspect of our research can contribute to its wider application.
Tell me what you eat and I will tell you who you are
Brillat-Savarin, Physiologie du Goût (1825)
Though it is so often quoted that it has become commonplace, Brillat-Savarin’s words do strike at the heart of food’s centrality to humans’ very nature. This goes far beyond the biological necessity of fuelling our bodies. Food (including drink) ‘plays a central role in our societies, and provides us as much with intricate symbols and metaphors as with nutritional substance’ (MacClancy 1992: 5). These symbols and metaphors play a major role in social group identification and people’s sense of belonging to such groups. The centrality of foodways to culture has been stressed by leading anthropologists (e.g. Bourdieu 1992: 250–2; Douglas 1984; Sahlins 1976; Wolf 1982). The ways in which food is consumed have enormous potential impact on the day-to-day experience of humans and therefore on the habitus underlying their experienced cultural differences.
Thus, foodways, with their social codes and symbols, are good indicators of cultural differences, relevant to past people’s group identifications. Culinary differences between self-identifying groups can be expressed by the choice of particular ingredients in preparing food. More often, though, in situations where there exists no significant differential access to crucial ingredients, culinary differences are expressed in what is done with the ingredients to transform them into a proper meal (Dietler 2010: 205–213). Ways of preparing food, but also ways of serving it, or choices concerning what food is consumed when, how and with whom are often employed to mark boundaries between groups, be they religious (e.g. Freidenreich 2011), or cultural, social, or class groups (e.g. Goody 1982; Jurafsky et al. 2016). That is, it is often what is done with food, rather than the ingredients themselves, that is important for demarcating culinary boundaries.
While archaeology often has limited access to actual ingredients used and food prepared in the past, it does have access to the material culture which past people used in preparing and consuming their food and drink. The material culture used for food consumption is informed by and helps to shape social practices (Bourdieu 1984; Appadurai 1986; see also Dietler 2010). In particular, ceramics remains, the bread and butter of much archaeology, hold important information (e.g. Woolf 1998: 181–205; for discussion see also: Allison 2015; 2016; 2018; Allison and Sterry 2015).
Despite this, much study of ceramics and socio-cultural practices around foodways in Roman archaeology has focused on investigating the distribution of food products and the Roman economy (i.e. the study of amphorae: e.g. Teichner et al. 2008; González Cesteros and Berni Millet 2018) or on investigating different food-preparation practices (i.e. the study of cooking wares: e.g. Swan 1999; Fentress 2010). When it comes to investigating food-consumption practices throughout the Roman world there has often been a tendency to assume that the ancient textual sources provide the blueprint for some sense of conformity across that world (see Allison 2016: 153; 2017: 187). Roman archaeologists have traditionally shied away from using ceramics for detailed investigations of food consumption despite the intuitive potential they have. Ceramic finewares, typically employed in food consumption, have instead been used to answer questions related to their production and, again, to trade and the Roman economy (e.g. Mees 2007; 2016; Polak 2017; Willet 2014), or merely for dating sites. The fact that ceramic tablewares are found in their millions across the Roman world makes them particularly valuable artefacts in this regard, notably the ubiquitous wheel-made and mass-produced, red-polished terra sigillata (see Mees and Polak 2013). This focus notwithstanding, consumption is ontologically prior to production and trade, certainly at the scale witnessed around the Roman world (Dietler 2010: 205). That is, producers are likely to first secure, or cultivate, a market for their goods before taking the risk of (large-scale) production. In this sense, the existence of large-scale production and exchange of finewares like terra sigillata should have prompted more intensive investigation into the consumption end of these fine ceramics’ life course (see van Oyen 2016).
Indeed, these millions of Roman ceramic tablewares found in different parts of, and in different socio-cultural contexts in, the Roman world do not necessarily symbolise homogeneity of practice across this world but potentially document diverse eating and drinking practices in different regions, sites and socio-economic contexts. As has been noted (e.g. Bes and Poblome 2007), though, until quite recently (e.g. Cool 2006; Pitts 2014; 2015), despite their great potential to answer more socially oriented questions (van Oyen 2015), there has been a lack of vision in traditional scholarship on Roman ceramic tablewares, and questions concerning their end use have been overlooked.
As part of a recent shift in attention towards such end-use questions, the ‘Big Data on the Roman Table’ (BDRT) research network (2015–16), involving archaeologists, heritage professionals, computer scientists and mathematicians, set out to explore new ways of harnessing and analysing the extensive datasets of ceramic finewares, notably tablewares, from around the Roman world for consumption-orientated approaches to this rich archaeological resource (Allison, Pitts and Colley 2018). Among the issues discussed by this network were the poor quality of much of the data, often ‘legacy data’ (see Allison 2008), and also the inconsistencies and often incompleteness of current recording approaches that prohibit intra- and inter-site analyses to investigate diversity of cultural practices among foodways across this world.
The perceived poor quality of archaeological data – i.e. lacking contextual information for end use or form identification – has to some extent been used to justify a focus on vessel fabric to investigate production and distribution, which is a limited exploitation of the potential of this ostensibly rich resource. Answering questions concerning end use usually requires more detailed recording of a number of characteristics of each artefact (e.g. residue evidence and usewear patterns), of their precise findspots, and of the deposition of their assemblages. Moreover, ceramic remains are often found in re-deposited contexts. Such secondary deposition, while still allowing the traditional production- and trade-oriented approaches, is often used as justification for not exploring questions related to end use. It has been judged more expedient to pursue a scholarly vision that focuses on dating and trade patterns requiring only a sample of the remains to be collected.
This leads to a further rationale for not paying as much attention to consumption-oriented approaches – that related to cost. Identifying and recording ceramic remains – i.e. classifying the fabric and often the form of individual sherds – is a repetitive and time-consuming task but one that requires a significant amount of expertise, meaning that a conceptually simple task is expensive in terms of specialist time, and therefore in cost to the project (see Anichini et al. 2020, Section 1.1). The sizes of many of the ceramic assemblages, not to mention the vastness of all such assemblages across the Roman world, make this task prohibitively expensive. Because a study of ceramics’ end use requires more comprehensive recording of all ceramic material, in greater detail, it is even more time consuming and more costly than an approach that requires only a sample of this material. The arguments of poor data quality and cost effectiveness mutually reinforce each other, with cost effectiveness leading to incomplete datasets and sampling strategies that drive further economising through exclusively pursuing questions based on samples and ensuring that comprehensive study of the end use of the material is not viable.
Sampling strategies are often developed to make the data handling processes more efficient and targeted for the particular questions to be addressed. These strategies, discussed most explicitly in literature concerning survey (e.g. Alcock and Cherry 2004) but also applicable to excavation, range from collecting only proportions of the ceramic remains, to recording only limited characteristics of each sherd, to recording only selected ‘diagnostic’ sherds, to a combination of all these strategies, and often to further selective approaches to the publication of the resulting data. While the reasons behind such strategies are understandable (see Timby 2020: 530), such selective processes lead to limited subsets of the archaeological record being available for any further analyses (cf. Fentress 2000; Poblome et al. 2013: 152–154; Alcock and Cherry 2004; Whitelaw 2012). Different vessel forms have generally only been employed to trace their typological development to assist with site dating (e.g. Ettlinger et al. 1990; Oxé et al. 2000), rather than to investigate how the various forms would have been used. Obviously, while such sampling measures might be appropriate for answering the questions of traditional ceramic research, they entail the loss of the types of information required for answering questions related to end use and socio-cultural practices.
For such reasons, consumption-oriented studies of Roman ceramics have to date focused on relatively small assemblages from particular excavations in settlement sites (e.g. Allison 2009: 19–25; 2016: 175–179; 2020; Cooper, Johnson and Sterry 2018; Luley 2018), or from excavations of similarly small grave assemblages that do not necessarily document lived practices (e.g. Biddulph 2018; Dananai and Deru 2018). Such studies are limited in their usefulness for broader comparative consumption approaches (e.g. inter-community) to foodways. If assemblages in such small-scale studies could be recorded so that they could be used for more comprehensive comparisons for larger scale inter-site analyses, however, we might be able to use them for wider applications to identify varying cultural practices among a range of sites. Recently, to compensate for current limited and inconsistent recording, but to still carry out large-scale investigations, some studies have taken (or proposed) fairly broad-brush approaches. For example, the functional categories proposed by Bermejo Tirado (2018) and Van der Veen (2018) for recording ceramic data to allow consumption-oriented investigations are, of necessity, rather coarse grained. More importantly, perhaps, Marshall and Seeley (2018) analysed the morphologies and limited recorded evidence of end use (e.g. burning, usewear etc.) of large volumes of open tableware remains from Roman sites across London to investigate functional variation among these vessels. Nevertheless, their studies have been restricted by the limited and sporadic recording of such characteristics for every artefact in their dataset.
So, while potentially comparable datasets of Roman ceramic remains of sufficient size do exist, due to cost and traditional recording approaches they are, to date, rarely recorded to a sufficient level of comprehensiveness required for in-depth consumption-oriented research (Polak 2000: 1; 2016: 397), and rarely in the same consistent manner to allow analyses across a number of sites.
The Arch-I-Scan project has been developed to address some of the questions raised by the BDRT network. Its main concerns are to develop a process by which more ceramics can be recorded more comprehensively and in a more detailed and consistent manner so that they can indeed be used to enable greater understandings of the socio-cultural practices associated with the consumption of food and drink. By taking a more holistic approach that records all these artefacts, and greater detail of their characteristics and contexts, scholars can investigate, for example, the proportions of different fabrics and forms in specific assemblages, whether at the household, settlement, or regional level, to discern patterns of ceramic use within and among sites. These patterns can then be analysed for information about various eating and drinking practices throughout the Roman world and among differing socio-economic groups or cultural milieux.
The Arch-I-Scan project is focusing on the ubiquitous terra sigillata tablewares that were mass produced at different manufacturing centres across the Roman world, that have been distributed in vast quantities across that world, and that have long been an object of inquiry for investigations of production and trade, and for site dating, rather than of the end uses of these vessels. For such inquiries, a detailed typological classification system for the forms of terra sigillata vessels has been developed since the late nineteenth century (Dragendorff 1895), and many of these Dragendorff forms continue to be the basic classification system for these tablewares (see Tyers 2014). While there are identifiable differences in fabric (and to a lesser extent form) between terra sigillata produced at different centres of manufacture, the form range is relatively small and variability within identified forms is equally limited. For terra sigillata from Gaul, the main type found in Britain, where it is known as Samian ware, some fifty forms have been identified (see Webster 1996, though other rarer variations exist: see e.g. Bird 1993).
Such a limited and fairly regular form spectrum, as well as its distribution, makes terra sigillata ideal, not only for investigating consumption practices across a wide area, but also for automated recording. Automated recording and classification of terra sigillata remains could remove, or at least ease, some of the current limitations concerning the comprehensive and consistent recording of the sizeable datasets of these tablewares. And if comprehensive recording and dataset size are no longer mutually exclusive, consumption-oriented analyses and inter- and intra-site comparisons can become more feasible. A variety of strategies for such automated recording have been developed in archaeology, tapping into the ever-increasing power of computers (e.g. Bell and Croson 1998; Di Ludovico et al. 2013; Durham et al. 1995; Gibson 1996; Huffer and Graham 2018; Kampel and Sablatnig 2001; 2002; 2006; Kampel et al. 2006; Karasik and Smilansky 2011; Makridis and Daras 2012; Li-Ying and Ke-Gang 2010; Lucena et al. 2014; Ma et al. 2000; Piccoli et al. 2015; Smith et al. 2010; Wang et al. 2017). Many of these apply technology from the field of Artificial Intelligence (see Barceló 2004 and Baxter 2014 for overviews of archaeological applications of AI). This burgeoning field is also what the Arch-I-Scan project has turned to for its initial phase of developing a tool for the automated recording of Roman tablewares.
Thus, the Arch-I-Scan project is developing a machine-learning AI tool to classify remains of terra sigillata more efficiently, particularly in their most commonly found, fragmentary condition. The goal is an automated classifier that is trained to identify ceramic sherds from photos taken on smartphones. Taking photographs with a smartphone, even if it requires guidelines on how to take them effectively, is not a specialist task (in contrast, for instance, to 3D scanning; see e.g. Maiza and Gaildrat 2005). This can be carried out by anyone holding a phone, be they an archaeologist in the field, a student, or a member of the public. This means that larger datasets can be recorded and that specialist time can be freed up for more analytical tasks fitting the levels of skill of the pottery experts, such as investigating more complex social questions about the pottery.
Arch-I-Scan is therefore developing a tool based on convolutional neural networks with the aim that it will use the images of a particular sherd to identify the particular form, and also the size and possibly fabric, of the vessel from which the sherd originates. Convolutional neural networks are a subclass of deep learning approaches to machine learning and are conventionally used for image recognition tasks (LeCun et al. 2010). In recent decades, this field of Artificial Intelligence has made enormous advances, increasingly outperforming humans in situations where the problem can be sufficiently tightly circumscribed. That is, if the boundaries of a task are sufficiently clearly defined (such as clear and distinct categories for a classification task) and training data are available in sufficient quantity, neural nets are often superior to humans at image classification tasks, as demonstrated by the accuracy of object detection metrics in Figure 1. As they process images faster than humans, classifiers based on convolutional neural networks are certainly more efficient.
For archaeology, the anticipated results of this superior recording efficiency are that projects should be able to classify larger datasets within realistic timeframes and budgets. If archaeologists can classify and record their material more efficiently, consistently, and so more comprehensively, this would enable them to carry out types of analyses that require such recording to address questions that have until recently been considered unviable. This also means that using large datasets of Roman tablewares, recorded in relatively fine detail, to answer consumption-oriented questions, could become a more realistic ambition.
The Arch-I-Scan Project’s first requirement is a sufficiently large training dataset of images of Roman tablewares with which to train the AI tool. Through recording as many photographs of the Roman tablewares as is feasible, the project aims to produce sufficient images to develop an automated classifier. As noted above, there are some fifty main forms of terra sigillata in Britain, although with subdivisions within some of these forms, which can be used to test the quality of the classifier’s performance. That is, the form subdivisions can be used to assess what level of detailed class subdivision the classifier can produce.
While the artificial classifier is intended to significantly speed up the recording process so that Roman tableware datasets can be recorded in a manner appropriate for addressing consumption-oriented questions, and potentially more economically, the actual creation of such a classifier threatens to replicate, in a sense, the original problem of needing to record masses of images of ceramic remains. That is, the machine is data hungry and needs to be presented with a sizeable body of training material to build up the features that it uses to classify these images. Gathering a set of training data of sufficient size is in itself potentially prohibitively time consuming. If the classification task is sufficiently elaborate, the size of the training set of images required would quickly outstrip what can be realistically gathered within a practical time frame. Besides, it is not at all unthinkable that a large enough group of each and every individual form simply does not exist, as certain terra sigillata forms are very rare. Creating a sufficiently comprehensive training set as required to build an archaeological classifier is itself therefore a potential hurdle in the process.
A further hurdle lies in the fact that, archaeologically, we deal with sherds rather than complete vessels. Ceramic vessels do not break uniformly or in any predictable fashion. This means that a given vessel can break into a theoretically infinite number of differently shaped sherds. So, while the system of classes (i.e. vessel forms) into which images of artefacts need to be sorted is, in the case of Gaulish terra sigillata, relatively succinct, the variety within the set of artefacts (i.e. the sherds) and their images that need to be classified means that this task is not at all simple. The classifier needs to be able to reliably cope with the heterogeneity in these data (Itkin, Wolf and Dershowitz 2019). This means that the amount of training data needed requires an even greater number of sherds, and an even more vast number of images, than would be required for developing a training set of complete or near complete vessels.
To attempt to overcome these hurdles, and to help speed up the process of developing a large training set of images for the classifier, the project is involving students and volunteers in the photographing process, using hand-held mobile devices, specifically smartphones. The participation of students and volunteers using readily available equipment also serves to demonstrate that this recording process does not require specialists or specialist equipment. In addition, involving volunteers in the development phase is broadening the potential impact of Arch-I-Scan’s research, by introducing a wider audience to both AI and archaeological cataloguing processes. Once the classifier has been developed and trained with tens of thousands of images, we anticipate that it will be able to return a vessel classification of a particular sherd after being shown a limited number of images of this sherd, automating the identification process to a significant degree.
During the BDRT network, we experimented with an AI prototype (Tyukin et al. 2018). In addition, Penelope Allison, Fiona Seeley (Museum of London Archaeology), and Jeremy Levesley (School of Mathematics and Actuarial Sciences, University of Leicester) also discussed the potential usefulness of using the Museum of London’s (MoL) collections of complete/near complete Roman tablewares – largely terra sigillata but also London Ware and other Roman finewares – to build up a dataset of images of this pottery as a reference or training set for further datasets of sherds of these ceramic forms. This led to Arch-I-Scan’s partnerships with MoL and also with other institutions that have large Roman tableware collections of sherds – the Museum of London Archaeology Service (MOLA), the Vindolanda Trust, the University of Leicester Archaeological Services (ULAS), and the Colchester and Ipswich Museum Service (CIMS). Because many of tableware remains in these collections – both near complete and fragmentary – have already been classified to some extent by ceramic specialists, they are useful for first training then testing the automated classifier.
Thus, because of the initial plans and partnership with MoL, the Arch-I-Scan project commenced in December 2019 with the research team and student volunteers from the University of Leicester photographing the antiquarian collection of complete and near complete Roman finewares in this museum’s collection (Figure 2). As a pilot study, we used the vessels in this collection to compile our first dataset of images. These vessels had to be photographed in such a way that the information about them would be accessible to the machine. To this end, Daniël van Helden and Santos Núñez Jareño designed a specific standardised process for taking a number of images of each vessel from different angles. This ensured that the photographs taken by the different members of the team produced a set of comprehensive images of each vessel that were also fairly consistently photographed, while at the same time making the process practicable for non-specialists with simple equipment and, most important of all, rapid. Speed was of the essence, since this determined the all-important size of the training set that we could viably collect. In three days at MoL, and as our first attempt at such recording, with a team of up to six people, we took 12,395 photographs of 384 different Roman tableware vessels, including 8702 photos of 247 terra sigillata vessels. While the usefulness of this dataset of near complete vessels has been debated by the project team, it provided a useful starting point to bring together this interdisciplinary research team and familiarise them (particularly mathematicians) with Roman tablewares and with problems we were likely to encounter during the scanning of fragmentary material – for example, lighting conditions, use of scales, and using inexperienced recorders.
In January 2020, following quickly on from this initial data collection, the team commenced photographing sherds, mainly of terra sigillata, from the London excavations of MOLA. This part of the scanning programme again involved the project team and volunteers, largely from Birkbeck, University of London (Figure 3). To ensure that we had a standardised process for taking comparable images of each sherd, Daniël van Helden and Santos Núñez Jareño devised a photography protocol so that everyone would be able to take the same kind and number of photos of each sherd. The opportunity to be involved was greatly appreciated by the volunteers (see video: Colley 2021).
For three weeks, with fourteen different volunteers, we recorded images of Roman tableware sherds from the collection in the MOLA offices in London. At the end of these three weeks, however, the Covid-19 pandemic had spread to the UK so we could no longer continue to collect images of sherds in this manner to train the classifier as planned.
With the project’s planned data collection process severely hampered by the pandemic, Núñez Jareño and van Helden started to explore how an AI training regime could be developed which did not rely on huge amounts of real data (i.e. photos of pottery) being gathered. If we could not train the AI classifier with photos of real sherds or vessels, might we be able to train it with images of computer-generated vessels if these images were close enough to the real thing?
The relative uniformity of terra sigillata vessel forms makes them very suitable for computer simulation. That is, barring occasional decoration and handles, the vessel forms are clearly identified through, and well defined by, their profiles. Being mass produced, there is little variation among the profiles within each typological form, with the exception of size and sometimes the proportion of height versus diameter (see e.g. Polak 2000: 74–130; Monteil 2013), which means that with a limited number of simulated models, it is possible to cover much of the limited variation within a particular form. Simulation of Roman pottery is not wholly new. The ArchAIDE project developed a sherd classifier that was trained on simulated 3D models produced from standard printed vessel profile drawings by tracing and labelling the inside and the outside profile of the vessel separately (Anichini et al. 2020, Section 2.2.3; Banterle et al. 2017). In the use stage of the ArchAIDE system, users trace and label the sherd’s inner and outer profiles onto a digital image of the sherd (ArchAIDE Project 2022 - Shape Recognition)). Using this extra input from the user, the classifier then identifies the form from which the sherd originated.
Our aim is to train the Arch-I-Scan classifier in such a way that it does not require a specialist user to label the two profiles of the sherd. In this way we hope to minimise the amount of expertise required to use the tool. We therefore need the classifier to be able to identify such characteristics on its own. For this reason, and because the Covid pandemic has prohibited our collection of photos of real sherds, we need masses of realistic images of simulated vessels to prepare the machine for real photos of real sherds. The procedure we used is summarised below (for more details see Núñez Jareño et al. 2021a).
Because we did not know how complex it would be to simulate images of pottery sherds or how successfully we could use these simulated sherd images to train a classifier, and had only just started to collect such images, we decided to first create a classifier for whole terra sigillata vessels. Whole vessels are much less varied in their shape than sherds, so this should reduce the difficulty of this classification task. While a classifier that identifies whole vessels is not particularly useful in any archaeological sense, this whole vessel simulation experiment provides a good proxy with which to evaluate the different strategies that can inform further steps in the project’s process of using simulation to develop a sherd classifier. That is, it can potentially provide valuable information for the more complex task of training a sherd classifier using simulated data.
Because of the so-called reality gap (see Núñez Jareño et al. 2021a, section 3.5), we cannot simply train a classifier using only simulated data. We ultimately need a machine that can classify photos of actual archaeological material, taken in the real world. Therefore, we also need photos of actual archaeological vessels to prepare the classifier for these. For this, we used the dataset of photographs of complete/near complete terra sigillata vessels that we had taken in the MoL collection. From the set of vessels we photographed, we chose those Dragendorff forms that were most abundantly available (see Figure 4). In this set, only a small minority in the range of Dragendorff forms were represented by more than ten individual vessels, and only two had more than thirty individual vessels. Because of these low numbers for each particular form, we subsumed subdivisions within the Dragendorff forms under their parent classes (e.g. Dr27g under Dr27). In addition, as we ran initial analyses, it turned out that the range from Dr18 to Dr31 proved very difficult to disentangle for the machine because relatively shallow, flat-based plates of the classes Dr18, Dr18/31 and Dr31 are essentially part of a gradual spectrum without distinct cut-off points between these forms. As our simulations are based on a relatively limited number of archetypal drawings (see below), it is difficult to adequately catch the diversity within each form in the simulated vessels. This would have required many more intermediate drawings between the archetypes by which the three points on the spectrum (Dr18, Dr18/31 and Dr31) are defined and more research time than this particular pilot warranted. The machine also confused these forms with Dr15/17 (see Webster 1996, Figure 18). We therefore decided to aggregate the whole range of Dr18 to Dr31 (including their ‘R’ variants, which generally have a rouletted ring on the inside of the vessel) into a single class, Dr18, and to omit Dr15/17 from the dataset altogether. This obviously limits the utility of this classifier, as it was not trained to distinguish these classes, but since the aim of this experiment was to gauge the scale of the challenge of creating a classifier trained on simulated images, knowing about these limitations in itself is valuable information.
By this winnowing process we ended up with 5373 images of 162 different terra sigillata vessels, and nine classes (Table 1 and Figure 5). This is an extremely small sample size for an image classification task. This meant that we had to take extra care to reliably measure the classifier’s errors to assess its accuracy in classification (see Núñez Jareño et al. 2021a, esp. Sections 2.5–3.3).
|CLASS||NO. OF VESSELS|
Nevertheless, with our dataset of images of real vessels defined, we set about augmenting the numbers with images of simulated vessels. To create the 3D models, we started with the drawings from Webster (1996). We traced the profiles of the types included in the experiment. Using the Python package Matplotlib (Matplotlib 2022) and 3D modelling software Blender (Blender 2022), the resulting contours of their traced profiles were rotated 360 degrees around the central axis of the vessel to generate a 3D point mesh (see Núñez Jareño et al. 2021a, Figure 6). We simulated breaks in the resulting 3D models, similar to the ones present in our dataset of real vessels, to prepare the machine for what it would be faced with, as well as to increase variety among the simulated training sets. Finally, the models were ‘photographed’ from a variety of angles, resulting in a single image of the 3D model, mirroring the photos of the real vessels. Three such sets of images of simulated vessels were created: the first using Matplotlib, and two using Blender - one ‘plain’ and one where the surface of the model was treated to appear more like that of terra sigillata and where some simulated decoration was added to those Dragendorff forms that are decorated (see Figure 6). This resulted in the creation of three different datasets of simulated images (9,000 using Matplotlib, 12,600 using ‘plain’ Blender [blender1], and 11,700 including simulated decoration [blender2]) across the nine vessel classes used in this experiment.
We then used four commonly used, off-the-shelf neural network architectures (VGG19, Mobilenet v2, Resnet 50 v2, and Inception v3), each initialised with the so-called ImageNet starting weights (the standard initialisation procedure in image recognition tasks – for references see Núñez Jareño et al. 2021a, Section 2.2), and then trained them to create four experimental conditions each: one trained using images of real vessels only (the control group); three initially pre-trained with simulated images (each with one of the three simulated datasets) and then trained a second time with photos of real vessels (Figure 7). This allowed us to evaluate the impact of training with simulated data as well as that of the quality (i.e. verisimilitude) of the simulations on the classifier’s performance. Because we used a very small number of real photos with which to test the machine, standard accuracy measurements (see e.g. Hossin and Sulaiman 2015) were not very reliable, since small perturbations can have big impacts. To mitigate this, we ran each experimental condition twenty times with a different random configuration of photos (i.e. of different vessels within a particular class) for training, validation, and test sets. We took the average accuracy across these twenty training-validation-test ‘splits’ as the indicator of the machine’s performance. Because vessels were not all equally difficult to identify for the classifier, whether or not a specific vessel was in the test set (the set on which the accuracy metrics are based) could influence the performance indicator. By alternating between splits whether a set of photos of a specific vessel was in the training, validation, or test set, we could circumvent this impact, as it is the average across the splits that was used as the accuracy measurement. As is evident from Figure 7, accuracy of identification improved remarkably when training with simulated data, i.e. a classifier pre-trained with simulated images performed better than those with only the ImageNet starting weights. More realistic simulation seemed to improve it further (the least realistic, matplotlib, being outperformed by blender1, which was in turn outperformed by the more realistic blender2), though the results of the more detailed blender2 simulation set are within the error band of blender1, so this needs to be treated with caution. Training with good quality simulated data not only improved the accuracy, but the error bands were also narrower, meaning that the classifiers were also more stable.
Using confusion matrices, we can look at the results of the experiment in more detail. A confusion matrix shows the percentages of identifications that the classifier made, given that the actual vessel is of the type in the row header. A row in the confusion matrix (see Tables 2 and 3) therefore shows, in each column, the percentages of identifications the classifier gave to vessels of the class in the row header, both correct (on the diagonal) and incorrect (off the diagonal). If we compare the two extreme situations of the Inception v3 architecture, i.e. not pre-trained with simulated data (Table 2) and pre-trained with the blender2 dataset (Table 3), we can see that the percentages of correct identifications have increased across all the classes, though for some classes the increase is rather marginal. We can also see that almost all the incorrect identifications became less likely as we moved from no pre-training to pre-training with simulated images of high verisimilitude (i.e. from just Imagenet in Table 2 to blender2 in Table 3). If we assume that incorrect classifications are uniformly distributed across all possible incorrect classes, then a class that was wrongly selected by the classifier more than twice as often as expected can be seen as a particularly difficult class for the classifier to distinguish from the correct class: a major source of confusion. A decrease in such difficulties is an indication of a more successful classifier (see Núñez Jareño et al. 2021a, Section 2.5). Comparing Table 2 with Table 3, in the latter there are notably fewer off-diagonal cells, marked in red, that were major sources of confusion. What we can also see, though, is that not all such major sources of confusion were cleared up by pre-training with simulated images. Classes Dr35 and Dr36 remained a problem for the machine. This should not come as a surprise, as these are two very similar forms, only really differing in size and some proportions, the profile of Dr36 being that of Dr35 stretched along the horizontal axis. As the images were simulated without scales, the machine did not have this information to help with identification. Furthermore, we can speculate that the confusion between classes Dr29 and Dr37 was due to the fact that these are both decorated forms. Because of the low numbers of vessels, it could be that the machine focussed too much on the decoration in the cases where it was confused. Other instances of confusion are more difficult to interpret, as we are not able as yet to ascertain what the machine’s classifications were based on.
We reiterate that this experiment was based on a very small sample. While each of the simulated datasets was larger than the original dataset of 5373 images, the overall number of images was still very small for image classification procedures for which millions of images are regularly employed. As such, this experiment needs to be repeated at a larger scale to be substantiated. As we are not interested in a machine that classifies whole vessels into only nine classes, however, we are instead using these results to inform our designs and decisions in developing an archaeologically useful classifier: one for fragmentary material.
Through this experiment, necessitated by the Covid-19 pandemic, we have shown that simulation can be an important addition to the training of AI classifiers for archaeologically complete ceramic vessels. As discussed, though in itself a classifier of whole vessels is of limited use, this initial dataset and this application are very informative as a demonstration of the impact of simulated data for the construction of a classifier in the context of an extremely small, real, training set. Since our sample of real-world exemplars was too small to create an adequate training set for the neural network, pre-training with simulated vessels provided a viable alternative.
The comparison between different simulations in terms of classification performance is also enlightening. The quality of the simulations used to pre-train an AI image classifier had a considerable impact on its performance in classifying complete/near complete ceramic vessels. Enhancing simulations with the Blender software more than doubled the effect that simulation had on the performance of the neural net. Not only was the average accuracy of neural nets higher when pre-trained with higher quality simulations, the variance of the accuracy was also lower. Accuracy is therefore consistently better when an AI image classifier of terra sigillata vessels is pre-trained with high quality simulated examples.
The process of developing the classifier of whole ceramic vessels has indeed proved promising for avenues being explored in the project. That is, these conclusions are informing further exploration and testing of the potential of simulation in constructing tools to improve the efficiency of identifying and collating the more fragmentary terra sigillata remains. As already mentioned, the size of the datasets required to train AI classifiers is potentially humongous. Creating a dataset of photographs of terra sigillata sherds that is large enough to train a neural network which can not only identify all the different forms of terra sigillata, but is also able to deal with the immense variability that breakage introduces was always a tall order, further exacerbated by the Covid-19 pandemic. The results from our pilot study of near complete vessels suggest, however, that the problems of limited datasets can be mitigated by using simulated data to train artificial classifiers.
Therefore, the project is currently focusing on creating simulated sherds with which to train a sherd classifier. The process that we used to simulate the breaks and missing parts of the near complete vessels is being inverted to generate simulated sherds that are being used to develop new classifiers. As a parallel but intersecting track towards producing useful simulated sherds, we are also exploring ways to understand better why the machine confuses certain forms with others, as shown in the confusion matrices in Tables 2 and 3. If we want to produce and work with simulated sherds, we need to be strategic from the outset. Because the potential variability among sherds is nigh infinite, it is important to know relatively early in the process why the machine makes the mistakes it makes. If we know on which elements of an image the machine is basing an erroneous classification, we can attempt to correct this by training with images designed to represent that specific aspect as accurately as possible. To gain such information, the project is exploring approaches to explainable AI, which visualise the inner workings of the neural networks on the classified image, indicating which part of the image was diagnostic for the machine (Chen et al. 2019; Selvaraju et al. 2019; Sun et al. 2019). Such an insight into the inner workings of the machine has the potential to not only make it more explicable, and therefore more trustworthy, but might also allow the kind of feedback or tailored training discussed in this paragraph.
As this project is still in its initial phases, we still have some way to go before we can use the resulting data from the collections of our project partners to investigate Roman foodways and particularly to analyse consumption practices within and among our datasets from different sites (see Allison 2020).
The material being collated in the Arch-I-Scan project consists of a dataset that, when completed, can be used to investigate food- and drink-consumption practices in the different contexts of Colchester, Leicester, London and Vindolanda. Together with the AI tool, this can provide an excellent body of data and a recording process for wider application at other Roman sites, in Britain and beyond. Once Arch-I-Scan is sufficiently trained and can classify the terra sigillata from these collections, as yet unclassified by ceramic specialists, the resulting datasets and machine-learning tool can be made freely available for other archaeologists, academic and professional, and also other people tasked with, or interested in, recording Roman pottery, so that they can use the resulting datasets for their own analyses and the machine-learning tool to record their own collections. The dataset of images of whole vessels from the Museum of London, real and simulated, is accessible on the Arch-I-Scan Repository (Núñez Jareño et al. 2021b). In addition, the potential for the wider application of the machine-learning tool is not limited to terra sigillata, or indeed to Roman pottery, at least in principle, though significant training data will be necessary for such expansions.
This project exemplifies the benefits of interdisciplinary collaborations for improving the recording of archaeological data to facilitate answering a greater range of questions than we can currently ask of such data. With dwindling public funds, government, charitable and professional cultural heritage and archaeological organisations have inadequate resources for detailed, comprehensive classification and digital collation, management, and analyses of the full artefact datasets in their care. This project aims to set agendas for more comprehensive, accurate, and cost-effective artefact collation (from past, recent and future excavations) to build more robust, analysable datasets that can be more easily digitally recorded, managed and disseminated for more effective interpretation (see Timby 2020: 530).
It is further anticipated that the outcomes of the Arch-I-Scan project have the potential to be used by other groups interested in archaeology, such as community archaeological groups, archaeological volunteers, and students. The involvement of such groups in the project to date is testimony to their keenness to engage with digital, state-of-the-art archaeological recording processes and to develop skills in this area. And there are possibilities for further groups, such as visitors to museums and archaeological sites, to benefit from this project through the possibilities it provides for greater accessibility to information from, and interpretation of, the collections of our project partners, and further users of the Arch-I-Scan service.
The development of the simulated 3D models of terra sigillata adds a further dimension to the potential wider applications and impact of this project. As is evident from the discussion above, once the datasets produced by Arch-I-Scan are made freely available, they can also provide reference collections for teaching or training specialists. This is equally pertinent to the simulated data, and particularly that of complete/near complete terra sigillata vessels (see ArchAIDE 2019 – reference collection of simulated whole amphorae), but potentially also simulations of sherds. Where students and professional archaeologists are currently trained using paper catalogues, such a reference collection of simulated vessels can be a valuable tool. With 3D printing becoming increasingly available, one could potentially 3D print the simulations as life-size replicas as a further teaching and training aid (compare Anichini et al. 2020).
In the first instance, the Arch-I-Scan Project is very grateful to the Art and Humanities Research Council for providing the funding that has made this research possible. We also wish to give particular thanks to Santos Núñez Jareño for his invaluable research for this project and especially for the inspiration of and work on the 3D simulations. We are also grateful to our project partners, particularly to the Museum of London for access to the material used in this paper and to Roy Stephenson who set up this partnership and organised this access. We are similarly grateful to Nicola Fyfe (MoL) and Fiona Seeley (former Head of Finds and Conservation, MOLA) for their assistance with the material for our photographing programme and to the student volunteers, Alessandra Pegurri and Gabriel Florea, for helping photograph the pots. Other people to acknowledge are Prof. Jeremy Levesley, whose ideas and interest have been instrumental in setting up this interdisciplinary project, and Victoria Szafara, an important project team member who organised the logistics for our time in London.
The authors have no competing interests to declare.
Allison, PM. 2015. Everyday Eating and Drinking in Roman Domestic Contexts. In: Parr, BE, di Castro, A and Hope, C (eds.), Housing and habitat in the Mediterranean world: Cultural and Environmental Responses, 265–279. Babesch Supplements 26. Leuven: Peeters.
Allison, PM. 2016. Everyday foodways and social connections in Pompeian houses. In: Steel, L and Zinn, K (eds.), Exploring the materiality of food “stuffs”: Transformations, symbolic consumption and embodiments, 152–186. London and New York: Routledge (Taylor and Francis).
Allison, PM. 2017. Naming tablewares: using the artefactual evidence to investigate eating and drinking practices across the Roman world. In: Minchin, E and Jackson, H (eds.), Text and the material world: essays in honour of Graeme Clarke, SIMA – Studies in Mediterranean Archaeology, 186–198. Uppsala: Astrom editions.
Allison, PM. 2018. Meals and the Roman military. In: Ivleva, T, de Bruin, J and Driessen, M (eds.), Embracing the Provinces: Society and Material Culture of the Roman Frontier Regions. Essays in honour of Dr. Carol van Driel Murray, 103–110. Oxford: Oxbow Books. DOI: https://doi.org/10.2307/j.ctv13nb8qs.18
Allison, PM. 2020. ‘The ceramics from the Insula del Menandro and the significance of their distribution’. In: Ossana, M and Toniolo, L (eds.), ‘Fecisti cretaria’: Dal frammento al contesto: studi sul vasellame del territorio vesuviano, 199–209. Studi e richerche del parco archeologico di Pompei 40: Rome: L’erma di Bretschneider.
Allison, PM, Pitts, M and Colley, S (eds.). 2018. Big Data on the Roman Table: New approachts to tablewares in the Roman world. Internet Archaeology, 50. https://intarch.ac.uk/journal/issue50/index.html.
Allison, PM and Sterry, M. 2015. “Family” meals? Who ate with whom and where in Roman military bases? In: Vagalinski, L (ed.), Proceedings of the XXII Limes Congress, Ruse, Bulgaria Sept. 2012, 487–494. Sofia: National Archaeological Institute with Museum – Bulgarian Academy of Sciences.
Anichini, F, et al. 2020. Developing the ArchAIDE Application: A digital workflow for identifying, organising and sharing archaeological pottery using automated image recognition. Internet Archaeology, 52. DOI: https://doi.org/10.11141/ia.52.7
Appadurai, A (ed.). 1986. The Social Life of Things: Commodities in Cultural Perspective. Cambridge: Cambridge University Press. DOI: https://doi.org/10.1017/CBO9780511819582
ArchAIDE. 2019. ArchAIDe Portal for publications and outputs. Archaeological Data Service. https://archaeologydataservice.ac.uk/archives/view/archaide_2019/downloads_amphora.cfm, last accessed 1 August 2022.
ArchAIDE Project. 2022. Desktop, Shape Recognition. http://archaide-desktop.inera.it/shape-recognition, last accessed 1 August 2022.
Banterle, F, Callieri, M, Scopigno, R, Itkin, B, Wolf, L and Dershowitz, N. 2017. VASESKETCH: Automatic 3D Representation of Pottery from Paper Catalog Drawings. 14th IAPR International Conference on Document Analysis and Recognition; 683–690. DOI: https://doi.org/10.1109/ICDAR.2017.117
Baxter, M. 2014. Neural networks in archaeology. Available at: https://www.academia.edu/8434624/Neural_networks_in_archaeology [Last accessed 20 April 2020].
Bell, S and Croson, C. 1998. Artificial Neural Networks as a Tool for Archaeological Data Analysis. Archaeometry, 40(1): 139–151. DOI: https://doi.org/10.1111/j.1475-4754.1998.tb00829.x
Bermejo Tirado, J. 2018. Domestic patterns of tableware consumption in Roman Celtiberia. Internet Archaeology, 50. DOI: https://doi.org/10.11141/ia.50.3
Biddulph, E. 2018. From Table to Grave: Examining Table Settings in Roman Britain from Funerary Evidence. Internet Archaeology, 50. DOI: https://doi.org/10.11141/ia.50.7
Blender. 2022. 3D modelling software. https://www.blender.org/, last accessed 1 August 2022.
Chen, C, Li, O, Tao, C, Barnett, AJ, Su, J and Rudin, C. 2019. This Looks Like That: Deep Learning for Interpretable Image Recognition. 33rd Conference on Neural Information Processing System (NeurIPS 2019), Vancouver, Canada. arXiv:1806.10574.
Colley, S. 2021. Arch-I-Scan: Roman Pottery and Machine Learning, YouTube Video. https://www.youtube.com/watch?v=isA-APZGTG0.
Cool, HEM. 2006. Eating and drinking in Roman Britain. Cambridge: Cambridge University Press. DOI: https://doi.org/10.1017/CBO9780511489570
Cooper, NJ, Johnson, E and Sterry, MJ. 2018. Eating In and Dining Out in Roman Leicester: Exploring pottery consumption patterns across the town and its suburbs. Internet Archaeology, 50. DOI: https://doi.org/10.11141/ia.50.10
Dananai, A and Deru, X. 2018. Pottery Function, Dining and Funerary Assemblages. A comparative study from northern Gaul. Internet Archaeology, 50. DOI: https://doi.org/10.11141/ia.50.6
Dietler, M. 2010. Archaeologies of Colonialism: Consumption, Entanglement and Violence in Ancient Mediterranean France. Berkeley: University of California Press. DOI: https://doi.org/10.1525/9780520947948
Di Ludovico, A, Camiz, S and Pieri, G. 2013. Comparative Use of Mathematical Models in an Investigation on Mesopotamian Cylinder Seals. In: Contreras, F, Farjas, M and Melero, FJ (eds.), CAA2010 Fusion of Cultures Proceedings of the 38th Annual Conference on Computer Applications and Quantitative Methods in Archaeology, Granada, Spain, April 2010, 495–498. Oxford: Archaeopress.
Douglas, M. 1984. Standard Social Uses of Food: Introduction. In: Douglas, M (ed.), Food in the Social Order: Studies of Food and Festivities in Three American Communities, 1–39. New York: Russell Sage Foundation.
Durham, P, Lewis, P and Shennan, S. 1995. Artefact matching and retrieval using the Generalised Hough Transform. In: Wilcock, J and Lockyear, K (eds.), CAA 93. Computer Applications and Quantitative Methods in Archaeology 1993, 25–30. Oxford: Tempus Reparatum.
Ettlinger, E, Hedinger, B, Hoffmann, B, Kendrick, P, Pucci, G, Roth-Rubi, K, Schneider, G, von Schnurbein, S, Wells, CM and Zabhelicky-Scheffenegger, S. 1990. Conspectus Formarum Terrae Sigillatae Italico Modo Confectae. Bonn: R. Habelt.
Fentress, E. 2010. Cooking pots and cooking practice: an African bain-marie? Papers of the British School at Rome, 78: 145–50. DOI: https://doi.org/10.1017/S0068246200000842
Freidenreich, D. 2011. Foreigners and their Food: Constructing Otherness in Jewish, Christian, and Islamic Law. Berkeley: University of California Press. DOI: https://doi.org/10.1525/9780520950276
González Cesteros, H and Berni Millet, P. 2018. Roman Amphorae in Neuss: Augustan to Julio-Claudian Contexts. Oxford: Archeopress. DOI: https://doi.org/10.2307/j.ctvndv94n
Goody, J. 1982. Cooking, Cuisine and Class: A Study in Comparative Sociology. Cambridge: Cambridge University Press. DOI: https://doi.org/10.1017/CBO9780511607745
Huffer, D and Graham, S. 2018. Fleshing Out the Bones: Studying the Human Remains Trade with Tensorflow and Inception. Journal of Computer Applications in Archaeology, 1(1): 55–63. DOI: https://doi.org/10.5334/jcaa.8
Hossin, M and Sulaiman, MN. 2015. A Review of evaluation metrics for data classification evaluation. International Journal of Data Mining & Knowledge Management Process, 5(2): 1–11. DOI: https://doi.org/10.5121/ijdkp.2015.5201
Itkin, B, Wolf, L and Dershowitz, N. 2019. Computational Ceramicology, arXiv preprint. https://arxiv.org/abs/1911.09960.
Jurafsky, D, Chahuneau, V, Routledge, BR and Smith, NA. 2016. Linguistic Markers of Status in Food Culture: Bourdieu’s Distinction in a Menu Corpus. Journal of Cultural Analysis. DOI: https://doi.org/10.22148/16.007
Kampel, M and Sablatnig, R. 2001. Automated 3d Recording of Archaeological Pottery. In: Bearman, D and Garzotto, F (eds.), Proceedings of the International Conference on Cultural Heritage and Technologies in the Third Millennium Milan/Italy, Vol. 1: 169–182.
Kampel, M and Sablatnig, R. 2002 Computer Aided Classification of Ceramics. In: Niccolucci, F (ed.), Virtual Archaeology: Proceedings of the VAST2000 European Conference, Arezzo, 77–82. Oxford: Archaeopress.
Kampel, M and Sablatnig, R. 2006. 3D Data Retrieval for Archaeological Pottery. In: Pan, HZZ, Thwaites, H, Addison, AC and Forte, M (eds.), Interactive Technologies and Sociotechnical Systems, Proceedings of 12th International Conference, VSMM 2006, Xi’an, China, October 18–20: 387–395. New York: Springer. DOI: https://doi.org/10.1007/11890881_42
Kampel, M, Mara, H and Sablatnig, R. 2006. Automated Investigation of Archaeological Vessels. In: Luise, M (ed.), Proceedings of EUSIPCO2006: 13th European Signal Processing Conference, Florence/Italy, Sept. 4–8
Karasik, A and Smilansky, U. 2011. Computerized morphological classification of ceramics. Journal of Archaeological Science, 38: 2644–2657. DOI: https://doi.org/10.1016/j.jas.2011.05.023
LeCun, Y, Kavukcuoglu, K and Farabet, C. 2010. Convolutional networks and applications in vision. In: Proceedings of 2010 IEEE International Symposium on Circuits and Systems, (ISCAS), Institute of Electrical and Electronic Engineers (IEEE), 253–56. DOI: https://doi.org/10.1109/ISCAS.2010.5537907
Li-Ying, Q and Ke-Gang, W. 2010. Kernel Fuzzy Clustering Based Classification of Ancient-Ceramic Fragments. In: Proceedings of the IEEE Conference on Information Management and Engineering, 348–350. DOI: https://doi.org/10.1109/ICIME.2010.5477818
Lucena, M, Martínez-Carrillo, AL, Fuertes, JM, Carrascosa, F and Ruiz, A. 2014. Applying mathematical morphology for the classification of Iberian Ceramics from the upper valley of Guadalquivir river. In: Martínez-Trinidad, JF, Olvera-Lopez, JA, Salas-Rodríguez, J and Suen, CY (eds.), Pattern Recognition. MCPR 2014. New York: Springer. DOI: https://doi.org/10.1007/978-3-319-07491-7_35
Luley, B. 2018. A Terra Sigillata Revolution? Terra sigillata consumption in first-century AD Roman Mediterranean Gaul. Internet Archaeology, 50. DOI: https://doi.org/10.11141/ia.50.8
Ma, Q, Yan, A, Hu, Z, Li, Z and Fan, B. 2000. Principal Component Analysis and Artificial Neural Networks applied to the Classification of Chinese Pottery of Neolithic Age. Analytica Chimica Acta, 406: 247–256. DOI: https://doi.org/10.1016/S0003-2670(99)00764-3
Maiza, C and Gaildrat, V. 2005. Automatic Classification of Archaeological Potsherds. In: Dimitri, P (ed.), The 8th International Conference on Computer Graphics and Artificial Intelligence, 3IA’2005, Limoges, France, 11/05/2005-12/05/2005, 135–147. Limoges: Université de Limoges.
Makridis, M and Daras, P. 2012. Automatic classification of archaeological pottery sherds. Journal on Computing and Cultural Heritage, 5(4). DOI: https://doi.org/10.1145/2399180.2399183
Marshall, M and Seeley, F. 2018. From the Spreadsheet to the Table? Using ‘spot-dating’ level pottery records from Roman London to explore functional trends among open vessel forms. Internet Archaeology, 50. DOI: https://doi.org/10.11141/ia.50.9
Matplotlib. 2022. Python plotting package. https://pypi.org/project/matplotlib/, last accessed 1 August 2022.
Mees, AW. 2007. Diffusion et datation des sigillées signées et décorées de La Graufesenque en Europe. L’influence de l’armée sur l’évolution du pouvoir d’achat et du commerce dans les provinces romaines. Actes du Congrès de Langres. Marseille: Société Française d’Étude le Céramique Antique en Gaule, 145–208.
Mees, AW. 2016. Was there a Difference between Roman ‘Civil’ and ‘Military’ Samian (terra sigillata) Market Supply? Finding answers with statistical distribution analysis methods. Internet Archaeology, 50. DOI: https://doi.org/10.11141/ia.50.16
Mees, AW and Polak, M. 2013. Scattered Pots. Exploring Spatial and Chronological Aspects of Samian Ware. In: Fulford, M and Durham, E (eds.), Seeing Red: New Economic and Social Perspective on Terra Sigillata, 36–48. London: Institute of Classical Studies.
Monteil, G. 2013. Samian plain vessels – a review of their development in the 2nd c. AD. In: Fulford, M and Durham, E (eds.), Seeing Red: New economic and social perspectives on Gallo-Roman terra sigillata, BICS Supplement 102, London, 349–367.
Núñez Jareño, SJ, van Helden, DP, Mirkes, EM, Tyukin, IY and Allison, PM. 2021a. Learning from Scarce Information: Using Synthetic Data to Classify Roman Fine Ware Pottery. Entropy, 23(1140). DOI: https://doi.org/10.3390/e23091140
Núñez Jareño, SJ, van Helden, DP, Mirkes, EM, Tyukin, IY and Allison, P. 2021b. Arch-I-Scan data repository. https://github.com/ArchiScn.
Piccoli, C, Aparajeya, P, Papadopoulos, GT, Bintliff, J, Leymarie, FF, Bes, P, van der Enden, M, Poblome, J and Daras, P. 2015. Towards the automatic classification of pottery sherds: two complementary approaches. In: Traviglia, A (ed.), Across Space and Time: Papers from the 41st Conference on Computer Applications and Quantitative Methods in Archaeology, Perth, 25–28 March 2013, 463–474. Amsterdam: Amsterdam University Press.
Pitts, M. 2014. Reconsidering Britain’s first urban communities. Journal of Roman Archaeology, 27: 133–73. DOI: https://doi.org/10.1017/S1047759414001196
Pitts, M. 2015. Globalisation, circulation and mass consumption in the Roman world. In: Pitts, M and Versluys, MJ (eds), Globalisation and the Roman World: World History, Connectivity and Material Culture, 69–98. Cambridge: Cambridge University Press. DOI: https://doi.org/10.1017/CBO9781107338920.006
Poblome, J, Willet, R, Firat, N, Martens, F and Bes, P. 2013. Tinkering with Urban Survey Data: How Many Sagalassos-es do we Have? In: Johnson, P and Millett, M (eds.), Archaeological Survey and the City, 146–175. Oxford: Oxbow. DOI: https://doi.org/10.2307/j.ctvh1dwwz.10
Polak, M. 2017. Old Wine in New Bottles. Reflections on the Organization of the Production of Terra Sigillata at La Graefesenque. In: Bird, J (ed.), Form and Fabric: Studies in Rome’s Material Past in Honour of B.R. Hartley, 115–121. Oxford: Oxbow. DOI: https://doi.org/10.2307/j.ctvh1dtvk.18
Selvaraju, RR, Gogswell, M, Das, A, Vedantam, R, Parikh, D and Batra, D. 2019. Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization. International Journal of Computer Vision, 128: 336–359. DOI: https://doi.org/10.1007/s11263-019-01228-7
Smith, P, Bespalov, D, Shokoufandeh, A and Jeppson, P. 2010. Classification of Archaeological Ceramic Fragments Using Texture and Color Descriptors. In: Proceedings of the Symposium on Electronic Imaging: Science and Technology – Storage and Retrieval for Image and Video Databases IV. Vol. 2670, 426–437. DOI: https://doi.org/10.1109/CVPRW.2010.5543523
Sun, Y, Chockler, H, Huang, X and Kroening, D. 2019. Explaining Image Classifiers using Statistical Fault Localization, arXiv preprint. https://arxiv.org/abs/1908.02374.
Teichner, F and Pons Pujol, L. 2008. Roman Amphora Trade across the Straits of Gibraltar: an ancient ‘anti-economic practice’. Oxford Journal of Archaeology, 27(3): 3030–314. DOI: https://doi.org/10.1111/j.1468-0092.2008.00309.x
Timby, J. 2020. Review of P. M. Allison, M. Pitts, and S. Colley (eds) Big Data on the Roman Table: New approachts to tablewares in the Roman world. Internet Archaeology, 50(2018). Britannia, 51: 529–530. DOI: https://doi.org/10.1017/S0068113X20000161
Tyers, PA. 2014. Potsherd. Atlas of Roman Pottery. https://potsherd.net/atlas/types/sigillata/gallery. (Accessed 25/01/2022).
Tyukin, I, Sofeikov, K, Levesley, J, Gorban, AN, Allison, P and Cooper, NJ. 2018. Exploring Automated Pottery Identification [Arch-I-Scan]. Internet Archaeology, 50. DOI: https://doi.org/10.11141/ia.50.11
Van der Veen, V. 2018. Form Follows Function. A new approach to determining vessel function. Internet Archaeology, 50. https://intarch.ac.uk/journal/issue50/2/index.html. DOI: https://doi.org/10.11141/ia.50.2
van Oyen, A. 2015. The Roman City as Articulated through Terra Sigillata. Oxford Journal of Archaeology, 34(3): 279–299. DOI: https://doi.org/10.1111/ojoa.12059
van Oyen, A. 2016. How Things make History. The Roman Empire and its Terra Sigillata Pottery. Amsterdam: Amsterdam University Press. DOI: https://doi.org/10.5117/9789462980549
Wang, H, He, Z, Huang, Y, Chen, D and Zhou, Z. 2017. Bodhisattva Head Images Modeling Style Recognition of Dazu Rock Carvings based on Deep Convolutional Network. Journal of Cultural Heritage, 27: 60–71. DOI: https://doi.org/10.1016/j.culher.2017.03.006
Whitelaw, T. 2012. Collecting Cities: some Problems and Prospects. In: Johnson, P and Millett, M (eds.), Archaeological Survey and the City, 70–106. Oxford: Oxbow. DOI: https://doi.org/10.2307/j.ctvh1dwwz.8
Willet, R. 2014. Experiments with diachronic data distribution methods applied to Eastern Sigillata a in the Eastern Mediterranean. Journal on Hellenistic and Roman Material Culture, 3(1): 39–69. DOI: https://doi.org/10.11116/HEROM.3.3
Woolf, G. 1998. Becoming Roman: The Origins of Provincial Civilization in Gaul. Cambridge: Cambridge University Press. DOI: https://doi.org/10.1017/CBO9780511518614