umap

This is a data map providing a view of all Hacker News stories with a score of at least 2 (to remove most of the spam). Stories are close together in the map if they have semantically similar titles. In the bottom left is a histogram of stories over time. Hovering on a bar will select stories from that year, and dragging a selection allows selecting multiple years. A keyword based text search is in the upper left. Hold down the shift key and drag to lasso-select points and get a word cloud generated from the selection. Clicking on a point will open the URL for that story. The dataset was filtered from https://huggingface.co/datasets/OpenPipe/hacker-news Stories were embedded in a vector space via nomic-embed: https://huggingface.co/nomic-ai/nomic-embed-text-v1.5 A 2D representation was generated using UMAP: https://github.com/lmcinnes/uma

This is a data map providing a view of all HackerNews stories with a score of at least 2 (to remove most of the spam). Stories are close together in the map if they have semantically similar titles. In the bottom left is a histogram of stories over time. Hovering on a bar will select stories from that year, and dragging a selection allows selecting multiple years. A keyword based text search is in the upper left. Hold down the shift key and drag to lasso-select points and get a word cloud generated from the selection. Clicking on a point will open the URL for that story. The dataset was filtered from https://huggingface.co/datasets/OpenPipe/hacker-news Stories were embedded in a vector space via nomic-embed: https://huggingface.co/nomic-ai/nomic-embed-text-v1.5 A 2D representation was generated using UMAP: https://github.com/lmcinnes/umap

Looks like there is a little motion on this topic: https://github.com/lmcinnes/umap/pull/1103

Indeed, TensorFlow likes pushing everything to the GPU by default whereas many PyTorch DL implementations encourage feeding data from the CPU to the GPU as needed with a DataLoader. There have been attempts at a PyTorch port of Parametric UMAP (https://github.com/lmcinnes/umap/issues/580) but nothing as good.

side note: instead of t-SNE consider UMAP - provides better results (and it's much faster) https://github.com/lmcinnes/umap

PSA: Try UMAP instead of t-SNE https://github.com/lmcinnes/umap It's much faster and usually results in better clustering / representation.

In my experience, unsupervised learning is more broadly useful (and easier to implement) than supervised learning. Use it for data exploration, data validation, anomaly detection, topic modeling, recommender systems, cluster analysis, etc. Recommended algorithms: - UMAP (https://github.com/lmcinnes/umap/blob/master/README.rst) - HDBSCAN (https://github.com/scikit-learn-contrib/hdbscan/blob/master/...) - MatrixProfile (https://github.com/target/matrixprofile-ts/blob/master/READM...) - NMF (https://scikit-learn.org/stable/modules/generated/sklearn.de...)

Here’s a pretty comprehensive answer on the topic from the original UMAP author: https://github.com/lmcinnes/umap/issues/25 Clustering the output of UMAP is also given a nice tutorial in the docs: https://umap-learn.readthedocs.io/en/latest/clustering.html Basically, the answer is yes you can do this, but verify and analyze the output to ensure it makes sense (e.g. coloring points by known features/labels). For example, if you have a small number of points in the dataset (<1000), UMAP tends to display a dense cluster that is quite separated from the remaining data. However, this apparent cluster is spurious and contains noisy data points that UMAP couldn’t “figure out what to do with” (they are similar in their dissimilarity to the other data).

Uniform Manifold Approximation and Projection https://arxiv.org/abs/1802.03426 https://github.com/lmcinnes/umap

Indeed, there is the unfortunate fact that directly pointing out that a particular person has not read the paper is against HN's rules. Let us not violate the Prime Directive further. If I understand the origin of this paper correctly, it is mainly the idea and work of an undergraduate student. From this perspective, several problems become more understandable. The limited computation done in support of the paper may have been done by a setup as meager as a lone student's laptop. The naïve view of machine learning as a field could be explained by undergraduate eagerness; indeed, perhaps they did "recently [find] out about least squares". I agree with your thoughts on SGD. The authors did do PCA for some of their tests, but this entire setup should be buildable in a more efficient and incremental way. I think, however, that the reason that

The name UMap is a bit unfortunate, as that is also an abbreviation for Uniform Manifold Approximation and Projection, which is a kind of mathematically more robust tSNE. In fact, for a split-second I was kind of hoping someone had figured out an interesting way of applying that technique to OSM data. Also, the demo doesn't load for me in any browser I throw at it - at first I thought it was my overly strict security settings, but I'm getting 404s even when running Chromium without any privacy plugins, adblockers. [0] https://github.com/lmcinnes/umap

There's a longer explanation of how the images were produced: https://johnhw.github.io/umap_primes/index.md.html I had not heard of UMAP before, but it seems to be a tool to visualise correlation in high dimensional datasets: https://github.com/lmcinnes/umap

Not sure if a major program but here is a recent and promising one: https://github.com/lmcinnes/umap