The Role of Features and Representations
What you'll learn: How machine learning models work with numerical representations rather than directly perceiving raw reality.
The Bridge Between Reality and Algorithms
Machine learning models don't "see" images, "hear" sounds, or "read" text the way humans do. They work exclusively with numbers—specifically, vectors and matrices of numerical values. Your job as an ML practitioner is to translate real-world data into meaningful numerical representations called features.
What Are Features?
Features are measurable properties or characteristics extracted from your data. For a house price prediction problem, features might include:
- Square footage (1,850)
- Number of bedrooms (3)
- Age in years (12)
- Distance to downtown in miles (4.2)
Each house becomes a vector: [1850, 3, 12, 4.2].
Why This Matters
The quality of your features determines the ceiling of your model's performance. A sophisticated algorithm with poor features will underperform a simple algorithm with well-chosen features. This is often called "garbage in, garbage out."
Think of it like cooking: even the best chef (algorithm) can't make a gourmet meal without quality ingredients (features). The raw data (unprocessed vegetables) must be prepared into usable forms (features) before the model can learn patterns.
Representation Learning
Different representations of the same data expose different patterns. An image might be represented as:
- Raw pixel intensities
- Edge strengths and orientations
- Statistical texture measures
Deep learning's power lies partly in learning these representations automatically rather than hand-crafting them.
Key Takeaway: Machine learning models learn from numerical features, not raw reality—choosing or learning good representations is fundamental to ML success.