Section: New Results

Machine Learning for High-dimensional Data

Uncertainty in Fine-grained Classification

Participants : Titouan Lorieul, Alexis Joly.

Uncertainty is critical when considering classification problems that involve thousands of domain specific labels. A picture of a plant, for instance, contains only a partial information that is usually not sufficient to determine its scientific name with certainty. We first work on the modelling of such uncertainty in the context of crowdsourcing systems involving experts as well as non expert annotators. We rely on Bayesian inference to learn the annotators’ confusion and to optimally assign them new items to be validated. In particular, we work on a non-parametric version of this model allowing to combine annotators’ suggestions even when the number of possible labels is undetermined and might change over time [33]. In mirror to this research, we also work on the uncertainty of automatic classifiers, in particular deep convolutional neural networks trained on massive amounts of plant images. We conduct an experimental study aimed at evaluating quantitatively the intrinsic data ambiguity of image-based plant observations [64], and we started working on new methods for estimating the uncertainty of ensembles of deep neural networks by fitting a Dirichlet distribution on the set of their predictions. Besides, we study the use of different taxonomic levels as a source of potential reduction in prediction uncertainties [66].

Species Distribution Modelling based on Citizen Science Data

Participants : Christophe Botella, Alexis Joly.

Species distribution models (SDM) are widely used for ecological research and conservation purposes. Given a set of species occurrence, the aim is to infer its spatial distribution over a given territory. Because of the limited number of occurrences of specimens, this is usually achieved through environmental niche modeling approaches, i.e. by predicting the distribution in the geographic space on the basis of a mathematical representation of their known distribution in environmental space (= realized ecological niche). The environment is in most cases represented by climate data (such as temperature, and precipitation), but other variables such as soil type or land cover can also be used. In [24], we study for the first time the relevance of a species distribution model computed from automatically identified plant observations made by citizens rather than from classical inventories made by experts. The results show that the resulting models have a great potential for the early detection of new invasions. In [65] and [60], we propose a deep learning approach to species distribution modelling in order to improve the predictive effectiveness in the context of massive amount of occurrence data. Non-linear prediction models have been of interest for SDM for more than a decade but our study is the first one bringing empirical evidence that deep, convolutional and multilabel models might participate to resolve the limitations of SDM.

Evaluation of Species Identification and Prediction Algorithms

Participants : Alexis Joly, Hervé Goëau, Christophe Botella, Jean-Christophe Lombardo.

We ran a new edition of the LifeCLEF evaluation campaign [45] with the involvement of 13 research teams worldwide. The main novelties and outcomes of the 2018-th edition are the following:

• GeoLifeCLEF: a new challenge [71] dedicated to the location-based prediction of species based on spatial occurrences and environmental data tensors. The evaluation concludes that deep environmental convolutional neural networks perform better than spatial models or ponctual environmental models.

• Man vs. Machine plant identification: To evaluate how far automated identification systems are from the best possible performance, we organize a challenge involving 19 deep-learning systems implemented by 4 different research teams and 9 of the best expert botanists of the French flora. The main outcome of this work is that the performance of state-of-the-art deep learning models is now very close to the most advanced human expertise.

• Bird sounds identification: the 2018-the edition of the BirdCLEF challenge reveals impressive identification performance when considering bird sounds recorded by the Xeno-Canto community. Identifying birds in raw, multi-directional soundscapes, however, remains a very challenging task.

Towards the Recognition of The World's Flora: When HPC Meets Deep Learning

Participants : Hervé Goëau, Jean-Christophe Lombardo, Alexis Joly.

Automated identification of plants and animals have improved considerably in the last few years, in particular thanks to the recent advances in deep learning. In 2017, a challenge on 10,000 plant species (PlantCLEF) resulted in impressive performances with accuracy values reaching 90%. One of the most popular plant identification application, Pl@ntNet, nowadays works on 18K plant species. It accounts for million of users all over the world and already has a strong societal impact in several domains including education, landscape management and agriculture. Now, the big challenge is to train such systems at the scale of the world’s biodiversity. Therefore, we built a training set of about 12M images illustrating 300K species of plants. Training a convolutional neural network on such a large dataset can take up to several months on a single node equipped with four recent GPUs. Moreover, to select the best performing architecture and optimize the hyper-parameters, it is often necessary to train several of such networks. Overall, this becomes a highly intensive computational task that has to be distributed on large HPC infrastructures. Therefore, we experiment two french national supercomputers through an access offered by GENCI (Occigen@CINES, a 3.5 Pflop/s Tier-1 cluster based on Broadwell-14cores@2.6Ghz nodes and Joliot-Curie»@TGCC, a BULL-Sequana-X1000 cluster integrating 1656 nodes Intel Skylake8168-24cores@2.7GHz). To implement the synchronized stochastic gradient descent on the CPU cluster Joliot-Curie, we are using the deep learning framework Intel CAFFE coupled with Intel MLSL library (in the context of a collaboration with Intel).

Evaluation of Music Separation Techniques

Participants : Antoine Liutkus, Fabian-Robert Stöter.

After the groundbreaking advent of deep learning, we feel the music processing community needs to step back and think about what had been accomplished and what remains challenging in the problems of musical signal processing and filtering. Therefore, we give a complete overview of the state of the art in music demixing in [32] comprising more than 350 references, as well as two chapters in dedicated books [68], [67]. These references may be considered as complete overviews of the state of the art in music demixing. Furthermore, we introduce the topic to non-expert researchers and engineers in [26].

Apart from this effort in presenting the most recent advances in music processing to the community, we organize yearly a systematic evaluation of state of the art. We report the results of the 2018 Signal Separation Evaluation Campaign in [58], gathering a record number of participants. A perceptual evaluation of the results obtained through this campaign is presented in [59], in collaboration with researchers from the Surrey University.

Robust Probabilistic Models for Time-series

Participants : Antoine Liutkus, Fabian-Robert Stöter.

Processing large amounts of data for denoising or analysis comes with the need to devise models that are robust to outliers and that permit efficient inference. For this purpose, we advocate the use of non-Gaussian models for this purpose, which are less sensitive to data-uncertainty. Most of our effort on this topic is split in two subtasks.

First, we develop new filtering methods that go beyond least-squares estimation. In collaboration with researchers from RWTH, Aachen, Germany, we introduce a new model based on mixtures of Gaussians for filtering in [50]. It combines tractability with a better account of phase consistency for complex data. Along with researchers from IRISA, Rennes and Telecom ParisTech, we also work on filtering $\alpha$-stable processes [44], [46], [57], which enjoy important applications in robust signal processing.

Second, we work on large amounts of musical archives. This includes an original way to scale up interference reduction in live musical recordings in collaboration with the managers of the Montreux Jazz Festival data at EPFL (Switzerland).