Embedded Files
Towards Explainable Neural Network Models of Mosquito Abundance
Towards Explainable Neural Network Models of Mosquito Abundance
Advisor: Dr. Joceline Lega
Vector-borne disease outbreaks are closely tied to vector abundance, which makes knowledge of population dynamics useful in preventing future outbreaks. My dissertation work focuses on developing human-understandable neural network models of Aedes aegypti populations.
The video to the right introduces the Aedes-AI project, and below are descriptions of the model development and model interpretability stages.
Video 1.
I participated in the University of Arizona Graduate Center's Grad Slam in 2021. The campus-wide competition involves creating 3-minute presentation of research projects for a general audience.
Model Development
Model Development
Collaborators: Dr. Joceline Lega, Sean Current
Aedes-AI is a collection of neural network models of Aedes aegypti mosquito abundance. The neural networks are trained on synthetic data generated from a mechanistic model, in contrast to other models of mosquito abundance that rely on noisy, real world trap data for training. We show the networks predict mosquito abundance with high skill, and assess the impact of model architecture and data oversampling.
Citation: [1] Kinney, Adrienne C., Sean Current, and Joceline Lega. "Aedes-AI: Neural network models of mosquito abundance." PLoS computational biology 17.11 (2021): e1009467.
Figure 1.
Architecture of the Aedes-AI models. Figure reproduced from [1].
Figure 2.
Average performance metrics for the testing locations show high skill. Figure reproduced from [1].
Figure 3.
Abundance predictions generated from an Aedes-AI model (GRU HI) and a mechanistic model (MoLS). Figure reproduced from [1].
Model Interpretability
Model Interpretability
Collaborators: Dr. Joceline Lega
In progress.
The Aedes-AI framework presents an opportunity to study and advance interpretability methods because the models are trained on data generated from known dynamics. We focus on sensitivity to input and understanding the behavior of specific network components.
GitHub: https://github.com/akinney1/aedesAI_interpretability
Figure 4.
Model skill as a function of the number of filters in the convolutional layers, Figure 1. These results guide model reduction.
Figure 6.
Model sensitivity to input data. We see the model is most sensitive to temperature nearest the prediction point (top right box of each plot).
Figure 7.
Visualizing the forward pass of input samples through ConvLayer 1, Figure 1. We see the convolutional filters activate for different portions of the humidity, temperature plane.
Page updated
Report abuse