AIF-C01 Practice Question 52: Deep Learning Feature Learning
AWS Certified AI Practitioner (AIF-C01) · Test 6 · question 2 of 10
An engineering team classifies photographs of manufactured parts as defective or acceptable. Their current approach requires a data scientist to hand-craft numeric features (edge counts, contrast ratios, blob sizes) before a classical algorithm can use them, and accuracy has stalled. A colleague proposes a deep learning model instead. Which characteristic of deep learning makes it the better fit for this problem?
Show answer and explanation
Answer:
Deep learning neural networks comprise many layers of artificial neurons that work together to solve complex problems. Because each layer learns progressively richer representations from the layer below, deep learning models can comprehend unstructured data and make general observations without manual feature extraction. Images are exactly the unstructured data that classical methods find hard to process, which is why the hand-crafted features were the bottleneck (option B).
- A is wrong, and backwards. Deep learning generally needs *more* labelled data and far more compute than a classical algorithm on engineered features. That cost is the trade-off you accept in exchange for dropping the feature engineering. - C is wrong because "deep" describes the architecture, not the learning paradigm. A defect classifier trained on photos labelled defective or acceptable is supervised learning; the labels are still required. - D is wrong, and again backwards. Deep networks are the classic "black box": a hand-crafted feature set is usually *easier* to reason about. Explaining a deep model's individual predictions needs a separate explainability technique.
Reference: What is Deep Learning?