Deep Learning and Transformers: concepts and vocabulary | Complete Generative AI: From Foundations to Production
A self-paced 26-week learning path with 15 modules: AI foundations, Python, machine learning, transformers, LLMs, prompts, APIs, embeddings, RAG, frameworks, agents, customization, multimodal AI, Spring Boot with React, and production LLMOps. Includes 15 practical assignments and three portfolio projects.
Prerequisites: Begin with basic computer skills. Programming is introduced in Weeks 2–3. Use a local Python environment for coding labs; Java and React experience helps in Weeks 21–22. Model downloads, provider accounts and deployment resources are optional learner-managed dependencies. The 26-week schedule is suggested; reading estimates exclude independent project work.
Learning goals
Neural network architecture
Word embeddings and tokenization
Attention and self-attention
Transformer architecture
Encoders, decoders and positional information
Pretraining, instruction tuning and inference
Understand the foundations
Tokenization maps text into units from a vocabulary and then into IDs; tokens are not necessarily words. Embeddings map token IDs to learned vectors. A transformer uses attention to mix information across positions, feed-forward layers to transform representations and position information to distinguish order. Encoder-style attention can use a whole input sequence, while a causal decoder masks future positions during next-token prediction. Architecture and masking determine which information each position may access.
Connect the ideas
Scaled dot-product attention forms scores from queries and keys, applies an attention mask, normalizes scores with softmax and combines value vectors. Multiple heads can learn different relationships, but an attention weight should not be treated as a complete explanation of a model decision. Pretraining learns broad statistical patterns from an objective such as next-token prediction. Instruction tuning changes behavior using instruction-response examples. Inference applies fixed parameters to new inputs; it is distinct from updating model weights.
Checkpoint: explain these concepts in your own words, then describe one failure that the design must detect. Use synthetic or permissioned data throughout the course.
01. Introduction to Artificial Intelligence · Week 1
02. Python Programming for Gen AI · Weeks 2–3
03. Mathematics and Machine Learning Foundations · Week 4
04. Deep Learning and Transformers · Weeks 5–6
05. Large Language Models (LLMs) · Weeks 7–8
06. Prompt Engineering · Weeks 9–10
07. Generative AI APIs and Application Development · Week 11
08. Embeddings and Vector Databases · Weeks 12–13
09. Retrieval-Augmented Generation (RAG) · Weeks 14–15
10. LangChain, LlamaIndex and AI Frameworks · Week 16
11. AI Agents and Agentic AI · Weeks 17–18
12. Fine-Tuning and Model Customization · Week 19
13. Multimodal Generative AI · Week 20
14. Gen AI with Java Spring Boot and React · Weeks 21–22
15. Production Gen AI, LLMOps and Capstone · Weeks 23–26