Artificial intelligence: the machine is learning

22. Februar 2024

Core ML is a library by Apple and allows the integration of machine learning algorithms into apps. This is what we have done with bcAdmin4. We have trained various models with calls of bats on different levels. Starting with a first model classifying genus/groups we worked our way towards species specific models. All this was implemented directly into bcAdmin4, so you will not need an extra program like batIdent. And thanks to all the labor Apple has put into CoreML the classification is now roughly 5x faster than back with batIdent. We never believed it could sped up so much.

How is the new model performance?

The most asked question by users is about the performance of the algorithm and if all calls are classified correctly. Yet, it can't be answered easily. How to measure the performance? It heavily depends on the recorded calls, their quality, the species and the situation they were recorded. Of course a mathematical approach exists which is to use a subset of the training calls for validation. We went a step further and published a set of independent validation recordings. The following table shows on a per call basis (good+bad calls as per our classification) the results on genus level.

The results do look pretty well. Nearly any identification achieved was correct. The good result is somehow lowered since only a selection of calls were used and no results for all calls of a recording on a per recording level were calculated. How do really bad calls or calls of other species influence these results? We tested it and added a comparison with the results of a batIdent identification on genus level.

And since we can identify single species, why not try that as well? Again in comparison to batIdent. In the end, nearly 90% of correct classifications made us rather happy.

All within bcAdmin4

The models of the above tables are all included in bcAdmin4. The most current methodology will be made available with bcadmin4 1.4.1 and includes a specific set of probabilities on various levels.