Back

01_basic_definitions_and_concepts

Loading views...
Next →

01. Basic Definitions And Concepts

1. Overview

Machine learning (ML) is an approach employed when the input/output structure of a problem is known, but a complete explicit algorithm to map inputs to outputs is absent. Instead of hand-crafting rules, ML utilizes data to infer the underlying transformation automatically.

2. Core Concepts

  • Definition: Machine learning is programming computers to optimize a performance criterion using example data or past experience.
  • Algorithm Limitations: Explicit algorithms exist for structured tasks (e.g., sorting). However, complex real-world tasks (e.g., spam filtering, facial recognition) lack direct programmatic solutions.
  • Data-Driven Inference: In the absence of explicit knowledge, ML compensates by collecting labeled examples to learn the appropriate mapping.
  • Optimization: Learning fundamentally involves adjusting internal model parameters using training data to approximate the true underlying process.

3. The Role of Machine Learning

  • Adaptive Systems: Essential for artificial intelligence (AI) systems operating in dynamic environments where rigid rules fail.
  • Pattern Recognition: Excels at tasks where human intuition operates effectively but cannot be formally codified (e.g., recognizing that a face is symmetric with eyes, nose, and mouth, rather than a random pixel collection).
  • Data Mining (KDD): Applying ML to massive datasets to extract compact, high-value predictive structures. Analogy: Processing a massive volume of earth to extract a small amount of precious material. This helps predict hyper-specific consumer behaviors (e.g., buying spices for Glühwein in winter, clicking a specific web link, or choosing an ice cream flavor).
  • Model Purpose: Models can be predictive (making predictions about the future) or descriptive (gaining knowledge and explaining the data process), or both.
  • Interdisciplinary Nature: ML lies at the intersection of statistics (inference from finite samples) and computer science (efficient, scalable algorithms). Time and space complexity are often as critical as predictive accuracy.

4. Mathematical Abstraction

The core abstraction of an ML model can be expressed as:

Where:

  • x: Observed attributes or features.
  • y: Target output.
  • g(\cdot): Model family (the functional form).
  • \theta: Model parameters adjusted during the learning process.

The learning objective is typically defined as minimizing a loss function L over a dataset D:

\theta^* = \arg\min_\theta L(\theta; D)

5. System Architecture

       (Underlying Process)
                |
                v
          [Raw Data] -> (Preprocessing) -> [Features, x]
                                                 |
                                                 v
(Training Labels, y) <---- Loss L(θ) <---- [Model: y = g(x|θ)]
                                                 ^
                                                 |
                                      (Update Parameters θ)
Next →