Indrė Žliobaitė

In many scientific and data-driven applications, machine learning models are increasingly used as measurement instruments, rather than merely as predictors of predefined labels. When the measurement function is learned from data, the mapping from observations to quantities is determined implicitly by the training distribution and inductive biases, allowing multiple inequivalent mappings to satisfy standard predictive evaluation criteria. We formalize learned measurement functions as a distinct focus of evaluation and introduce measurement stability, a property capturing invariance of the measured quantity across admissible realizations of the learning process and across contexts. We show that standard evaluation criteria in machine learning, including generalization error, calibration, and robustness, do not guarantee measurement stability. Through a real-world case study, we show that models with comparable predictive performance can implement systematically inequivalent measurement functions, with distribution shift providing a concrete illustration of this failure. Taken together, our results highlight a limitation of existing evaluation frameworks in settings where learned model outputs are identified as measurements, motivating the need for an additional evaluative dimension.

Over the last few decades, the term “taxon-free” has been widely used in paleoecological literature to highlight reasoning via ecomorphological or other functional traits of organisms as opposed to reasoning via taxonomic affiliations and phylogenetic conservatism. In practice, however, “taxon-free” inferences are very rarely free from using taxonomic information and occasionally they are even using phylogenetic conservatism to infer environmental or ecological conditions of taxa in the past. While some publications acknowledge the mismatch between the semantics of the term and the practice, others highlight the lack of reliance on taxonomic information as an advantage of the methodology, even in practice that reliance is present in a masked form. Here I present a structured survey on the use of the term “taxon-free” in paleoecological literature with an emphasis on mammalian community paleoecology. Based on the results I advocate dropping the use of the term “taxon-free” to minimize the potential for misinterpretations and use the term “trait-based” approaches instead, or alternatively, “phylogeny-free” if absence of reliance on phylogenetic niche conservatism is to be highlighted.

Animal bones from archaeological contexts can reveal the interplay between past environments and human societies. Resource acquisition shaped many aspects of past societies and influenced the development of trade networks and migration. Fish have been a cornerstone of human subsistence for millennia, yet the rise of commercial fishing and trade was complex. Here, we synthesised a database of ∼1.9 million zooarchaeological fish records spanning 2000 years across Europe. Using machine-learning of catch compositions alongside fish thermal tolerances, we show that fisheries became less local over time, with homogenisation coinciding with Little Ice Age-associated cooling, a period of documented resource scarcity, concurring with growing trade. Moreover, increased proportions of marine taxa and more specialist marine fisheries were observed in the preceding Medieval Climate Anomaly, to sustain concurrent urban and population growth. Enhanced use of marine protein buffered food insecurity, whilst signalling the transition from localised to trans-regional trade networks.

The literature on machine learning in the context of data streams is vast and growing. This indicates not only an ongoing interest, but also an ongoing need for a synthesis of new developments in this area. Here, we reformulate the definitions of supervised data-stream learning, alongside consideration of contemporary concept drift and temporal dependence. Equipped with this, carry out a fresh discussion of what constitutes a supervised data-stream learning task; including continual and reinforcement learning; highlighting major assumptions and constraints. We carry out a fresh reconsideration of approaches and methods, with regard to their suitability to modern settings. But more than a categorization of state-of-the-art streaming methods, we provide a re-introduction to what is supervised stream learning, and our emphasis here is a survey of settings, and algorithmic settings. Our main goal is to pull theory and practice of supervised learning over data streams closer together. We conclude that practical stream learning does not mandate an online-learning regime. In the modern context, learning regimes should be selected and developed according to the factual data arrival mode, resource constraints, and maximum robustness and trustworthiness. We finish with a set of recommendations to this effect.

Brown et al. (2024) highlight that organisms invest a constant amount of energy into the production of viable offspring per unit of body mass per generation. This explains how diversity in life can exist. We interpret their result in relation to balancing offspring costs in real vs. physiological time.

This extended abstract was presented at the Nectar Track of ECML PKDD 2024 in Vilnius, Lithuania. The content supplements a recently published paper "Laws of Macroevolutionary Expansion" in the Proceedings of the National Academy of Sciences (PNAS).

Van Valen’s law of constant extinction postulates that in comparable ecological contexts, the probability for a taxon to survive to the next time interval is independent of how long it has already existed. The law implies that species do not age, that is, the survival probability from one time step to the next does not decrease with species’ age. While the original law describes survivorship over time, here we show that there is a counterpart to the law describing species dynamics in space. Indeed, the probability for a taxon to expand a step further in its range size does not depend on how large of a range the taxon already occupies. We show that such patterns of stochastically constant expansion are common in the global mammalian fossil record throughout the Cenozoic and are consistent with the mammalian species distribution in the present-day world. Intriguing is that the law in space holds best for the linear dimension of the range rather than for the range area. We interpret this as suggesting that species primarily expand via the frontier of their range; this process is known as leading-edge expansion in ecology. And while the difficulty of expanding in linear steps remains constant, the area gains diminish. In line with evolutionary theory collectively known as the Red Queen’s hypothesis, this suggests that the effectiveness of expansion saturates, even if species longevity does not. When interpreted in reverse, this explains from the macroevolutionary perspective why larger species ranges are associated with lower extinction risks.