How a computer scores two pictures

Research cited

1. Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger and (excerpt) ↗arxiv.org

Other sources

1. scikit-learn developers, "sklearn.metrics.pairwise.cosine_similarity" (documentation), scikit-learn ↗scikit-learn.org
2. Slow small ai (ddw.nl) ↗ddw.nl
3. Slow small ai (visualmethodologies.org) ↗visualmethodologies.org

Images, slide by slide

Slide 1

A dark gallery wall with a screen showing the number 0.67 beside a picture of a red building, and a second screen with a grid of many photosPhoto: RN Collins

Slide 2

Code-drawn diagram: a simple picture passes through a box labeled model and comes out as a bracketed column of numbersImage: RN Collins studio, code-drawn. Checked against Radford et al., "Learning Transferable Visual Models From Natural Language Supervision", arXiv 2103.00020, 2021 ()Museum or journal record ↗

Slide 3

Code-drawn diagram: two arrows from one point, picture A and picture B, with the angle between them marked, and a shorter black arrow along B. It shows that the score depends on direction onlyImage: RN Collins studio, code-drawn. Checked against scikit-learn documentation, sklearn.metrics.pairwise.cosine_similarity ()Museum or journal record ↗

Slide 4

Code-drawn diagram: a worked sum for the lists (1, 1, 1) and (1, 1, 0) ending in 0.816, above a scale from -1 to 1 with the result markedImage: RN Collins studio, code-drawn. Checked against scikit-learn documentation, sklearn.metrics.pairwise.cosine_similarity ()Museum or journal record ↗

Slide 5

Two screens on a dark wall: one shows a purple corridor and a building side by side, the other a mosaic of many small photosPhoto: RN Collins

Ring frame

HAL DS3100 ASR video display terminalFrame: Ulli1105, CC BY 3.0, croppedMuseum or journal record ↗