Dmitry Kobak

@hippopedoid

Born but to die and reas'ning but to err

Tübingen
Vrijeme pridruživanja: prosinac 2019.

Tweetovi

Blokirali ste korisnika/cu @hippopedoid

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

  1. Prikvačeni tweet
    prije 21 sat

    Our huge Patch-seq effort is out on bioRxiv: . 1320 neurons in mouse motor cortex patched, sequenced, and mapped to a scRNA-seq atlas. 642 reconstructed morphologies! Our running joke was that it was a bit like catching Pokemons. (1/n)

    Prikaži ovu nit
    Poništi
  2. proslijedio/la je Tweet
    prije 21 sat

    proud of all papers from my lab - but this one in particular! patch-seq on >1,300 neurons from mouse motor yields ephys & anatomy for >70 types. with and many others!

    Prikaži ovu nit
    Poništi
  3. prije 20 sati

    This was submitted back-to-back with a similar Patch-seq effort from the Allen institute, independently done in the visual cortex. We only saw each other's drafts two weeks ago. They only study interneurons but have more of them: . (6/n)

    Prikaži ovu nit
    Poništi
  4. prije 20 sati

    Perhaps most importantly, we find a lot of continuous morpho-electric variation *within* each family, as one traverses the transcriptomic landscape. Given that transcriptomically the t-types are often not well-separated either, we question that t-types are really discrete. (5/n)

    Prikaži ovu nit
    Poništi
  5. prije 20 sati

    We to t-SNE and kNN classification on electrophyisological, morphological, and combined features, and find that big neuronal families (Pvalb, Sst, Vip, Lamp5, IT, PT, CT) have largely distinct morpho-electric phenotypes with little overlap.

    Prikaži ovu nit
    Poništi
  6. prije 20 sati

    This dataset allowed us to provide a morphological and electrophysiological description of most transcriptomic cell types in mouse motor cortex. This is a figure from the paper itself, but there is also a supplementary PDF showing all our cells for each t-type. (3/n)

    Prikaži ovu nit
    Poništi
  7. prije 21 sat

    Most neurons were patched by in in Houston. All sequencing performed by in in Stockholm. Computational analysis done in the lab of in Tuebingen. Reference transcriptomic atlas provided by the Allen institute. (2/n)

    Prikaži ovu nit
    Poništi
  8. 30. sij

    I must confess that I could never *really* understand VAEs... so ended up thinking about them as autoencoders that inject noise into the bottleneck as some sort of implicit regularization strategy. Well, turns out this might actually be a reasonable way to think about them!

    Poništi
  9. 30. sij

    A fascinating account of the life in Bell Labs in the late 1990s and the Schön scandal (a postdoc fabricated data and published 16 [!!] first-author Nature+Science papers on semiconductors within two years; his PI later claimed he did not know and did not realize).

    Poništi
  10. proslijedio/la je Tweet
    23. pro 2019.
    Prikaži ovu nit
    Poništi
  11. proslijedio/la je Tweet
    14. sij

    My blog post from 2017 which I converted to a bioRxiv pre-print last year has now been peer reviewed and published in Nature Biotechnology : "Droplet scRNA-seq is not zero-inflated"

    Poništi
  12. proslijedio/la je Tweet
    11. sij

    Brownian tree with one billion nodes.

    Prikaži ovu nit
    Poništi
  13. proslijedio/la je Tweet
    28. pro 2019.

    Here's the corrected version for those following!

    Prikaži ovu nit
    Poništi
  14. 8. sij

    Somebody posts a catchy video of a chimp with a "super-human" short-term memory performance; 100K likes, 35K retweets. Somebody else replies that this paper has been debunked, humans can easily do better; 150 likes, 10 retweets. THREE ORDERS OF MAGNITUDE less attention.

    Poništi
  15. proslijedio/la je Tweet
    25. pro 2019.

    I created an interactive visualization of 150000+ butterflies in the using deep learning, t-SNE and data from ! Click here for the interactive version:

    Prikaži ovu nit
    Poništi
  16. proslijedio/la je Tweet
    20. pro 2019.

    This is my experience too. tSNE is pretty good given a suitable initialization. And given proper parameters () it can be really great. UMAP is pretty good too but often uses too much whitespace.

    Poništi
  17. proslijedio/la je Tweet
    20. pro 2019.

    Not surprised that UMAP's preservation of global structure is entirely due to initialization. Force layouts (used by UMAP & t-SNE) do local updates only. A contrario, Laplacian eigenmaps and PCA use the global structure. Init globally and fine-tune locally is the wining strategy.

    Poništi
  18. 20. pro 2019.

    But to decide which algorithm is more faithful to the single-cell data, further research is needed. Our Comment argues that Becht et al. paper does not answer that. (10/10)

    Prikaži ovu nit
    Poništi
  19. 20. pro 2019.

    Just to be clear: this is *not* an attack on UMAP! I think UMAP is great :-) But I also think t-SNE is great. And there is plenty of room for further improvements and for better conceptual understanding of this whole family of embedding methods. (9/10)

    Prikaži ovu nit
    Poništi
  20. 20. pro 2019.

    This is in agreement with the recommendation to use PCA initialization (rather than random initialization) for t-SNE made in the recent paper by and me: . (8/10)

    Prikaži ovu nit
    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:

    ·