Section: New Results

Machine Learning for Biodiversity Informatics

Herbarium based phenological research offers the potential to provide novel insights into plant diversity and ecosystem processes under future climate change. The goal of this study [11], conducted in collaboration with US and French ecologists, is to automate the scoring of reproductive phenological stages within a huge amount of digitized herbaria and provide significant resources for the ecological and organismal scientific communities. Specifically, we address three questions: 1) Can fertility, i.e., the presence of reproductive structures, be automatically detected from digitized specimens using deep learning? 2) Are the detection models generalizable to different herbarium data sets? and 3) Is it possible to automatically record stages (i.e., phenophases) within longer phenological events on herbarium specimens? This is the first time that such an analysis has been conducted at this scale, on such a large number of herbarium specimens and species. The results obtained for 7782 species of plants representing angiosperms, gymnosperms, and ferns suggest that it is possible to consider large-scale phenological annotation across broad phylogenetic groups.

Species distribution models (SDM) are widely used for ecological research and conservation purposes. Given a set of species occurrences and environmental data (such as climatic rasters, soil occupation, altitude, etc.), the aim is to infer the spatial distribution of the species over a given territory. In a previous work, we showed that using deep convolutional networks significantly improved predictive performance compared to conventional punctual approaches. We have deepened this methodology with two main contributions. The first one is to extend the model to explicitly take into account species co-occurrences [22]. This is achieved through a new multimodal architecture that allows the joint learning of biotic and abiotic patterns in a common representation space. The second contribution is to experiment deep SDMs at the scale of several tens of thousands of species and tens of millions of occurrences. These contributions were made possible thanks to the use of supercomputer supercomputer Jean Zay (more than 1000 GPUs) of the GENCI national infrastructure.

We run a new edition of the LifeCLEF evaluation campaign [29] with the involvement of 16 research teams worldwide. The main outcomes of the 2019-th edition are:

• GeoLifeCLEF. The main result of the second edition of this challenge [24] is that deep convolutional models outperform the most efficient machine learning models used in ecology (such as random forests or boosted trees). In particular, they are able to transfer knowledge from animals distribution to plant distribution, which had never been shown before.

• PlantCLEF. The 2019-th edition of the plant identification challenge [26] was designed to evaluate automated identification on the flora of data deficient regions, tropical ones in particular. It is based on a dataset of 10K species mainly focused on the Guiana shield and the Northern Amazon rainforest, an area known to have one of the greatest diversity of plants and animals in the world. The results reveal that the identification performance in this context is considerably lower than the one obtained on temperate plants of Europe and North America. The performance of convolutional neural networks fall due to the very low number of training images for most species and the higher degree of noise that is occurring in such data.

• Bird sounds identification. The 2019-the edition of the BirdCLEF challenge [41] focuses on the difficult task of recognizing all birds vocalizing in omni-directional soundscape recordings. Therefore, the dataset of the previous year has been extended with more than 350 hours of manually annotated soundscapes that were recorded using 30 field recorders in Ithaca (NY, USA). The main outcome is that the recognition performance can be significantly improved thanks to sophisticated data augmentation methods adapted to the problem.

In addition to organizing these challenges, we published a synthesis of the LifeCLEF evaluation campaign since its inception in 2011. This synthesis [44] is part of a larger book published on the occasion of the 20th anniversary of the CLEF international research forum. It highlights the rapid progress that automatic identification has made over the past decade, and allows us to take a step back on the future challenges of this discipline.

In this work [47], done in collaboration with the REAL-OPT team, we introduce a new activation checkpointing method which allows to significantly decrease memory usage when training Deep Neural Networks with the back-propagation algorithm. Similarly to checkpointing techniques coming from the literature on Automatic Differentiation, it consists in dynamically selecting the forward activations that are saved during the training phase, and then automatically recomputing missing activations from those previously recorded. We propose an original computation model that combines two types of activation savings: either only storing the layer inputs, or recording the complete history of operations that produced the outputs (this uses more memory, but requires fewer recomputations in the backward phase), and we provide an algorithm to compute the optimal computation sequence for this model, when restricted to memory persistent sequences. We provide a PyTorch implementation that processes the entire chain, dealing with any sequential DNN whose internal layers may be arbitrarily complex and automatically executing it according to the optimal checkpointing strategy computed given a memory limit. Through extensive experiments, we show that our implementation consistently outperforms existing checkpointing approaches for a large class of networks, image sizes and batch sizes.