Most of the AutoML examples I see online are for simple CSV files. Can automated machine learning effectively handle Neural Architecture Search for computer vision tasks? I want to build a custom classifier for medical imaging but I’m not an expert in designing deep CNN architectures from scratch.
3 answers
Yes, this is often referred to as Auto-DeepLearning or NAS (Neural Architecture Search). While automated machine learning for images is much more computationally expensive than for tabular data, tools like AutoKeras make it very accessible. It will automatically search for the best number of layers, filter sizes, and dropout rates for your specific image set. However, for medical imaging, I would suggest a hybrid approach. Start with a pre-trained model like ResNet or EfficientNet and use AutoML specifically for the fine-tuning stage. This is usually much faster and yields better results than letting an algorithm build a network from zero.
Have you considered the hardware requirements? Running a full architecture search for automated machine learning on images usually requires multiple high-end GPUs.
Google’s Vertex AI has a very powerful vision-based automated machine learning tool, but it can get quite pricey if you aren't careful with your budget.
Peter is right. The "pay-as-you-go" model can surprise you with a massive bill if the architecture search takes longer than expected.
We have access to a small cluster at the university, so I’m hoping we can let the search run over a weekend without hitting a wall.