Adversarially-regularized mixed effects deep learning (ARMED)

A framework for building mixed effects neural networks for clustered and non-iid data, as reported in our paper:

Nguyen KP, Treacher A, Montillo A. "Adversarially-regularized mixed effects deep learning (ARMED) models improve interpretability, performance, and generalization on clustered (non-iid) data." in press at IEEE Transactions on Pattern Analysis and Machine Intelligence.

Our preprint is available at http://arxiv.org/abs/2202.11783v1.

Setup

Global directory paths should be editted in armed.settings:

1. RESULTSDIR: where experimental results will be stored
2. DATADIR: where downloaded and simulated datasets are stored

Add the repository root to the PYTHONPATH. If using Visual Studio Code, this can be done by modifying the .env file, which is read by the Python extension when running any code interactively.

Dependencies

See conda_environment.yml for Python dependencies. A new environment with these dependencies can be created using

conda env create -f conda_environment.yml --prefix </path/to/environment/location>

Table of contents

The main armed package contains the general-purpose tools for building ARMED models. The random effects layers can be found in armed.models.random_effects. Below are links to specific applications of ARMED models included in the above manuscript.

• Synthetic datasets: dense feedforward neural network applied to simulated spiral classification problems with random effects
• MCI conversion: dense feedforward neural network applied to classification of stable vs. progressive mild cognitive impairment
• AD diagnosis: convolutional neural network applied to classification of Alzheimer's Disease vs. cognitively normal subjects from T1w MRI
• Melanoma cell image compression and classification: convolutional autoencoder applied to compression of melanoma live cell images with simultaneous phenotype classification