Skip to main content

Advanced Vision

CNNs and ResNets are the workhorses of image classification, but they aren't the whole story. This track covers two architectures that go beyond the convolutional backbone: the Vision Transformer, which brings attention to images, and the U-Net, which produces a full-resolution output for dense prediction.

What you'll build

  • Vision Transformer (ViT) — split an image into patches, embed them as tokens, and classify with a transformer encoder. A global receptive field from the very first layer, using the built-in nnx.MultiHeadAttention.
  • U-Net Segmentation — an encoder-decoder with skip connections that labels every pixel, using nnx.ConvTranspose to upsample back to full resolution.

Prerequisites

You should have built a CNN and understand transformers (ViT reuses the encoder) and ResNet skip connections (U-Net generalizes them).

Next steps