Matt Cornillon: How I found my Pokémon cards thanks to Postgres: an AI journey (PGConf.EU 2023)

Explore the journey of using Postgres and a machine learning model to detect Pokémon cards from images, with a demo of the real-life application and discussion on vector storage and manipulation.

Key takeaways

The speaker used a machine learning model to detect Pokémon cards from images and stored the embeddings in Postgres using the PG Vector extension.
The embeddings were generated using a convolutional neural network (CNN) and cosine distance was used for similarity search.
The speaker mentioned that the number of dimensions in the generated embedding is determined by the machine learning model, which in this case is 768 dimensions.
The PG Vector extension allows for storing and manipulating vectors, including inserting, deleting, and updating data.
The speaker also mentioned that the machine learning model can be improved continuously by reusing the pictures and filling the model with new data.
The speaker used Hugging Face to generate the embedding and also mentioned that it’s a company offering open-source machine learning models and data sets.
The speaker used label studio to create a data set for machine learning and mentioned that it’s a tool that enables creating a data set from pictures in a good format.
The speaker also mentioned that the PG Vector extension offers three different distance methodologies: cosine similarity, L2 distance, and IVF flat.
The speaker showed a demo of the application and mentioned that it’s a real-life use case and not just a theoretical example.
The speaker mentioned that the similarity search using PG Vector is exact nearest neighbor search if no indices are used, but approximate nearest neighbor search if indices are used.
The speaker also mentioned that storing pixels inside Postgres might not be efficient and that the embeddings are a better way to store the data.

Matt Cornillon: How I found my Pokémon cards thanks to Postgres: an AI journey (PGConf.EU 2023)

More talks