In this sense, "putting on the evolutionary pair of glasses" is nothing else but to account for the inevitable theory-ladenness of observation. There seems in fact no good reason to assume why machine learning should be exempt from theory-ladenness, which is one way to read the NFL theorems.

A #philscy perspective on our work: As machine learning is underdetermined (following the no-free-lunch theorems), the training process has little chance to infer an internal representation that is phylogenetically plausible - unless we introduce explicit inductive bias.

...in particular, we can consider the splits diagrams of all quartets of taxa and train the neural network so that the splits diagrams for all quartets degenerate to trees.

If we pass a (knwon and well-established) phylogenetic tree to the training process of the neural network, we can utilize mathematical structures behind evolutionary distances...

As has been observed by other authors before, the internal representations of neural networks are usually incompatible with known phylogenies. So, how could a neural network learn to infer features that are evolutionarily plausible?

Link to the full paper: www.sciencedirect.com/science/arti...

Sieht sehr böhmisch aus!

In Henk de Regt’s book on Understanding Scientific Understanding, skill takes center stage for understanding as a primary goal of science.

…and most importantly: Calculators could actually *do* arithmetic, and did not imitate the process of doing arithmetic.

There is also the "classical" paper by Zech et al. which identifies potential shortcuts in pneumonia detection (see e.g. Fig. 2)

journals.plos.org/plosmedicine...

AI predicts beer consumption from knee x-rays www.nature.com/articles/s41...

cdn.bsky.app/img/feed_thu...

Here's what we've been missing in the meantime. Better tell the owner of the "Hammer of Justice" not to hit his own thumb.