CSAI Advance: Project Cycle, Evaluation and Neural Networks
What this domain tests
Advance CSAI moves into the full AI project cycle, data preparation, model building and evaluation, neural-network concepts, and deployment with ethical checks. Questions are often scenario-based: choose the right metric, spot a flawed pipeline, or explain why a model fails on new data.
Core rules to keep straight
- Project cycle: problem scoping → data acquisition → data exploration → modelling → evaluation → deployment.
- Confusion matrix terms: TP, FP, FN, TN. Precision = TP/(TP+FP); recall = TP/(TP+FN); F1 is the harmonic mean of precision and recall.
- Overfitting means low training error but high test error; remedies include more data, regularization and simpler models. Underfitting means high error on both.
- Preprocessing includes handling missing values, encoding categories and scaling features. Split the data before scaling to avoid leakage.
- A neural network has weights, biases, activation functions such as ReLU or sigmoid, a loss function and training epochs.
Worked example
A disease-screening model produces TP=30, FP=10, FN=5 and TN=55. Precision = 30/(30+10) = 0.75. Recall = 30/(30+5) ≈ 0.857. Accuracy = (30+55)/100 = 0.85. In screening, a false negative means a missed case, so recall is the priority; still report precision because false positives cause unnecessary follow-up. The reasoning is that each metric answers a different question about errors.
Traps that cost marks
- Reporting only accuracy on imbalanced data; a high score can hide poor minority-class detection.
- Swapping precision and recall formulas under exam pressure.
- Scaling or imputing the full dataset before the train-test split, which leaks test information.
- Assuming a model is unbiased because accuracy is high; bias can persist at high accuracy.
- Treating deployment as the end; monitoring, feedback and retraining are part of the cycle.