Markus Frey

@CYHSM

PhD candidate in the labs of & . in vitro -} in vivo -} in silico

Vrijeme pridruživanja: kolovoz 2015.

Tweetovi

Blokirali ste korisnika/cu @CYHSM

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

  1. Prikvačeni tweet
    11. pro 2019.

    Happy to share my first PhD project as a . We introduce DeepInsight: A general framework for interpreting wide-band neural activity. Can be applied on unsorted neural data + calcium imaging to decode continous behaviour or stimuli. Paper:

    Prikaži ovu nit
    Poništi
  2. proslijedio/la je Tweet
    25. sij

    The dendritic spatial code: branch-specific place tuning and its experience-dependent decoupling

    Poništi
  3. 24. sij

    Congratulations Dr. Nau! You set the bar very high with a fantastic defense.

    Poništi
  4. proslijedio/la je Tweet
    22. sij

    This Friday, Jan 24, I will be defending my . Want to join the livestream? Trial lecture, 10:15am, CET: The Role of the Hippocampus in Imagination & Dreaming Defense, 12:15pm: Perception & the Cognitive Map: deriving a stable world from visual inputs Stream below👇

    Prikaži ovu nit
    Poništi
  5. proslijedio/la je Tweet
    13. sij
    Poništi
  6. 12. sij

    As a side note, using the cosine as an activation function is also sufficient to get 1% error on MNIST (see book section 6.3.3). Plotting done in python with

    Prikaži ovu nit
    Poništi
  7. 12. sij

    However, a periodic activation function like cos, sin or tan finds the solution to the XOR problem, possibly because of the many local minimas (loss landscape shown for tangens)

    Prikaži ovu nit
    Poništi
  8. 12. sij

    Interestingly the same network using the rectified linear unit activation (ReLU) does not find the solution. The minima shown in this loss landscape is above the critical value of 0.37.

    Prikaži ovu nit
    Poništi
  9. 12. sij

    This is the loss landscape of the neural network solving the XOR problem with the dCaAPs activation function. The plateau explains why gradient descent has difficulties finding the minima using randomly initialized weights.

    Prikaži ovu nit
    Poništi
  10. 10. sij

    To run the code directly in the browser, use this colab notebook: For more about dCaAPs see the paper by Gidon et al. and :

    Prikaži ovu nit
    Poništi
  11. 10. sij

    A recent Science paper by Gidon et al. describes calcium-mediated dendritic action potentials (dCaAPs). Here is a TensorFlow/Keras implementation solving the XOR problem in a single artificial neuron using dCaAPs activation function:

    Prikaži ovu nit
    Poništi
  12. 16. pro 2019.
    Poništi
  13. proslijedio/la je Tweet
    11. pro 2019.

    Together with , we developed DeepInsight, a framework for decoding continuous behaviour and stimuli from raw neural data. This model is able to discover new neural correlates in hippocampal recordings and outperforms traditional decoding methods.

    Poništi
  14. 11. pro 2019.

    Code is available here: Thanks to all the people involved in this project, , , Alice O'Leary, , , , and

    Prikaži ovu nit
    Poništi
  15. proslijedio/la je Tweet
    31. lis 2019.
    Poništi
  16. 21. lis 2019.

    I am also happy to talk about how we can interpret convolutional neural networks and make machine learning more accessible for neuroscience researchers.Come by from 8-12 at BB79 .Work together with

    Poništi
  17. 7. lis 2019.

    This was a fun side project with on Tumor segmentation. Code for the model architecture will be available here soon:

    Poništi
  18. proslijedio/la je Tweet
    30. ruj 2019.

    Excited to share our new : Behavior-dependent directional tuning in the human visual-navigation network: . Environmental coding in V1-RSC-PHC-EC-HPC pathway depends on behavior & successful cognitive mapping. /w , &

    Poništi
  19. 28. ruj 2019.

    I really enjoyed this service after some really fun but also exhausting weeks on the Camino del Norte.

    Poništi
  20. 11. lip 2019.

    Forgot to scale the decoding errors while plotting them across the arena and this is how you end up as a modern artist. Highest bid gets the code.

    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:

    ·