Jo Kristian Bergum

@jobergum

Sr Principal Engineer at working on . Follow for updates from Vespa team. Git profile

Trondheim, Norge
Vrijeme pridruživanja: srpanj 2009.

Tweetovi

Blokirali ste korisnika/cu @jobergum

Jeste li sigurni da želite vidjeti te tweetove? Time nećete deblokirati korisnika/cu @jobergum

  1. This is an outstanding paper on various sentence embedding techniques . Great accuracy on STS (semantic text similarity) benchmarks, almost sota accuracy but at a fraction of the cost. Also enables semantic retrieval using approximate neighbor search.

    Poništi
  2. proslijedio/la je Tweet
    30. sij

    Awesome ( serving engine) updates: tensor functions, new sizing guides, performance improvement for matched elements in map/array-of-struct, and boolean field query optimization. Explore here: .

    Poništi
  3. 29. sij

    Almost forgot, and LightGBM integration.Fantastic to work with so many talented engineers all pulling in the same direction!

    Prikaži ovu nit
    Poništi
  4. 29. sij

    So many great things in progress at hq this sprint, super energized! Approximate neighbor search over HNSW index evaluation. Parameters, accuracy loss and performance. Great progress. Also Deep BERT ranking integration coming closer with tensor GEMM optimizations.

    Prikaži ovu nit
    Poništi
  5. 29. sij
    Poništi
  6. proslijedio/la je Tweet
    20. sij

    Hi all, your regular reminder that the Call for Papers for Haystack US, the search relevance conference, ends Jan 31st - please send us your search and relevance talks!

    Poništi
  7. 21. sij
    Poništi
  8. 19. pro 2019.
    Poništi
  9. 18. pro 2019.

    One of many applications which uses in . Check out this presentation from AWS re:Invent 2019: How Verizon Media implemented push notification using Amazon DynamoDB. 150K/s push notifications, resolving of targets done by Vespa.

    Poništi
  10. proslijedio/la je Tweet
    11. pro 2019.

    Today I had the pleasure to talk about the bug bounty community on a panel with and a few folks from at ’s conference! 🙏🏼

    Poništi
  11. 5. pro 2019.

    Interesting to see launch Kendra enterprise search with question answering capabilities using deep learning models. We at the Vespa team recently described how to do semantic question answering using ,

    Poništi
  12. proslijedio/la je Tweet
    3. pro 2019.

    Getting started with machine-learned ranking using Vespa

    Poništi
  13. 29. stu 2019.

    Also makes me think we have failed at informing the world about the feature set in Vespa and the scale we operate it at Verizon Media.

    Prikaži ovu nit
    Poništi
  14. 29. stu 2019.

    Came across this and realized that the features which amazon team contributed upstream to Lucene has been in Vespa for ages. Multi threaded indexing, custom term freq contribution and multiple threads per query. I’ve written about these features and more

    Prikaži ovu nit
    Poništi
  15. proslijedio/la je Tweet
    28. stu 2019.

    If you have it's not too late to make yourself a state of the art and infinitely scalable e-commerce site in time for Black Friday. explains how in

    Poništi
  16. 27. stu 2019.

    Working on new sample applications for this morning. A simple neural network trained with , exported in format and deployed to Vespa for serving. Text embeddings from Google's Universal Sentence Encoder. Work building on

    Poništi
  17. 26. stu 2019.

    Very nice introduction to text classification using the [cls] token output embedding from distilled BERT as the text feature tensor for binary classification task. Similar technique can be used in LTR (Learning to Rank).

    Poništi
  18. proslijedio/la je Tweet
    18. stu 2019.

    “With these GPU optimizations, we were able to use 2000+ Azure GPU Virtual Machines across four regions to serve over 1 million BERT inferences per second worldwide” Bing, using (distilled, 3-layer) BERT in production. via

    Poništi
  19. proslijedio/la je Tweet
    5. stu 2019.

    Awesome updates - Nearest Neighbor and Tensor Ranking, Optimized JSON Tensor Feed Format, Matched Elements in Complex Multi-value Fields, Large Weighted Set Update Performance, & Monitoring Support. Learn more: .

    Poništi
  20. 5. stu 2019.

    It might fix your search problems. BERT based ranking models is dominating the leaderboard for passage ranking, beating the official BM25 baseline significantly (0.39 versus 0.16) on MRR. . Tensors + Neural Network evaluation => .

    Poništi

Čini se da učitavanje traje već neko vrijeme.

Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.

    Možda bi vam se svidjelo i ovo:

    ·