Course Details

Courses Descriptions

MODULE OBJECTIVE

What you'll achieve

Master NLP fundamentals, build neural network intuition from MLP to LSTM, follow the evolutionary path through Attention to Transformers, and understand how LLMs are trained.

SECTION 01 — NLP Foundations

Text as data — processing the written word

• Corpus, vocabulary, OOV words
• Tokenisation: character, word, sub-word (BPE)
• Stemming, Lemmatisation, Stopword removal
• Vectorisation: One-Hot, BoW, TF-IDF, N-grams
• Practical: IMDB sentiment with TF-IDF + LR

SECTION 02 — Word Embeddings

Dense semantic representations

• Why one-hot fails: sparsity, no semantics
• Word2Vec: CBOW and Skip-Gram architectures
• GloVe: global co-occurrence statistics
• Vector arithmetic: King - Man + Woman = Queen

Create & Compare Sparse vs Dense Representation

SECTION 03 — Deep Learning — MLP to RNN

Neural networks for sequence data

• Perceptron → MLP: forward pass and backprop
• Activation functions: ReLU, sigmoid, softmax
• CNNs for text classification
• RNNs and LSTMs: variable-length sequences
• Vanishing gradient — why Attention was needed

SECTION 04 — Attention & Transformer

The breakthrough that changed everything

• Seq2Seq limitations and the attention fix
• Self-attention and multi-head attention
• Positional encoding: giving order to tokens
• BERT (encoder) vs GPT (decoder)

SECTION 05 — From Transformer to LLM

How modern AI came to exist

• Scaling laws: data + parameters = emergent capability
• GPT-1 → GPT-2 → GPT-3 → GPT-4 → GPT-5 lineage
• Instruction tuning and RLHF
• Tokens, context windows, temperature, top-p
• Open-source LLMs: LLaMA, Mistral, Falcon

KEY TAKEAWAYS

1. Every LLM starts with tokenisation and embeddings

2. Attention solves the bottleneck RNNs couldn't

3. The Transformer is an architecture, not a model

4. Scaling laws: capability emerges from data + compute

5. RLHF turns raw LLMs into helpful assistants