AI0-001 AI Implementation and Operations Practice Question

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  Add more training data.

  Why it's wrong here

  Adding training data increases dataset size and compute load, worsening rather than relieving memory pressure during training. More data is the right remedy when the model underfits or lacks coverage, not when the exhibit shows memory exhaustion during execution.

• ✗

  Increase the learning rate.

  Why it's wrong here

  Raising the learning rate changes gradient step size, not memory footprint; larger steps can even worsen instability without freeing a single byte. It is tempting because learning rate is the usual first lever for convergence problems, and it would be the right adjustment when training loss plateaus or diverges rather than when memory is exhausted.

• ✗

  Switch to a CPU.

  Why it's wrong here

  Switching to CPU changes the compute device, not memory demand; the model and its tensors still occupy the same memory, so the issue persists, only slower. CPU execution suits small models or inference where GPU acceleration is unnecessary, not memory-constrained training.

• ✓

  Reduce the batch size.

  Why this is correct

  Reducing batch size lowers peak activation memory because fewer samples are held simultaneously during the forward and backward passes. This directly addresses the memory exhaustion constraint shown in the exhibit without altering model architecture or precision.