Analyze vision model failures

For a computer vision product, address the points below:

1. Describe the machine learning fundamentals most relevant to vision work: bias versus variance, overfitting, class imbalance, evaluation metrics, threshold selection, calibration, and data leakage.
2. Work through this probability problem: in a photo stream, 5% of images are actually blurry. A blur detector catches 90% of blurry images and falsely marks 10% of sharp images as blurry. Given that an image is flagged, what is the chance it is truly blurry?
3. After launch, a production vision model begins to perform worse. Outline a systematic approach for identifying root causes and choosing fixes. Include data drift, labeling quality, train-serving skew, preprocessing bugs, feature issues, model capacity, threshold problems, and offline-online metric mismatch. Overview: This question assesses understanding of core computer vision ML concepts (bias–variance tradeoff, overfitting, class imbalance, evaluation metrics, threshold selection, calibration, and data leakage), probabilistic reasoning for interpreting performance, and diagnosing and root-causing production model issues.