I've had a lot of experience with dealing with exactly this problem in ML.
There is no replacement for visualizing the data and intermediates. You have to see what is going on, it's the only way to catch bugs and issues and make good performance improvements.
Like actually looking at the gradients and weights and activations and predictions and stuff often shows you that something janky is going on, its great for adjusting architectures. Looking into specific high loss test samples and mispredictioms and stuff will show you that there are problems with your data or normalization and whatnot.
The issue is that there are basically an infinite number of intermediates you could potentially look at. So ignoring almost all of them is the only thing that scales and you have to be extremely deliberate.
There is no replacement for visualizing the data and intermediates. You have to see what is going on, it's the only way to catch bugs and issues and make good performance improvements.
Like actually looking at the gradients and weights and activations and predictions and stuff often shows you that something janky is going on, its great for adjusting architectures. Looking into specific high loss test samples and mispredictioms and stuff will show you that there are problems with your data or normalization and whatnot.
The issue is that there are basically an infinite number of intermediates you could potentially look at. So ignoring almost all of them is the only thing that scales and you have to be extremely deliberate.