Hi all, I'd like to take this opportunity to add my two-pennies worth on this discussion, and make the most of the opportunity to share my recent preprint on one of the challenges faced when using AI on our particular type of data (https://arxiv.org/abs/2601.18448).
>From my perspective, the most promising applications aren’t necessarily in automating landmark placement, but earlier in the processing chain - automated segmentation. We could discuss the point of the use of AI to great lengths but having seen some earlier conversations here I'm not sure that's the most valuable use of our time... but for most data scientists AI is a powerful tool to automate difficult, mundane, or time consuming tasks, that normally require human precision; for me that has always been more segmentation than landmark placement, and there are some pretty cool advances in the world of data science about segmentation, including CT data. On another note, I have noticed that few are talking about the current popularity in some research to use AI as a classification tool. I dont particularly want to discuss the epistemological and algebraic concerns with LDA/CVA here, these issues are already well published, but for those who are using algorithms for the purpose of classifying new specimens, I have a preprint currently under review that I would like to share (and would love feedback as long as its constructive). The link is above but I paste it here too: https://arxiv.org/abs/2601.18448 I recently explored the impact that procrustes alignment itself has on the preparation of our data for the purpose of machine learning, and advise caution when using any form of AI on landmark data without taking appropriate precautions to ensure we are complying with the most basic requirements of computational learning - avoiding data leakage into the test or validation sets. Train:Test:Validation splits have to be done before alignment itself! A mistake I made in my early research that I admit I did not know about and am trying to correct now. I additionally found that the stability and power of predictive models, for tasks even as basic as simple regression to test for allometric signals, is weakened considerably by questions of sample size - something we already knew, but thought it was worth reflecting upon again in more detail in the lense of AI applications too. I also show that whatever algorithm we use, if we cannot find a way to incorporate an understanding of the spatial autocorrelation embedded into the definition of a landmark, the accuracy of whatever model we use will decrease substantially. This is incredibly relevant for models that assume linearity in our data. When talking about spatial autocorrelation, most simple linear models are unable to truly account for these properties. I have had discussions with some people in this group in person before about this topic, when coinciding in different conferences etc, and have many ideas about how to look into overcoming it, however haven't had much time to work on it properly and formalise these ideas. I definitely have data though that shows that there are much better ways of training algorithms on landmarks that substantially improve those models that are blind to the interaction that exists between landmarks in space. If anyone is interested in discussing or even collaborating, please let me know, and I am more than happy to discuss these points in more detail here. Anyway, this is my first time posting on this group, I thought it was a good moment to share my thoughts. Have a great day! *Lloyd A. Courtenay* *--* *PhD* *Archaeologist: Specialised in Microscopy, Data Science and Applied Statistics* *Postdoctoral Fellow* *ANR Access-ERC* *BSMART* Centre National de la Recherche Scientifique (CNRS) PACEA UMR 5199 Bât B2, allée Geoffroy Saint Hilaire - CS 50023 33615 Pessac cedex, France Work Email: [email protected] Mobile: (+34) 633 647 825 Website: https://quanta.eva.mpg.de/ Member of the Editorial Board of* Scientific Reports * *https://www.nature.com/srep/ <https://www.nature.com/srep/>* [image: EBM signature banner] Member of the Editorial Board of* Quaternary Environments and Humans * https://www.sciencedirect.com/journal/quaternary-environments-and-humans On Fri, 1 May 2026 at 14:15, Fabio Machado <[email protected]> wrote: > Dear all, > > I would like to revisit the topic of AI-related issues in morphometrics > from a slightly different lens. > > It is undeniable that there is an increasing push for the use of AI in > society at large and in science. An example of this last one is the recent > PRFB program focused on the integration of AI into biology ( > https://www.nsf.gov/funding/opportunities/prfb-postdoctoral-research-fellowships-biology/nsf26-504/solicitation). > I think it’s fair to say that very soon we will be under pressure to > include AI tools in most, if not all, aspects of our practice (those > teaching might already be experiencing that). > > We discussed, in general terms, how AI can be used for coding, writing, > and other tasks, but I wonder if there are clearer examples of AI-related > tools (Machine Learning) that can benefit morphometrics in general. > > The most obvious example are ML automated landmarking methods, which I > feel are mostly beneficial. But other than that, what are the potential > beneficial uses? > > I've been seeing a lot of suggestions for ML approaches that produce > embedded spaces that serve as stand-ins for standard ordination methods > such as principal component analysis or canonical variate analysis. In my > opinion, I always thought that these standard ordination approaches are > valuable because they rely on simple algebraic formulations that allow us > to traverse between multiple morphometric representations. However, from > the few examples I've seen, the axes of these ML embedded spaces are > usually later correlated with standard geometric morphometric measurements > to generate shape descriptors. This, however, severs the connection between > multiple representations (e.g., figure-space and shape-space) that makes > morphometrics so powerful and appealing. > > So, are there any other potential ways for us to “fit in” into this newly > constructed necessity more directly, or is there a fundamental gap between > classical/geometric morphometrics approaches and new ML methods? I > personally don’t see the appeal of any post-data-processing ML use cases, > but I am always curious to hear other folks' opinions. > > Best, > > [image: EMORPH2.png] > > *Fabio Machado, D.Sc.* > > *Assistant Professor and Curator of Vertebrates* > > > *Department of Biology Oklahoma State University* > > *https://MachadoFA.github.io <https://MachadoFA.github.io>* > > > > > > -- > You received this message because you are subscribed to the Google Groups > "Morphmet" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to [email protected]. > To view this discussion visit > https://groups.google.com/d/msgid/morphmet2/CH3PR12MB9284A2FD0A3CD1D8D571DB3DA2372%40CH3PR12MB9284.namprd12.prod.outlook.com > <https://groups.google.com/d/msgid/morphmet2/CH3PR12MB9284A2FD0A3CD1D8D571DB3DA2372%40CH3PR12MB9284.namprd12.prod.outlook.com?utm_medium=email&utm_source=footer> > . > -- You received this message because you are subscribed to the Google Groups "Morphmet" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To view this discussion visit https://groups.google.com/d/msgid/morphmet2/CANxkrZruPGT5gt0hRhm3NSUTA818DXdXAWPUgSei-iRvbsQ8Lw%40mail.gmail.com.
