Comment by anon291
2 years ago
Let's restate. I'd imagine you end up in local minima that are difficult to escape using model generated data. So sure, non-zero gradients, but if you plot the gradients, I would expect them to orbit at that point. But it seems like they diverge.
Mini-batches and dropout mean that you are constantly jumping out of and into other minima during training of any type (highly-redundant solution space is an important feature of deep learning). This is deliberate and necessary to explore the gigantic parameter space of these huge LLM models.
Sure, but one might think that training on self-generated data would keep you in a constrained subset of minima, but that is not the case.