Transformer Encoder Decoder Architecture Explained Masked Attention Cross Attention

In this video we continue learning the Transformer Architecture from the famous research paper “Attention Is All You Need” (2017). This lecture explains the complete Encoder–Decoder Transformer pipeline including Multi-Head Attention, Add & Norm, Feed Forward layers, Masked Attention, Cross Attention and Autoregressive decoding. GitHub Repository https://github.com/switch2ai You can download all code, scripts and documents from the repository. Evolution of Sequence Models 2014 – Encoder Decoder Architecture (Google) Models could convert one sequence to another such as machine translation. 2015 – Attention Mechanism Attention allowed models to focus on important parts of the input sequence instead of compressing everything into a single vector. 2017 – Transformer Architecture Transformers removed recurrence completely and relied entirely on attention mechanisms, enabling parallel processing and long-range dependency learning. Example Task: Machine Translation English We are learning transformer Hindi Hum transformer sikh rahe hai Transformer has two main components. Encoder Processes the source sentence. Decoder Generates the target sentence step by step. Encoder Architecture Input Sentence "We are learning transformer" Step 1 Tokenization ["we","are","learning","transformer"] Step 2 Convert Tokens to IDs Example [987,10,300,765] Step 3 Input Embedding Token IDs are converted into dense vectors using an embedding layer. Embedding dimension used in the original Transformer paper = 512 Step 4 Positional Encoding Since Transformers process tokens in parallel, positional encoding helps the model understand word order. Example Ind beats NZ NZ beats Ind Same words but different meaning because of word order. Final vectors W+P1 A+P2 L+P3 T+P4 Step 5 Multi-Head Attention Language contains multiple complexities like syntax, grammar, references and meaning. A single attention head cannot capture all relationships. Multi-Head Attention runs attent