https://github.com/mbjoseph/neuralecology

Code for the paper "Neural hierarchical models of ecological populations"
https://github.com/mbjoseph/neuralecology

bayesian deep-learning ecology hierarchical-model occupancy research-compendium science-based-deep-learning

Last synced: 2 months ago
JSON representation

Code for the paper "Neural hierarchical models of ecological populations"

• Host: GitHub
• URL: https://github.com/mbjoseph/neuralecology
• Owner: mbjoseph
• License: bsd-3-clause
• Created: 2019-09-05T18:58:19.000Z (over 5 years ago)
• Default Branch: main
• Last Pushed: 2022-11-06T21:57:55.000Z (about 2 years ago)
• Last Synced: 2024-06-11T17:06:18.991Z (7 months ago)
• Topics: bayesian, deep-learning, ecology, hierarchical-model, occupancy, research-compendium, science-based-deep-learning
• Language: Jupyter Notebook
• Homepage:
• Size: 13.3 MB
• Stars: 23
• Watchers: 2
• Forks: 10
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# Neural hierarchical models of ecological populations

[![DOI](https://zenodo.org/badge/206636551.svg)](https://zenodo.org/badge/latestdoi/206636551)

Paper: https://onlinelibrary.wiley.com/doi/full/10.1111/ele.13462  

Preprint: https://www.biorxiv.org/content/10.1101/759944v3

### Key idea

Parameterize a hierarchical model (an observation + process + parameter model) with a neural network, creating a **neural hierarchical model**.

![Alt text](./fig/fig2.svg)

Here, (a) shows linear regression, mapping input x to an output y. 

In (b) a neural network inserts hidden layers between x and y.

Analogously, an ecological model (c) maps an input x to parameters of a hierarchical model.

A neural version of model (d) would similarly involve hidden layers between x and these parameters. 

Deep models (e) can also be constructed that use more complex neural architectures, especially when data are structured in time, space, and/or over networks. 

A variety of neural network components can be readily used in neural hierarchical models. 

For example, you might parameterize a hidden Markov model of animal movement using a convolutional neural network that takes remotely sensed imagery as input (see Appendix S2 for details). 

![convHMM](fig/conv_hmm_edited.png)

## Hardware requirements

- 20+ GB of RAM

- 4 or more CPU cores

- GPU recommended

## Setting up the environment

This project uses [conda](https://docs.conda.io/en/latest/) to install python dependencies.

```

conda env create -f environment.yml

```

Once installed, activate the environment via:

```

conda activate neural-ecology

```

To install R dependencies:

```

R -e "devtools::install_deps(dependencies = TRUE)"

```

## Running the toy models

[![Binder](https://mybinder.org/badge_logo.svg)](https://mybinder.org/v2/gh/mbjoseph/neuralecology/master)

The `notebooks/` subdirectory contains toy models in Jupyter notebooks:

- [A neural occupancy model](notebooks/simple-occupancy.ipynb)

- [A neural dynamic occupancy model](notebooks/dynamic-occupancy.ipynb)

- [A neural N-mixture model](notebooks/n-mixture-model.ipynb)

- [A deep Markov model for capture-recapture data](notebooks/hidden-markov-model.ipynb)

## Building the paper

The workflow for building the paper is handled with GNU Make. 

To build the paper (including running the models for the case study) takes ~ 5 hours with 6 CPU cores and a GPU. 

```

make

```