01 - AI and machine learning fundamentals¶
How the terms nest¶
Artificial intelligence is the broad field of systems performing tasks that would require human intelligence. Machine learning is a subset where systems learn patterns from data rather than following explicitly programmed rules. Deep learning is a subset of machine learning using multi-layer neural networks. Generative AI is an application of deep learning that produces new content.
Task categories Oracle uses: language, speech, vision, and decision.
Types of learning¶
| Type | Data | Learns | Examples |
|---|---|---|---|
| Supervised | Labelled | To predict a label | Classification, regression |
| Unsupervised | Unlabeled | Structure in the data | Clustering, dimensionality reduction, anomaly detection |
| Reinforcement | Feedback from an environment | A policy maximizing reward | Game playing, robotics, control |
Supervised splits into: - Classification: predict a category. Spam or not spam, which of five defect types, will the customer churn - Regression: predict a continuous value. House price, tomorrow's demand, time to failure
The tell in a question: if the answer is a number on a scale, it is regression; if it is one of a fixed set of labels, it is classification.
Unsupervised includes: - Clustering (k-means): group similar items with no predefined groups - Dimensionality reduction (PCA): reduce features while retaining information - Anomaly detection: identify points that do not fit the learned pattern
The machine learning workflow¶
- Problem definition - what are we predicting, and what does success mean
- Data collection
- Data preparation - cleaning, handling missing values, encoding, normalizing
- Feature engineering - creating the inputs the model actually learns from, often the highest-leverage step
- Split - training, validation, and test sets
- Model selection and training
- Evaluation on held-out data
- Deployment
- Monitoring - watching for drift, where the live data diverges from the training data and performance degrades
Why three splits: the training set fits the model, the validation set tunes hyperparameters, and the test set is held back untouched so the final estimate is honest. Tuning against the test set contaminates it.
Overfitting and underfitting¶
| Training performance | Test performance | Cause | Fix | |
|---|---|---|---|---|
| Overfitting | High | Low | Model memorized the training data, including its noise | More data, regularization, simpler model, early stopping, cross-validation |
| Underfitting | Low | Low | Model too simple to capture the pattern | More complex model, better features, train longer |
| Good fit | High | High | - | - |
The bias-variance trade-off describes this: high bias means the model is too simple (underfitting), high variance means it is too sensitive to the training data (overfitting).
Evaluation metrics¶
For classification, start with the confusion matrix:
| Predicted positive | Predicted negative | |
|---|---|---|
| Actually positive | True positive (TP) | False negative (FN) |
| Actually negative | False positive (FP) | True negative (TN) |
| Metric | Formula | Answers |
|---|---|---|
| Accuracy | (TP + TN) / all | What proportion did I get right overall |
| Precision | TP / (TP + FP) | Of what I flagged, how much was correct |
| Recall (sensitivity) | TP / (TP + FN) | Of what I should have found, how much did I find |
| F1 | Harmonic mean of precision and recall | A single balanced score |
Accuracy misleads on imbalanced data. A fraud model on a dataset that is 99.9% legitimate can predict "not fraud" every time and score 99.9%, while catching nothing.
Choose by the cost of the error: recall when a miss is expensive (disease screening, fraud), precision when a false alarm is expensive (blocking legitimate email).
For regression: mean absolute error (MAE), mean squared error (MSE), root mean squared error (RMSE), and R-squared. MSE and RMSE penalize large errors more heavily.
Common algorithms¶
| Algorithm | Type | Used for |
|---|---|---|
| Linear regression | Supervised | Predicting a continuous value |
| Logistic regression | Supervised | Binary classification, despite the name |
| Decision tree | Supervised | Classification and regression, highly interpretable |
| Random forest | Supervised | Many trees combined, more accurate and less prone to overfitting |
| Support vector machine | Supervised | Classification with a clear margin between classes |
| k-nearest neighbors | Supervised | Classification by similarity to nearby examples |
| k-means | Unsupervised | Clustering into k groups |
| PCA | Unsupervised | Dimensionality reduction |
Responsible AI¶
Oracle's framing, and a source of exam questions:
- Fairness - the system does not produce discriminatory outcomes across groups
- Transparency - users know they are interacting with AI and understand its role
- Explainability - decisions can be explained to the people they affect
- Accountability - a human is responsible for outcomes
- Privacy - personal data is protected throughout the lifecycle
- Robustness and safety - the system behaves reliably, including under unexpected input
Bias enters through the data (historical bias reflected in training examples), the sampling (unrepresentative data), or the labeling (subjective human judgement).
Key terms¶
- Artificial intelligence - systems performing tasks that would otherwise require human intelligence
- Machine learning - systems that learn patterns from data rather than following explicit rules
- Deep learning - machine learning using multi-layer neural networks
- Supervised learning - learning from labeled data to predict a label
- Unsupervised learning - finding structure in unlabeled data
- Reinforcement learning - learning a policy from reward feedback in an environment
- Classification - predicting a discrete category
- Regression - predicting a continuous numeric value
- Clustering - grouping similar items without predefined labels
- Feature engineering - creating the input variables a model learns from
- Training set - the data used to fit the model
- Validation set - the data used to tune hyperparameters
- Test set - held-out data used once for an honest performance estimate
- Overfitting - a model that memorizes training data and generalizes poorly
- Underfitting - a model too simple to capture the underlying pattern
- Bias-variance trade-off - the tension between a model being too simple and too sensitive to its training data
- Confusion matrix - the table of true and false positives and negatives used to evaluate a classifier
- Precision - the proportion of positive predictions that were correct
- Recall - the proportion of actual positives that were correctly identified
- F1 score - the harmonic mean of precision and recall
- Model drift - degradation of model performance as live data diverges from training data
- Explainability - the ability to explain a model's decision to an affected person