tohms

I went back over backprop recently and it was humbling how much this stuff had faded, especially nowadays with large models that encourage passive learning in my opinion. So writing it up in public was my way of relearning it.

I tried to make it intuitive, interactive and with some details and proofs for the curious, and a few exercises at the end to practise actively.

Feedback welcome, especially if a step is unclear or something is wrong.

Thank you, the interactive parts were the actual reason I started the article, glad it helped.

Thanks ! the article is trying to explain several other papers like the Product Quantization paper which was peer‑reviewed in IEEE but i totally agree.

yeah, my read is most production users stay on CPU and if they need heavier/faster workdloads in they might often move toward managed vector services like pinecone, mongodb atlas, or pgvector, and attention follows usage. Also FAISS came out of FAIR's team in Paris, Meta’s recent massive reorganization and layoffs which splitted FAIR's team likely have an impact in its maintenance.

Author here. I wrote this as a visual companion to the 2017 FAISS paper (Johnson, Douze, Jégou), focused on the parts I found hardest to grok from text alone. The article covers a subset of what FAISS does, with the paper as the source of truth. NSG, FastScan, IMI are not covered here, they'll get their own articles. I'd be especially interested in feedback on:

- the IVFPQ / IVFADC explanation, particularly the LUT reuse argument - whether the GPU part captures enough of the actual complexity

Happy to answer questions.