1.ML-Ingredients
Introduction to Machine Learning System Ingredients
Overview of ingredients: 5 essential components in a machine learning system.
Problem Illustration: Classifying Cats and Dogs
Example of a binary classification problem.
Supervised learning using labeled training data (images of cats and dogs).
Goal: Classify unseen images as either cat or dog.
Key Components of the System
1. Input
Input can vary in complexity; in this case, it is an image.
Need to convert the image into a format suitable for the system.
Images represented as a matrix of pixel values.
Example: A 28x28 image corresponds to 784 pixel values (P1 to P784).
2. Output
For binary classification, there are two outputs:
Probability of belonging to the 'Dog' class.
Probability of belonging to the 'Cat' class.
Probabilities range from 0 to 1:
If the probability of 'Dog' is 0.7, then 'Cat' will be 0.3 (1 - 0.7 = 0.3).
Two output values simplify extension to multi-class classification.
3. Mapping Function
The mathematical function that maps inputs to outputs.
Also referred to as the hypothesis (F).
Different classification techniques (logistic regression, SVM, neural networks) will have different hypotheses.
The core function remains: mapping input (784 vector) to output (class probabilities).
4. Cost Function
Measures the error between predicted outputs and true outputs.
Also known as the loss function.
Essential for evaluating the performance of the mapping function.
True labels known from training data allows calculation of errors based on predictions.
Provides a mechanism to assess how well the system is performing.
5. Learning Procedure
The iterative process of minimizing the cost function by adjusting the mapping function.
Introduces flexibility into the mapping function via trainable parameters (W):
Changes in W result in changes in the mapping function, enabling minimization of error.
The goal of the learning process is to find parameter values (W) that reduce the cost function and thereby improve classification accuracy.
Example of Learning Process
Illustrative binary classification problem with two features (S1 and S2).
Hypothesis involves finding a straight line to separate classes.
W represents slope and b represents intercept.
Initial guess might yield high error; iterative adjustments lead to improved mappings:
Adjust W and b to minimize the cost function.
The process continues until an optimal line (and parameters) is found, resulting in minimal error.
Conclusion
Learning in machine learning involves minimizing error on training data by dynamically adjusting the mapping function through trainable parameters.
Common practice in programming is using a 'fit' method (e.g., logistic regression, etc.) to perform this learning process.