PhD: Aligning Deep Learning with Ecological Concepts — Explainable Methods for Biodiversity Modelling

PhD at the Inria centre of Université Grenoble Alpes (Thoth team) and the Laboratoire d’Écologie Alpine (LECA), supervised by Karteek Alahari (thesis director), Hadrien Hendrikx and Sara Si-Moussi.

Motivation

New technologies now collect massive amounts of biodiversity data — camera traps, acoustic recorders, aerial and satellite imagery, environmental DNA, vegetation surveys, in-situ sensors. Deep learning improves some predictions, such as species distributions, but explains little about the underlying ecological processes: latent dimensions are rarely tied to ecologically meaningful concepts. In ecology, however, the central question is often not prediction but understanding interactions — which components of the environment influence which traits of organisms, and at which scales?

Guiding question: how can we infer ecological interactions from heterogeneous data represented by deep features, i.e. latent representations learned by deep models?

Research directions

1. Interpretability of deep models for ecological data.
   • Emerging concepts in unimodal foundation models: probing self-supervised models of aerial imagery, bioacoustics and other ecological modalities to test whether ecologically relevant concepts (habitat structure, biophony vs. anthropogenic noise, species compatibility…) emerge, and whether they can be separated from observation conditions and spatio-temporal context.
   • Concepts used by multimodal biodiversity models: concept bottleneck models and sparse-autoencoder concept dictionaries, e.g. using species functional traits as intermediate concepts in species distribution models.
2. Intrinsically interpretable representation learning. Co-learning latent spaces and concepts through concept-based regularisation and explainable pretraining objectives, and studying the effect on interpretability, downstream performance, robustness and generalisation.
3. Evaluation and ecological inference. Defining metrics suited to ecological applications, and integrating the discovered concepts into hierarchical probabilistic graphical models of environment–species interactions across scales.

Keywords: representation learning, concept learning, explainability, self-supervised learning, species distribution modelling, multimodal data (images, sound, text).