Machine Learning 201 - Neural Networks and Image Recognition

In this session you will learn to train a Machine Learning model to do Image Classification. One of the original classic problems that machine learning was trying to solve for decades was how to tell a dog from a cat in pictures – something that even a small child could do but computers found it very difficult. After many decades that problem was solved and has paved the way for ML to now be good at reading radiology images, identifying faces, identifying object types for self-driving cars, identifying deforestation from satellite images, and all sorts of other use cases. We will learn how this is done in this hands-on session. In particular, we will work through the problem of identifying handwritten digits. We will build successively more sophisticated models to improve the accuracy of this task including Logistic Regression, a Feed Forward Neural Network, and a Convolutional Neural Network.

This is a 2 hour session, which was recorded live via a virtual session with a few attendees. No ML or python experience is required, but comfort with coding would be useful.

The links that you will need as part of this class are:
Notebooks at this link:

Enjoy!

[chrisstewart]

Great video! One thing that may trip up anyone following along is that you have to phone verify to get access to the setting that allows the environment to access the Internet. It's not essential but you would have to manually download the data file and import it to kaggle, then amend the code to access from there instead

[AngriestEwok]

If you're following along in 2024, the data url in the attached notebook doesn't work any more because it's been removed from Don's website. Not to worry though, you can simply click 'add input' and select 'MNIST in CSV'' - update the url to the location of the file instead and it works just fine.