6.13 Modern vision architectures
You can pick an architecture for a given compute budget.
Before:05. Classical Machine LearningUnlocks:07. Natural Language Processing11. The ECE Bridge — Signals, Edge & Embedded AI
Modern vision is a menu with a budget axis: ResNets as the reliable default, EfficientNet and MobileNet when compute is tight, vision transformers when data and scale justify them, ConvNeXt as the convergence point. Residual connections are the idea that made depth trainable everywhere. It follows the CNN foundations as the survey with judgement. The default to resist is the biggest model — on modest data, a pretrained ResNet fine-tuned well is very hard to beat.
Work through these
ResNet and residual connections
Connections that skip layers let gradients reach the early layers of very deep networks, which is what made depth beyond a few dozen layers practical. It is the single most consequential architectural idea here.
EfficientNet, MobileNet and scaling
Families designed for a given computation budget, including ones intended for phones. Choosing by budget rather than by leaderboard position is the working approach.
Vision Transformers (ViT) and patch embeddings
Applying the attention architecture to images by cutting them into patches, which brought the two halves of the field together. It needs more data than convolutional networks to train from scratch.
ConvNeXt and the convergence of the two families
A convolutional design updated with the lessons from attention architectures, which performs comparably. The convergence of the two families is the point of this item.
Sign in to keep your progress.
Free resources
We haven't checked most of these for screen reader use yet.
Links last checked 29 Aug 2026.
Stuck here?
Ask a mentor. A real person answers, and they can see exactly which topic you're on. Usually within a couple of working days.
Checking your session…
Topics shown in module order.