Ankur Handa

@ankurhandos

Robotics. Previously, Research Scientist at , post-doc at Cam, and PhD at Imperial.

London
Vrijeme pridruživanja: prosinac 2010.

Medijski sadržaj

  1. 28. sij

    source The whole speech is amazing to watch.

    Prikaži ovu nit
  2. 28. sij

    Infrastructure work can be tiring and demanding but it is very important to do so that interesting questions can be asked or answered. Apollo astronaut armstrong had a very funny way to put it.

    Prikaži ovu nit
  3. 26. sij
    Odgovor korisnicima
  4. 26. sij

    This is how they obtain the matrix from two 3D vectors.

    Prikaži ovu nit
  5. 26. sij

    They do a series of experiments to highlight the accuracy of this 6D representation. source code in pytorch is here:

    Prikaži ovu nit
  6. 26. sij

    Quaternions and Euler angles are discontinuous and difficult for neural networks to learn. They show 3D rotations have continuous representations in 5D and 6D, which are more suitable for learning. i.e. regress two vectors and apply Graham-Schmidt (GS).

    Prikaži ovu nit
  7. 25. sij

    It still amazes me that someone came up with such a crazy way of landing a space shuttle. This is a nice insightful and hilarious video explaining how they used to land a space shuttle.

    Prikaži ovu nit
  8. 24. sij

    Interestingly they also try Pareto optimal multi-task learning for their problem.

    Prikaži ovu nit
  9. 24. sij

    They combine a robust (shift and scale invariant) function that compares prediction against ground truth together with another loss function that penalises the gradients.

    Prikaži ovu nit
  10. 24. sij

    Impressive results on single image depth estimation by Ranftl et al

    Prikaži ovu nit
  11. 19. sij
    Odgovor korisniku/ci

    If you plug t = 1/pi in the last two equations you arrive at the result. This one extends from the Basel problem also used to obtain the identity that sum of inverse squares up to infinity = pi^2/6 source:

  12. 14. sij
    Odgovor korisniku/ci

    I saved it from a long video here I knew this was such a gem that it would be worth sharing in future.

  13. 13. sij
    Odgovor korisniku/ci
  14. 7. sij

    This is a nice short tutorial on normalising flows by

  15. 5. sij
    Odgovor korisniku/ci

    This is for procedurally generated terrains (I found it on HN)

  16. 3. sij
    Odgovor korisnicima

    So, the trick is that since alpha is also a part of normalising constant (Z) you have to also estimate that so that you can compute the derivatives otherwise the loss function collapses. Without the Z it doesn't work.

  17. 3. sij
    Odgovor korisniku/ci
  18. 3. sij

    Robust functions are useful in regression because they can cull the outliers from the optimisation process. Since the alpha is a free parameter they can learn that.

    Prikaži ovu nit
  19. 3. sij

    similarly the well known cauchy

    Prikaži ovu nit
  20. 3. sij

    however, huber is generally more resilient to outliers.

    Prikaži ovu nit

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:

    ·