Neural Networks

Topics in Econometrics and Statistics

Summer 2022, Tim Mensinger

This repository contains (source) material for

1. A custom implementation of a vanilla MLP neural network using JAX. The network is trained on the MNIST data set to recognize hand-written digits.
2. Two presentations on the theoretical and practical side of neural networks. The theoretical version was presented in class.
3. A term paper that analyzes neural networks from an econometric perspective.

Paper

The tex-files are stored in the folder [root]/paper and compiled using (a relatively fresh install of) latex. You can view it here

• 🗎 Paper

The results in the paper (tables and figures) are generated in a reproducible fashion using pytask. The build process is described below in the section Code.

Presentation

There is a theoretical and a practical presentation. You can view them here

• 🗎 Theoretical
• 🗎 Practical

The source code is stored in the folder [root]/presentation and rendered using marp.

Code

Installation

To get started, create and activate the (conda) environment using

bash $ cd 'into project root' $ conda env create -f environment.yml $ conda activate nnet

This installs the package dependencies and the project itself. You can then import functions defined in the repository as

python from nnet.network import build_network from nnet.plotting import plot_examples

Reproduce Paper

To reproduce the results from the paper (tables and figures) you only have to call pytask in an active environment

bash $ cd 'into project root' $ conda activate nnet $ pytask

This creates a folder [root]/bld which contains all tables and figures used in the paper in the subfolder paper.

Custom Neural Network Implementation

For this project I implemented a multilayer-perceptron neural network and its fitting procedure using Jax. The code is found in the module [root]/src/nnet/network.py. It works in both regression and classification settings and is tested on the MNIST dataset that contains hand-written digits.

Warning Jax is not supported on windows.

If you want to run my custom implementation locally I suggest running the notebook [root]/src/nnet/neural_network.ipynb ---or simply view it on nbviewer.

Credits

This project was created with cookiecutter and the cookiecutter-pytask-project template.