A couple of weeks ago I wrote again about determining the gender of the maker of a handprint by ratios of finger lengths. Well, staying with the hand, this column is about a project that attempted to train a maching learning (ML) program to determing the gender of the makers of finger fluting. We are all probably aware of finger fluting in caves, it is found all over the world, but it has always been somewhat peripheral to the subject of cave art itself. It is, however, purposeful markings made by people on the cave walls so it needs to be covered in any consideration of cave art. Various examples have been attributed to Neandertals, as well as Homo sapiens men, women and children. Now, a team in Australia is using artificial intelligence to try to clarify the makers of these marks.
“Flutings have the potential to reveal information about age, sex, height, handedness and idiosyncratic markmaking choices among unique individuals who form part of larger communities of practice. However, previous methods for making any determination about the individual artist from finger flutings have been shown to be unreliable4. Accordingly, we propose a novel digital archaeology approach to begin understanding this enigmatic form of rock art by leveraging machine learning (ML) as a tool for uncovering patterns from two datasets, one tactile and one virtual, collected from a modern population. We aimed to determine whether ML can reveal subtle differences in the sex of the artist based on their finger-fluted images.” (Jalandoni et al. 2025:1) In other words they will attempt to have machine learning programs learn to distinguish information like gender and age by analyzing finger fluting created by volunteers. If successful, this could then be applied to finger fluting in cave walls to learn more about the persons who originally created the marks.
“Experiments were conducted - both with adult participants in a tactile setup and using VR headsets in a custom-built program – to explore whether image-recognition methods could learn enough from finger fluting images made by modern people to identify the sex of the person who created them.” (Lock and Egan 2025:1) The team had participants actually make finger flutings in clay as well as virtually while being videotaped. “Two controlled experiments with 96 adult participants were conducted with each person creating nine flutings twice: once on a moonmilk clay substitute developed to mimic the look and feel of cave surfaces and once in virtual reality (VR) using Meta Quest 3. Images were taken of all the flutings, which were then curated and two common image-recognition models were trained on them. (Lock and Egan 2025:1-2)
Disappointingly, the tests did not produce reliable results. “The VR images did not yield reliable sex classification; even when accuracy looked acceptable in places, overall discrimination and balance were weak. But the tactile images performed much better. ‘Under one training condition, models reached about 84% accuracy, and one model achieved a relatively strong discrimination score.’ Dr. Tuxworth said. However, the models did learn patterns specific to the dataset; for example, subtle artifacts of the setup, rather than robust features of fluting that would hold elsewhere, which meant there was more work to be done.” (Lock and Egan 2025:1-2) Doctor Gervase Tuxworth is one of the experimental team that conducted this study. His statement suggests that the test results were highly variable.
“Overall, the deep learning models achieved high accuracy during training, with AUC values exceeding 0.85 for certain tactile image conditions. These results suggest that the models effectively learned patterns within the tactile dataset and demonstrated strong discrimination between male and female-generated finger fluting images. However, the relatively lower AUC values for virtual images, coupled with their unstable test accuracy, indicate that they do not provide sufficiently distinct features for reliable sex classification. This discrepancy highlights the greater robustness of tactile images over virtual images in capturing relevant classification features. Despite the promising performance on tactile images, deep learning models exhibited a pronounced disparity between training and test performance. While training accuracy consistently increased, reaching near-perfect levels in the later epochs, test accuracy remained unstable and showed no substantial improvement over time. This pattern indicates overfitting, where the models effectively learn dataset-specific features but fail to generalize to unseen test data.” (Jalandoni et al. 2025:10) I find the previous paragraph somewhat confusing. It states “accuracy consistently increased, reaching near-perfect levels” and “accuracy remained unstable and showed no substantial improvement” in two contiguous sentences. In any case, the team did not get reliable results.
There are a number of possible sources of inaccuracy in the test results. “The instability in test accuracy further suggests that the models struggle to extract robust and generalizable patterns from the finger fluting images, ultimately limiting their reliability for sex classification. A possible contributing factor to this challenge could be individual variation in hand size and fluting characteristics. For example, some females may have larger hands and exhibit stronger fluting patterns resembling those of males, while some males may have smaller hands and display lighter, less pronounced fluting strength. This variability could confuse the model, making it difficult to accurately differentiate between sexes and ultimately hindering its performance on the test set. These results underscore the critical need to increase the dataset size to alleviate overfitting and improve the model’s generalizability. Moreover, the inherent variability in finger fluting images may impose fundamental limitations on the feasibility of using deep learning for sex classification, suggesting that alternative approaches or additional contextual data may be necessary to enhance classification accuracy. The limited success of the tactile data in sex prediction underscores the importance of material-based approaches in understanding finger flutings. While the VR data failed to provide useful results, it opens up new and exciting possibilities for exploring the dynamic aspects of fluting and artistic intent in the future. While a modest achievement, this study highlights the potential of ML to enhance traditional archaeological methods”. (Jalandoni et al. 2025:10) Not every try is guaranteed success.
So, this test did not manage to display reliable accuracy, too many variables in the creation of finger fluting seemingly overwhelmed the software. Also, the experiment apparently did not include children, and it is thought that much finger fluting, at least in European cave contexts, was created by children. If successful, this project would have been a really wonderful development but, alas, it was not to be. Better luck next time.
NOTE: Some images in this column were retrieved from the internet with a search for public domain photographs. If any of these images are not intended to be public domain, I apologize, and will happily provide the picture credits if the owner will contact me with them. For further information on these reports you should read the original reports at the sites listed below.
REFERENCES:
Andrea Jaladoni, Robert Haubt, Calum Farrar, Gervase Tuxworth , Zhongyi Zhang , Keryn Walshe and April Nowell, 2025, Using digital archaeology and machine learning to determine sex in finger flutings, Scientific Reports, 15:34842. https://doi.org/10.1038/s41598-025-18098-4. Accessed online 12 October 2025.
Lock, Lisa, and Robert Egan, 2025, VR experiments train AI to identify ancient finger-fluting artists,
16 October 2025, The GIST, by Griffith University, https://phys.org/news/2025-10-vr-ai-ancient-finger-fluting.html.
Posts
No comments:
Post a Comment