Master Transformers from Scratch: BERT, ViT & AI Search

Course Overview

This course provides a comprehensive understanding of transformer models, focusing on the transformer encoder. It covers foundational concepts such as self-attention, queries, keys, values, multi-head attention, positional encoding, and the transformer encoder block. The course includes practical coding exercises in PyTorch and fine-tuning pre-trained models like BERT and Vision Transformers using Hugging Face. It also covers building a semantic search engine with a two-tower architecture and prepares learners for common machine learning and NLP interview questions.

Key Takeaways

• Understand self-attention step by step with real numbers.
• Explain queries, keys, values, and multi-head attention clearly.
• Understand every part of the transformer encoder block, including positional encoding.
• Understand how BERT is built and pre-trained.
• Build key transformer parts in PyTorch including tensors, autograd, nn.Module, and data pipelines.
• Fine-tune BERT for text classification and named entity recognition with Hugging Face.
• Apply Vision Transformers to image classification.
• Build a semantic search engine with a two-tower architecture.
• Answer common transformer questions in machine learning and NLP interviews.

Prerequisites

• Basic Python programming.
• High-school level math is sufficient; vectors, dot products, and matrices are taught from the start.
• No previous deep learning or transformer experience needed.

Target Learners

• Developers and data scientists who use transformers but want to understand their inner workings.
• Students who found attention and transformers confusing before.
• Anyone preparing for machine learning or NLP interviews.
• Learners who prefer clear explanations, simple English, and step-by-step examples.