Computer VisionBeginner1:20 video6 chapters

Image Classification End to End — lecture notes

From a labelled dataset to a trained classifier: the full pipeline, softmax probabilities and how to judge the results.

0:001. Introduction

Introduction — Image Classification End to End

Image classification answers one question: what is in this picture? Let us walk through the whole pipeline, from data to prediction.

0:092. Picture to prediction

Picture to prediction — Image Classification End to End

An image goes in. A convolutional network extracts features layer by layer. The final layer gives a score for each class, and softmax turns those scores into probabilities: here 86 percent house, 8 percent barn, 4 percent castle and 2 percent tent.

0:273. The pipeline

The pipeline — Image Classification End to End

Collect and label images for every class. Split off validation and test sets. Augment the training images for variety. Train, usually by fine-tuning a pretrained network. Then evaluate, class by class, and study the mistakes.

0:434. Softmax

Softmax — Image Classification End to End

Softmax exponentiates each class score so it is positive, then divides by the total so everything sums to one. The largest score gets the largest probability.

0:555. Pitfalls

Pitfalls — Image Classification End to End

Watch for pitfalls. Imbalanced classes. Shortcut learning, where the model looks at the background instead of the object. Near duplicate images leaking into the test set. And overconfidence: a high probability is not a guarantee.

1:106. Recap

To recap. Image, features, scores, softmax. Collect, split, augment, train and evaluate. Start from a pretrained model, and always inspect the mistakes.

Key takeaways

A classifier maps an image to class probabilities via softmax.
The pipeline: collect/label, split, augment, train, evaluate.
Fine-tuning a pretrained network is the usual starting point.
Beware imbalance, shortcut learning, leakage and overconfidence.

Check yourself

What do softmax outputs add up to?
Show answer
1 (100%) — Softmax normalises the scores into a probability distribution.
A model recognises cows only when there is grass in the picture. This is…
Show answer
Shortcut learning — It relies on a spurious background cue.
What is usually the best way to start a new image classifier?
Show answer
Fine-tune a pretrained network — Transfer learning works well with limited data.

Go deeper

Building an Image Classification Pipeline End to End · The AI Lecture Hall
Transfer Learning and Fine-Tuning · The AI Lecture Hall
Softmax Regression and Multiclass Classification Strategies · The AI Lecture Hall