How I hacked UMAP and won at a plotting contest
summary
Hacks UMAP source to save intermediate embeddings, runs it on MNIST and draws every epoch, stitches frames with ffmpeg, and wrangles data with Polars, Plotnine, ffmpeg. The hack spits out many NumPy files and is messy but works; the animation shows surprisingly little movement after early epochs. Result wins runner-up in the 2024 Plotnine plotting contest and demonstrates a reproducible method for visualizing algorithm behavior.
AI-written summary · about
chapters (14)
- 0:00Welcome!
- 1:06Plotnine contest
- 2:18Hoi! I'm Jeroen
- 3:13Visualizing the behavior of an algorithm
- 7:58Keeping track of intermediate predictions
- 9:23Tool: UMAP
- 12:31Tool: Polars
- 13:15Python Polars: The Definitive Guide
- 13:40MNIST dataset
- 14:46Tool: Plotnine
- 17:04Tool: FFmpeg
- 18:58Final result
- 19:38To conclude
- 20:40Q & A
related talks
51mGoofing Off with Fun (Big) Data
48mMAP All the Things
28mUnderstanding UMAP Internal Workings of a State of the Art Clustering Algorithm
38mData Mapping for Data Exploration
19mmayAI? Exploring Graphical Tools for Model Visualization & Enh.
56mPicasso's terminal; data science and AI in the visual arts
40mTopological Techniques for Unsupervised Learning
41mLuxuries, necessities, and the challenges that remain