The model wasn't the problem. The data was.

I thought training a CNN to classify cats and dogs would be straightforward. I was wrong.

The model plateaued at about 70% accuracy, and no amount of hyperparameter tuning helped. The real problem was the data: mislabelled images, class imbalance and noisy samples.

Fixing that taught me more than any textbook. It wasn’t a research breakthrough, but it was the kind of learning that sticks, and it came straight back when I worked on my dissertation.