Projects

How can ecological interactions be inferred from the latent representations of deep models trained on heterogeneous biodiversity data? Interpretable, concept-aligned representations of multimodal ecological data — from post-hoc concept discovery in foundation models to representations that are interpretable by design.

Combining oscillatory neurons (AKOrN) with a hippocampus-inspired self-supervised objective (PhiNet) to learn encoders that resist adversarial perturbations without adversarial training. Accepted at the WiML workshop @ NeurIPS 2025; extended in a 2026 arXiv preprint.

Passive imaging: recovering the Green’s function between two sensors by cross-correlating ambient noise signals, and studying how limited noise-source illumination in 3D degrades the estimate.

Data challenge: classifying DNA sequences with string kernels implemented from scratch (spectrum, mismatch, gapped k-mer), combined through kernel–target alignment and plugged into an SVM solved with OSQP.

Re-implementation of PointNet for point cloud classification and segmentation, and application to an RNA dataset.

A neural network learns summary statistics of DNA sequences, which feed an adaptive sequential Monte Carlo ABC with random forests to infer the parameters of finite-sites evolutionary models fitted to cancer data.

Comparing logistic regression, random forests, XGBoost, LSTMs and CNNs to predict rainfall from daily weather observations, then simplifying the model through interpretability analysis.

Synthesis of ZnO and CdS nanoparticles and photocatalytic reactions to produce hydrogen from waste.

How the solution of a wave-propagation PDE depends on the geometry of the domain, and under which conditions reflection vanishes or transmission becomes perfect.