Alternaria spore exposure in Bavaria, Germany, measured using artificial intelligence algorithms in a network of BAA500 automatic pollen monitors.

Although Alternaria spores are well-known allergenic fungal spores, automatic bioaerosol recognition systems have not been trained to recognize these particles until now. Here we report the development of a new algorithm able to classify Alternaria spores with BAA500 automatic bioaerosol monitors. The best validation score was obtained when the model was trained on both data from the original dataset and artificially generated images, with a validation unweighted mean Intersection over Union (IoU), also called Jaccard Index, of 0.95. Data augmentation techniques were applied to the training set. While some particles were not recognized (false negatives), false positives were few. The results correlated well with manual counts (mean of four Hirst-type traps), with R2 = 0.78. Counts from BAA500 were 1.92 times lower than with Hirst-type traps. The algorithm was then used to re-analyze the historical automatic pollen monitoring network (ePIN) dataset (2018-2022), which lacked Alternaria spore counts. Re-analysis of past data showed that Alternaria spore exposure in Bavaria was very variable, with the highest counts in the North (Marktheidenfeld, 154 m a.s.l.), and the lowest values close to the mountains in the South (Garmisch-Partenkirchen, 735 m a.s.l.). This approach shows that in our network future algorithms can be run on past datasets. Over time, the use of different algorithms could lead to misinterpretations as stemming from climate change or other phenological causes. Our approach enables consistent, homogeneous treatment of long-term series, thus preventing variability in particle counts owing to changes in the algorithms.

Self-Replication in Neural Networks.

A key element of biological structures is self-replication. Neural networks are the prime structure used for the emergent construction of complex behavior in computers. We analyze how various network types lend themselves to self-replication. Backpropagation turns out to be the natural way to navigate the space of network weights and allows non-trivial self-replicators to arise naturally. We perform an in-depth analysis to show the self-replicators' robustness to noise. We then introduce artificial chemistry environments consisting of several neural networks and examine their emergent behavior. In extension to this work's previous version (Gabor et al., 2019), we provide an extensive analysis of the occurrence of fixpoint weight configurations within the weight space and an approximation of their respective attractor basins.