I thought I’d spend a small amount of time talking about my 4th years honours. This isn’t going to be a big post like some of my others, just a few paragraphs talking about what it was and where you can read about it!
Overview
A Bayesian Neural Network differs from a regular one by providing a distribution of values instead of discrete values. This is primarily used to help overfitting but, interestingly, we can also thing of this as a method of predicting uncertainty. By running our network multiple times and calculating the entropy over the range of the values, we can get an idea of how unsure our model actually is. This is what I investigated as part of my 4th years honours, and I found some rather promising results.
Talk, paper and code
I’ve made my actual paper availible online, but if you’re more of a visual person then then you can also watch the presentation I made on YouTube. If you’re thinking about continuing this youself then feel free to adapt the code (Don’t judge, I was very young when I wrote this Python)