Moreover, we can use explainable AI tools to understand the learned functional form of the replaced process. We demonstrated this using the Barro Colorado Island plot by replacing the growth process with a DNN. We found plausible dbh-growth and light-growth functions learned by the hybrid model 4/4

We introduce forest-informed neural networks (FINNs), a new DVM in which processes can be replaced by deep neural networks and the entire model is calibrated jointly. FINN can approximate the functional shapes of otherwise misspecified processes and achieve better predictive accuracy 3/4

DVM need precise functional forms but determining the correct functional form can be challenging. An automatic approach to this problem, such as DNNs, is compelling. However, previous work has shown that plug-in estimators of processes don’t work well, and joint calibration is necessary 2/4

In our new preprint, “Inferring processes within dynamic forest models using hybrid modeling” @ykaber.bsky.social and I present a new hybrid modeling approach for jointly calibrating a DVM with embedded #deepneuralnetwork arxiv.org/abs/2508.01228 1/4
#deeplearning #forestdynamics

In Nature, Delavaux et al., 2024 suggest that species richness on oceanic islands is reduced by mutualism rates on the mainland. In arxiv.org/abs/2411.15105, @maximilianpichler.bsky.social and I had another look at this question and conclude that this effect is likely a statistical artefact.

go.bsky.app/NiaWN5i

I'm happy to receive suggestions for this.

For more details and explanations on how to train and interpret DNNs, see our extensive documentation (including #SDM and #MSDM examples) that also covers advanced topics such as custom loss functions and residual checks (under articles on citoverse.github.io/cito/)!

cito can now train DNNs for count data using Poisson or negative binomial distributions. In addition, deep joint species distribution models (#jsdm #sdm) based on the multivariate probit model can be fitted:

An important new feature is hyperparameter tuning under cross-validation, which helps to train the #DNN. Hyperparameter tuning can be easily done by passing a "tune(...)" to the hyperparameters (cito also automatically returns the model with the best hyperparameters):

cito v1.1 #rstats package for deep neural networks (#DL #DNN) (with formula syntax) is now available on #CRAN. New features include likelihoods such as the negative binomial distribution and easy hyperparameter tuning: cran.r-project.org/web/packages...

Come work with us - I am looking to fill a 3-yr position for a statistical postdoc / scientific programmer to continue the development of the DHARMa #Rstats #CRAN package for #glmm residual diagnostics. Full job advertisement is here uni-regensburg.de/assets/biolo...