MLS-C01 Question #78: Real Exam Question with Answer & Explanation

Question

During mini-batch training of a neural network for a classification problem, a Data Scientist notices that training accuracy oscillates. What is the MOST likely cause of this issue?

Options

• AThe class distribution in the dataset is imbalanced.
• BDataset shuffling is disabled.
• CThe batch size is too big.
• DThe learning rate is very high.

Explanation

{"question_number": 9, "correct_answer": "D", "explanation": "Oscillating training accuracy during mini-batch gradient descent is a classic symptom of a learning rate that is too high. With an excessively large learning rate, the parameter updates overshoot the loss function's minimum on each step, causing the model to bounce back and forth around the optimum rather than converging. Option A (class imbalance) would manifest as consistently poor accuracy on minority classes, not oscillation. Option B (disabled shuffling) can cause batch-to-batch variance but typically shows as slow convergence or bias toward recent batches, not sharp oscillation. Option C (large batch size) actually has the opposite effect - larger batches produce more stable gradient estimates, which reduces oscillation.", "generated_by": "claude-sonnet", "llm_judge_score": 4}

No community discussion yet for this question.