# CNN Architectures in Brief — Computer Vision

Source: https://www.skillbyai.com/en/computer-vision/k-arch

> From LeNet to ResNet and beyond.

## Deeper, wider, smarter

Convolutional neural networks stack convolution, normalisation, activation and pooling layers that learn edges, textures, parts and objects. Landmarks: **LeNet** (digits), **AlexNet** (2012 ImageNet breakthrough), **VGG** (deep, simple), **ResNet** (residual connections make very deep networks trainable), **MobileNet** and **EfficientNet** (efficient for mobile and edge), **ConvNeXt** (modernised CNN design). **Vision Transformers (ViT)** split images into patches and apply transformer layers; with enough data or pretraining they match or beat CNNs. In practice you start from a pretrained model of a suitable size.

## Pretrained networks do most of the work

Convolutional and transformer networks pretrained on large datasets classify images and transfer to new tasks.

![Three ideas: architectures, pretrained inference, transfer learning.](assets/figures/computer-vision/section-4-map.svg) — Figure 4.1 — Architectures, pretrained inference and transfer learning.

## Choosing a backbone

Size versus accuracy versus speed.

```text
family          strengths                                 typical use
ResNet-18/50    reliable, well understood, many weights   baselines, transfer learning
MobileNetV3     tiny and fast                              phones, embedded devices
EfficientNet    accuracy per compute                       balanced deployments
ConvNeXt        modern CNN, strong accuracy                server-side vision
ViT / DINO      strong pretrained features, scales well    large data, foundation models
```

## Start small

Try a small pretrained backbone first; move to larger ones only if validation results justify the extra latency and cost.

**Quiz:** What made very deep CNNs like ResNet trainable?

- [ ] Training without labels
- [ ] Removing all convolutions
- [ ] Using grayscale only
- [x] Residual (skip) connections

*Answer:* Residual (skip) connections. Skip connections keep gradients flowing.