One of the issues is that if you try to prove generalization bounds for anything that has an universal representation property (e.g. neural networks) you always get terrible bounds that don't reflect empirical performance on real-world learning tasks.

I think this is because there are probably some learning tasks which are intrinsically intractable (this is certainly the case if one-way functions exist), therefore any bound that quantifies over all learning tasks will be at least exponential.