This week, we started exploring machine learning with Diana. Initially, I was a bit worried because we were told it was highly technical and not very creative. However, once we started working with the Fashion MNIST dataset and building our own model, I found it surprisingly interesting. The code behind it wasn’t as complicated as I had expected, and seeing the model in action helped me understand the process better.
One of the most confusing parts was understanding neural networks, but working through an example with real data made it much easier to grasp. I found it especially helpful when Diana introduced us to the ResNet18 model, which simplifies some of the more complex steps and produces better results, particularly when using our own dataset.
Another key takeaway was learning how to improve model performance by applying different data transformations. Even though this actually made our model’s accuracy worse in some cases, it was still useful knowledge for refining models in the future. What I really liked about this topic is that once you understand the core code, most of it remains the same. For example, when we switched from Fashion MNIST to our own dataset, there were only a few changes needed to adapt the model.
What I enjoyed the most this week was the maths side of machine learning. Maths has always been something I enjoy, though it was a bit challenging when we started working with simultaneous equations, especially since I hadn’t done them since my GCSEs. However, once I understood Diana’s approach, it started to make sense again.
Looking ahead, I think I’ll enjoy Task 1 more than Task 2 in the brief, but that could change depending on how I approach Task 2. If I can find a way to incorporate something I’m passionate about, such as cats, into my custom image classification, I think I’ll find it much more engaging. I might explore ways to train a model to recognize different breeds or identify cat-related images, which would make the project feel more creative and personal.
Overall, this week’s introduction to machine learning was much better than I anticipated, and I’m excited to see how I can apply these concepts in the coming weeks.