Question 890
A deep-learning script accumulates gradients over four microbatches, but clips each microbatch gradient separately and produces different results from the intended global batch. Where should clipping occur?
Explanation
Correct Answer
Clipping the accumulated update controls the gradient actually applied and preserves the chosen effective-batch semantics.
Incorrect Answers
Inference handlers are unrelated to training optimizer steps.
No gradient exists before training examples are processed.
A checkpoint occurs after many updates and cannot constrain prior gradients.