The Landscape of Machine Learning on ArXiv

The Landscape of Machine Learning on ArXiv(lmcinnes.github.io)

31 points by lmcinnes 2 years ago | 10 comments

BenoitP 2 years ago |

Pretty interesting!

A big surprise for me was to find the explainability cluster quite far from the causality one. But I guess it stems from a cultural facet: causality is mainly the purview of statistics (with Pearl at the helm) with a strong medical sciences focus; when explainability is more of a reaction to algorithm used in the industry (trees, GLMs, neural networks, etc.); which you deploy for performance, and only then you care about knowing the why.

pyromaker 2 years ago |

This is cool! Always find these visualization helpful but it does get quite big sometimes.

[shameless plug]

I created a personalized newsletter for Arxiv (which I plan to expand to others) so you can receive the latest research papers on the topics. You can also filter by some keywords too (e.g. only give me LLM or RAG related papers).

https://app.scholars.io

lbeckman314 2 years ago |

Beautiful!

https://github.com/TutteInstitute/datamapplot

esafak 2 years ago | |

Thank you! What about the frontend? This is palpably faster than other graph libraries.

lbeckman314 2 years ago | | |

Interesting! Looks like it's bundled as a single HTML file (and then hosted on Github Pages). [0]

From the DataMapPlot Docs [1]:

"While the interactive plots work well in-line in a notebook, it is often useful to be able to send plots to others. The save method allows you to save the plot to a single HTML file that will embed (compressed) copies of the data. You can then share the HTML as required."

[0] https://github.com/lmcinnes/datamapplot_examples/blob/master...

[1] https://datamapplot.readthedocs.io/en/latest/interactive_int...

rdedev 2 years ago |

I created something like this for my data engineering class project. It was a temporal visualization of citation networks. It was fun to see different domains like computer vision and nlp be seemingly separate but then as time went on become pretty coupled with each other

fleischhauf 2 years ago | |

I suppose the underlying modality does not matter too much if it is about learning per se

HanClinto 2 years ago |

What is being used to build the data map -- how does one project these document vectors into 2D space?

BenoitP 2 years ago | |

OP is also the author of the popular dimensionality reduction algorithm UMAP.

I guess the pipeline was embedding documents with an LLM (or even plain old word2vec average over the abstract might do it), and then reducing that to 2 dimensions with a cosine similarity metric with the help of UMAP.

I have no idea about colors and local cluster naming though. Maybe that's handcrafted.

dkural 2 years ago |

Please add bioRxiv if you can, so many life-sciences relalted ML papers there.