03_supervised_learning
03. Supervised Learning
1. Overview
Supervised learning is the paradigm of learning a mapping from input attributes to corresponding known target values, utilizing a dataset containing explicit input-output pairs.
2. Core Methodologies
- Classification: The output is a discrete class code (e.g., binary or multiclass).
- Credit Scoring: Classifying applicants into low-risk vs. high-risk. A rule could be learned as: IF income > θ1 AND savings > θ2 THEN low-risk ELSE high-risk.
- Face Recognition: A highly complex multi-class task dealing with 3D nature, differences in pose, lighting, and occlusion (e.g., glasses, beards).
- Medical Diagnosis: Must handle missing inputs (due to tests being costly or time-consuming). Probabilistic outputs allow the classifier to reject and defer the decision to a human expert when in doubt.
- Speech Recognition: Extremely difficult due to temporal sequences and varying phoneme lengths; fundamentally requires a "language model" to improve accuracy.
- Regression: The output is a continuous numeric value.
- Used Car Pricing: Predicting the price based on attributes (brand, year, mileage) using linear formulations like y = wx + w0 or higher-order polynomials.
- Autonomous Navigation: Outputting the continuous steering angle for a vehicle using inputs from a video camera and GPS.
- Response Surface Design: Regressions can be used to optimize functions, such as finding the optimal temperature and time settings to maximize coffee roasting quality.
3. Model Dependencies
- Data Representation: Algorithms rely on representative training data. Missing values must be explicitly handled.
- Data Integrity: Quality heavily depends on ensuring that future data remains consistent with the training distribution.
4. Mathematical Formulation
The supervised learning mapping can be defined as a function bridging the feature space \mathcal{X} and the target space \mathcal{Y}:
f: \mathcal{X} \rightarrow \mathcal{Y}
Where the prediction \hat{y} for a given input x is:
Where \theta represents the optimized parameters derived from the training pairs (x_i, y_i).
5. Supervised Architecture
[Training Data: (X, Y)]
|
v
[Algorithm] <---- (Loss Minimization)
|
v
[Learned Function: f(X)]
|
v
[New Input, X'] ---> [ f(X') ] ---> [Prediction, Y']