What Happens Inside a Transformer Block ? Is Attention Really All You Need ?

[Chip (OP)]

Multi Head Attention:

How 8 Attention heads work in parallel and combine their output.

Skip Connections and Layer Normalization:

Why Transformers use shortcuts and how they help deep networks train effective.

The Feed Forward  Network:

How each token is processed independently to build richer representations.

Stacking Encoder Blocks:

How multiple blocks transform the input into increasingly useful representations.