This repository is a cleaned version of my research on the vanishing gradient problem with dead ReLUs. My research suggests an approach to combat dead ReLUs with residual connections called Late Residual Connections. These connections are added directly to the output of a neural network from each hidden layer. Please feel free to use the code and add on to the research! I can be reached here for any questions.

Below is an example of an abstracted view of traditional versus late residual neural networks. Each hidden layer has a late residual connection the output. To simplify, this connection allows for moregradient flow to each neuron in the network. My full thesis can be read here.

run_experiment_script.py is a script that can be used to run the experiments. Below explains how to create an environment and run the script.

You will need a Python version of 3.8 or higher to run the code in this repository. Run the following commands below to create your environment and install the required packages.

python3 -m venv env
source env/bin/activate
python3 -m pip install --upgrade pip
python3 -m pip install -r requirements.txt

To run the script, run run_experiment_script.py. This script looks at the environment variables file and runs the experiments based on the variables in the file. Change these to try different experiments. Below is a list of the variables and what they do.

- width: The widths of the networks.
- depths: The depths of the networks.
- learning_rates: The learning rates of the networks.
- optimizers: The optimizers of the networks.
- epochs: The number of epochs to train each network.

The code in this repository can also be run in a Docker container to avoid any issues with setting up the environment. To run the code in a Docker container, simple run the following command.

docker compose up

A prebuilt image can also be pulled from Docker Hub or GitHub Container Registry.

docker pull mfernst/late-residual-connections # Docker Hub
docker pull ghcr.io/matthewfernst/lateresidualconnections:main # GitHub Container Registry