Transformers Architecture
Self-attention, multi-head attention, encoder-decoder, positional encoding
Why Transformers? — The Problem with RNNs at Scale
intermediate~28minSequential bottleneck, long-range dependency failure, parallelism need
Encoder-Decoder Architecture Overview
intermediate~28minHigh-level: what encoder does, what decoder does, when used together
Self-Attention Mechanism Explained
advanced~35minQuery, Key, Value matrices, attention score calculation, softmax weighting
Attention Intuition: "Turn Off the Lights" Example
advanced~35minH1..H4 encoder states, C1..C4 context vectors, alpha weight meaning
Attention Weights — Alpha Formula Deep Dive
advanced~35minAlpha(i,j) = f(H_i, S_j), 16 alphas for 4x4 encoder-decoder, generalized formula
Multi-Head Attention
advanced~35minParallel attention heads, concatenation, what each head learns
Positional Encoding
advanced~35minWhy needed (no sequence order in attention), sine/cosine formula, intuition
Encoder-Only vs Decoder-Only vs Full Transformer
intermediate~28minBERT=encoder-only, GPT=decoder-only, T5=full encoder-decoder — when to use which
Transformation Blocks (N Stacked Layers)
advanced~35minLayer normalization, residual connections, feed-forward network in each block
CNN vs Transformer Architecture Comparison
intermediate~28minConvolution vs attention, local vs global context, computational differences