CSE559A Lecture 13

Positional Encodings

Set of sinusoids of different frequencies.

f(p,2i)=\sin(\frac{p}{10000^{2i/d}})\quad f(p,2i+1)=\cos(\frac{p}{10000^{2i/d}})

MLP is hard to learn high-frequency information from scaler input $(x,y)$ .

Example: network mapping from $(x,y)$ to $(r,g,b)$ .

Dependence on location, scaler, metadata, etc.
Can just be fully learned (use nn.Embedding and optimize based on a categorical input.)

In Vision Transformers, a special token called the class token is added to the input sequence to aggregate information for classification tasks.

PxP convolution with stride P (split the image into patches and use positional encoding)

Build a hybrid model that combines the vision transformer after 50 layer of ResNet.

At least for now, CNN and ViT architectures have similar performance at least in ImageNet.

General Consensus: once the architecture is big enough, and not designed terribly, it can do well.
Differences remain:
- Computational efficiency
- Ease of use in other tasks and with other input data
- Ease of training

Self attention as a key building block

Flexible input specification using tokens with positional encodings

A wide variety of architectural styles

Up Next: Training deep neural networks