Deep Learning (Data Scientist) 4 — Questions and Answers
Question 1: A data scientist deploys a fraud-detection network trained on 2022 data and observes accuracy decay in production during 2024. What is this phenomenon called?
- Data drift (distribution shift) (Correct answer)
- Gradient vanishing
- Catastrophic forgetting
- Mode collapse
Correct answer: Data drift (distribution shift)
When the production data distribution diverges from the training distribution over time, model performance degrades due to data drift.
Question 2: In a GAN, the generator produces nearly identical outputs regardless of the input noise vector. What is this failure mode called?
- Mode collapse (Correct answer)
- Overfitting
- Vanishing gradients
- Label leakage
Correct answer: Mode collapse
Mode collapse occurs when the generator maps diverse noise inputs to a narrow set of outputs, ignoring the data's variety.
Question 3: Which practice most directly reduces the memory footprint and latency of a trained network for edge deployment?
- Quantization of weights to lower-precision formats such as INT8 (Correct answer)
- Increasing the batch size at inference
- Adding more dropout layers
- Training with a larger learning rate
Correct answer: Quantization of weights to lower-precision formats such as INT8
Quantization stores and computes weights in lower precision, shrinking model size and speeding up inference on edge hardware.
Question 4: A team wants uncertainty estimates from a deployed classifier. Which approach provides an approximate Bayesian estimate at inference time?
- Monte Carlo dropout, keeping dropout active during multiple forward passes (Correct answer)
- Using a lower softmax temperature
- Removing the final activation function
- Averaging the training and validation losses
Correct answer: Monte Carlo dropout, keeping dropout active during multiple forward passes
Running several stochastic forward passes with dropout enabled approximates sampling from a posterior, yielding predictive uncertainty.
Question 5: Why is ReLU generally preferred over sigmoid for hidden layers in deep networks?
- ReLU avoids saturation for positive inputs, preserving gradient magnitude in deep stacks (Correct answer)
- ReLU outputs are bounded between 0 and 1
- ReLU has more trainable parameters
- ReLU guarantees convergence to a global minimum
Correct answer: ReLU avoids saturation for positive inputs, preserving gradient magnitude in deep stacks
Sigmoid saturates at both extremes and shrinks gradients, while ReLU's linear positive region keeps gradients flowing in deep networks.
Question 6: During hyperparameter tuning, a data scientist evaluates on the test set after every change and picks the best configuration. What is the main methodological flaw?
- Test-set leakage: the test set no longer provides an unbiased estimate of generalization (Correct answer)
- The learning rate was never tuned
- Too few hyperparameters were searched
- The training set was too large
Correct answer: Test-set leakage: the test set no longer provides an unbiased estimate of generalization
Repeatedly selecting on the test set overfits hyperparameters to it, invalidating it as a final unbiased benchmark.
Question 7: Which scenario is the strongest candidate for a 1D convolutional network rather than a 2D one?
- Classifying anomalies in a univariate sensor time series (Correct answer)
- Segmenting objects in satellite photographs
- Detecting faces in video frames
- Colorizing grayscale images
Correct answer: Classifying anomalies in a univariate sensor time series
1D convolutions slide along a single temporal axis, matching the structure of sequential sensor data.
Deep Learning (Data Scientist) 4
A data scientist deploys a fraud-detection network trained on 2022 data and observes accuracy decay in production during 2024.
What is this phenomenon called?