romanfll

11 ms·

1. 1.▲
   by romanfll 8mo ago

   Good catch—it's actually a hybrid approach: 1. Anomaly Detection & Visualisation: This runs 100% locally via WASM using the linear-time algorithms I shared here previously. 2. Forecasting: This hits our DriftMind API [1][2]. We des
2. 2.▲
   DriftMind – Anomaly detection and forecasting for Excel users (no signup)(thingbook.io)

   2 points by romanfll 8mo ago|3 comments
3. 3.▲
   by romanfll 8mo ago

   I previously shared my linear-time dimensionality reduction algorithm https://news.ycombinator.com/item?id=46285535 I've now wrapped that + my one-pass forecasting engine into a free tool for non-coders. Tool explaine
4. 4.▲
   Reflexive Memory: A CPU-Only Alternative to Transformers for Streaming Data(medium.com)

   3 points by romanfll 8mo ago|1 comments
5. 5.▲
   by romanfll 8mo ago

   This post details the architecture behind DriftMind, a streaming forecasting engine we built to replace Transformers at the edge.The Context:We monitor high-frequency industrial sensor streams where the cost of round-tripping data to a GPU
6. 6.▲
   by romanfll 9mo ago

   The shift from Explicit Reduction to GNNs/Embeddings is where the high-end is going in my view… We hit this exact fork in the road with our forecasting/anomaly detection engine (DriftMind). We considered heavy embedding models but
7. 7.▲
   by romanfll 9mo ago

   Not yet, but coming...
8. 8.▲
   by romanfll 9mo ago

   You are strictly correct for a single pass! log2(9000)~13, which is indeed much smaller than k=50. The missing variable in that comparison is Iterations. t-SNE and UMAP are iterative optimisation algorithms. They repeat that O(N log N) st
9. 9.▲
   by romanfll 9mo ago

   Thanks for your comment. To clarify: K is a fixed hyperparameter in this implementation, strictly independent of N. Whether we process 9k points or 90k points, we keep K at ~100. We found that increasing K yields diminishing returns very qu
10. 10.▲
    by romanfll 9mo ago

    The '2 seconds' figure comes from the end-to-end time on a standard laptop. I quoted 2s to set realistic expectations for the user experience, not the CPU cycle count. You are right that the core linear algebra (Ax=b) is milliseco
11. 11.▲
    by romanfll 9mo ago

    Thanks for your comment. You are spot on, that is effectively the standard Nyström/Landmark MDS approach. The technique actually supports both modes in the implementation (synthetic skeleton or random subsampling). However, for this br
12. 12.▲
    by romanfll 9mo ago

    Thanks! You nailed the intuition! Yes, it shares DNA with Landmark MDS, but we needed something strictly deterministic for the UI. Re: Publishing: We don't have a paper planned for this specific visualisation technique yet. I just want
13. 13.▲
    by romanfll 9mo ago

    Thanks for your comment! You are right, Barnes-Hut implementation brings UMAP down to O(N log N). I should have been more precise in the document. The main point is that even O(N log N) could be too much if you run this in a browser.. Thank
14. 14.▲
    by romanfll 9mo ago

    Author here. I built this because I needed to run dimensionality reduction entirely in the browser (client-side) for an interactive tool. The standard options (UMAP, t-SNE) were either too heavy for JS/WASM or required a GPU backend to
15. 15.▲
    A linear-time alternative for Dimensionality Reduction and fast visualisation(medium.com)

    118 points by romanfll 9mo ago|36 comments
16. 16.▲
    by romanfll 9mo ago

    Author here. I built this because I needed to run dimensionality reduction entirely in the browser (client-side) for an interactive tool. The standard options (UMAP, t-SNE) were either too heavy for JS/WASM or required a GPU backend to