Question 23
Content Domain 2: Exploratory Data AnalysisA Machine Learning Specialist is required to build a supervised image-recognition model to identify a cat. The ML Specialist performs some tests and records the following results for a neural network-based image classifier. Total number of images available = 1,000 Test set images = 100 (constant test set). The ML Specialist notices that, in over 75% of the misclassified images, the cats were held upside down by their owners. Which techniques can be used by the ML Specialist to improve this specific test error?
Correct answer: A
Explanation
The model is missing a feature pattern because “over 75% of the misclassified images” show cats “held upside down,” so the error is caused by insufficient rotation variation in training. Adding rotated examples increases the training data diversity and helps the neural network learn rotational invariance, which reduces this specific test error.
Why each option is right or wrong
A. Increase the training data by adding variation in rotation for training images.
The error pattern is concentrated in a specific pose: more than 75% of the misclassified test images are upside down, which indicates the classifier has not learned rotational robustness from the training distribution. Under the standard supervised-learning remedy, the fix is to augment the training set with rotated images so the network sees that orientation is not a label-defining feature; with only 1,000 total images and a fixed 100-image test set, this directly addresses the observed generalization gap rather than changing the test set.
B. Increase the number of epochs for model training.
More epochs mainly repeat learning on the same data; missing rotation examples may still remain unlearned.
C. Increase the number of layers for the neural network.
Deeper networks increase capacity, but they do not automatically create invariance to unseen image rotations.
D. Increase the dropout rate for the second-to-last layer.
Higher dropout is regularization for overfitting, not a targeted fix for absent upside-down training patterns.