AI Building AI

Ai building AI

I was building an image classifier for another country and had got to 70% validation accuracy and hit a plateau. So I thought I would try and make use of an AI to help me improve the AI I was building.

For 2 days I took guidance from Claude Code on how I could amend my AI training code to improve the overall accuracy of my model. This included getting me change the image preprocessor to something that was more suited to the Base Model I was using. Changing the image resolution I was using for the model, unfreezing layers of the base model, amending the model I was using to a smaller one for my small range of input data. These are all fairly large changes to a model. And I listened and amended the code accordingly.

After these many changes (with some back tracking and experiments) with the help of Claude code my validation accuracy dropped from 70% to just over 50%. Which is just as good as tossing a coin. The irony of an AI totally destroying another one is not lost on me. But at least it did it with the confidence of an experienced ML engineer.

I have gone back to my normal approach to these things. Small experimental changes to the hyper-parameters with small experimental runs with a small number of epochs. And then when one of those runs shows promise try it with a few more epochs and a little more patience.

It was worth the try but I can'l help think I have lost 2 days of work.