Linear Classifier

Linear Classification Network

Overview

Linear classification is a parametric approach to image classification that separates the process into two key components:

1. Score Function

Maps pixel values to confidence scores for each class

• Input: Raw image pixels
• Output: Confidence score for each class

D-DL4CV-Lec03a-Linear_Classifier

2. Loss Function

Evaluates predictions by score function and guides learning

• Input: Score function output + correct labels
• Output: Error measure that drives parameter updates in score function

D-DL4CV-Lec03b-Loss_Function

The highlight of linear classification network in contrast with kNN is that linear classifier can "learn" how to classify images by processing training data

How They Work Together

Repeat:

1. Score function turns raw image into class scores
2. Loss function computes loss based on class scores and the correct label
3. Update weights based on the loss

Regularization: Solution to Overfitting

D-DL4CV-Lec03c-Regularization