"Embeddings: How Computers Learned to Read"
summary
Shows that embeddings move from sparse co-occurrence matrices to neural 300-dimension vectors (word2vec) and then to context-aware BERT embeddings (2018), and demonstrates practical tooling: computes cosine similarity over session embeddings with SentenceTransformers + cross-encoder reranking, visualises clustered titles, and reranks an irrelevant Micropython hit down the list. Builds a small interactive retrieval demo — a live web app demo that retrieves nearest neighbours from a vector store and it works well enough to inspect and refine results.
AI-written summary · about
chapters (17)
- 0:00Embeddings introduction
- 2:30Distributional hypothesis
- 3:14Pre-2013 statistical methods
- 7:12Word2Vec and neural embeddings
- 8:57Neural network primer
- 10:23Training Word2Vec
- 11:48Dense embeddings properties
- 13:54Static embeddings limits
- 15:24BERT and dynamic embeddings
- 16:51Transformer architecture
- 17:55BERT’s contextual embeddings
- 18:15Sentence/document embeddings
- 20:03Embedding methods evolution
- 20:47Demo: sentence similarity
- 23:15Cross-encoder reranking
- 24:19Embeddings real-world apps
- 27:04Summary and demo
related talks
27mData with Purpose: The Power of Embeddings
16mEmpower AI: Building Smarter Solutions with Embeddings
46m"Embeddings - the lesser known Hero of AI"
31mEmbeddings... Embeddings Everywhere
47mThe Neural Search Frontier #bbuzz
48mUnderstanding Embedding Vectors and the RAG Pattern in AI Solutions
31mText and Code Embeddings
31mTaking AI/ML to the next level: Text!