Lake Plankton

Deep Learning Landscape and Dynamics

We study plankton diversity in lakes, from monitoring to modeling. An underwater camera deployed by the Pomati group provides images of plankton with high throughput. We use deep learning models to identify the species, and monitor the populations as a function of time. We then model these time series with recurrent neural networks, use community detection algorithms to infer interactions, and fit the dynamics to models that mimic the interaction between well-mixed species.

The learning of deep neural networks is a dynamical process which consists of finding an optimum in a non-convex loss surface (we call it landscape). We investigate the interaction between landscape and dynamics, in order to better understand how the learning takes place and how we can improve it. For example, we look at how landscape and dynamics are influenced by data imbalance, and we investigate the differences in the learning between diverse kinds of dynamics. Some nice descriptions of this research line can be found in E.Francazi's webpage.

Predicting Toxicity

High-dimensional activated dynamics

In a collaboration with the group of K. Schirmer in the department of Environmental Toxicology at Eawag and the Swiss Data Science Center, we use machine learning methods to predict the effects of chemicals on aquatic species, from fish to crustaceans to algae. Our main goal is to use a combination of data from in-vivo (whole organisms) and in-vitro (cell culture) experiments to infer the effects of chemicals on organisms for which no testing data is available (both for the chemical and for the organism). You can find here a nice blogpost by Christoph Schuer on our research.

Activated dynamics is a very slow process that takes place on exponentially large time scales. Usually it is associated to barrier hopping. However, activation can also be driven by entropy, which completely changes the way we should think of relaxation in complex systems. We study the emergence of entropically activated dynamics in several systems, from toy to machine learning models.

Invasions in blue-green ecosystems

Community Detection

We compare invasions in aquatic and terrestrial ecosystems primarily at large (national) spatial scales and among several higher-level taxa (insects, molluscs, crustaceans, all major vertebrate classes, and plants). We calculate native-exotic richness ratios and test the biotic resistance hypothesis while simultaneously accounting for other potential factors that could influence invasiveness and invasibility (i.e., species traits, propagule pressure, and the size of species pools).

Community detection consists of extracting the affinity between agents of a system, which is extracted from quantities such as the frequency of interactions. When analyzing datasets, however, the absence of a connection between agents might not be due to a lack of affinity, but rather to the fact that these agents never met. For example, Alex and Sam would like each other a lot, if they only met. We introduce exposure into community detection, as an additional mechanism to explain the lack of links among agents.