Deep learning

Image classification

product photos, defect detection, medical imaging, moderation

Assign a label to each image, or find where something is in it.

Linear algebra

Calculus & gradients

Optimisation

Foundations

Convolutions are linear maps; backpropagation is the chain rule; everything trains by gradient descent.

Frame the problem

Frame

Labeling & data collection

Exploring the data

Data & labels

Class balance. Shortcut features: backgrounds, watermarks, capture device.

Modalities → vectors

Represent

Pretrained embeddings are the representation until you fine-tune.

Split the data

Split

Group by source or session to keep near-duplicates on one side.

Dumb baseline

Logistic regression

MLP & backpropbackground

Fine-tuned backbone

Fine-tune a pretrained CNN or ViT with augmentation. From scratch only for unusual domains.

Convolutional networks

Vision transformers & CLIP

Where, not just what

Detection and segmentation when the answer is a location.

Detection & segmentation

Model

Pretrained embeddings plus logistic regression. Surprisingly strong.

Fine-tuning a pretrained model

Training craft

Train

Classification metrics

Interpretability

Evaluate offline

Confusion matrix by class. Look at the worst errors. Saliency to catch shortcuts.

Evaluate online

Nothing unusual here.

Serving & release

Inference optimisation

Ship & monitor

Model size and latency. Quantise.