Machine learning (ML) has become an interesting part of data analysis in the last decade. Many applications of ML have led to increased interests in its usage and development. As part of our citizen science course, we were honored to be joined by Laurens Hogeweg (Naturalis Biodiversity Center) and Mark Rademaker (PHD Candidate at Netherland Institute for Sea Research (NIOZ)).
Mark spoke on the topic, “Deep learning and Species Distribution Modelling (SDM)”. We began by describing the deep neural network and its architecture. We discussed the choice of hidden layers, weights and the various loss functions that are used. An application of the deep neural network in Mark’s research work was discussed and its relevance was compared to the MaxEnt. Another interesting feature was how the model could be improved by extending the number of observations and using the feature importance and interaction effects to determine the variable importance for each individual grid cell.
After a short break, Laurens presented on the topic: “AI for Biodiversity image recognition”. We began with what deep learning was. We learnt that deep networks contain millions of parameters and the problem-relevant properties are learned automatically as part of training the model. Large data sets and computational power are a necessity to determine the good values for the large number of parameters. Another important part of the discussion was how to make a large-scale biodiversity -wide image recognition model using the long tail and evaluating the large scale models (either using recall, precision or accuracy) and also using a fine-grained hierarchical classification (taxonomy and labeling). We discussed briefly about transfer learning and some applications of the model we discussed. Some of the applications noted are ObsIdentify and Artsorakel (both available on App Store and Play Store). Other applications included a smart insect camera, intermezzo, the Van Groenendael-Krijger collection, etc.
Was the session worth it? Well I hope you will come to the conclusion with me that it indeed was. We all came to a great understanding of the models being developed by Wouter Koch (a member of the project), and also appreciate the importance of Machine Learning in each of our individual projects.