Bayesnets¶

Welcome¶

Bayesnets is a mini-research project where we seek to look into neural nets by limiting scope in order to understand the deeper truths of neural networks.

For this project, we’re mostly limiting ourselves to

looking at the behaviour of many networks with only starting conditions affected
simple 1d functions
fully connected networks
visualisations (with a heavy focus on animations) to drive intuition.

Swarm¶

swarm is the python package of this repo you will see imported in the experiments. It contains utils to do batches of training and animation as well as things like setting rng seeds. Browse the source code here.

Terminology¶

We use the below terminology for experiments/code.

Bee : A neural net training sequence
Swarm : A group of networks trained the same way, with the only difference defined by starting conditions
Hive : A set of swarms with some some training/initialisation parameter varied.

About Martinis and Research¶

A group of three sydney siders who have too much time on their hands, especially during COVID and wanted to justify their love of alcohol.