Micha Heilbron

@m_heilb

PhD-candidate , / . Computational cognitive neuroscience

Nijmegen/Amsterdam, NL
Vrijeme pridruživanja: lipanj 2015.

Medijski sadržaj

  1. 24. sij
    Odgovor korisniku/ci
  2. 16. sij

    When we tested for such a hallmark, we found exactly this pattern of effects in VWFA, pMTG and IFG — all key areas of the reading network … suggesting these areas might be the source of the enhancement

    Prikaži ovu nit
  3. 16. sij

    We expected that in a source area the neural activity covaries with the amount of letter information in early visual cortex, such that letter information is higher when the area is more active, and vice versa. & this relationship should hold within, not just between conditions

    Prikaži ovu nit
  4. 16. sij

    Finally, we asked what the neural source of this enhancement effect might be Early visual cortex itself does not know anything about words or letters, so it has to come 'from the top-down' -- from some higher-order brain area But where?

    Prikaži ovu nit
  5. 16. sij

    (For the aficionados, yes we did make sure that the result was not a fluke created by our specific definition of ‘early visual cortex’. The same pattern was consistently found over a large range of ROI definitions)

    Prikaži ovu nit
  6. 16. sij

    Strikingly, we indeed find the enhancement predicted by the top-down model: letters are more easily decoded from early visual cortex when embedded in a word. Reassuringly we find the same subtle but consistent enhancement with both MVPA techniques

    Prikaži ovu nit
  7. 16. sij

    We also probe sensory letter representations using a complementary technique based on correlations between brain activity patterns

    Prikaži ovu nit
  8. 16. sij

    Our idea was that if word contexts enhance the perception of letter stimuli, then we should observe an enhancement of sensory information in visual cortex To probe sensory information, we try to predict the middle letter (U or N) based on brain activity patterns in V1-V2

    Prikaži ovu nit
  9. 16. sij

    We can quantify how sharp the network ‘sees’ the middle letter and call that ‘representational strength’ If we do that for every stimulus we see that only in the top-down model word context enhances letter representations Now, how can we test for such enhancement in the brain?

    Prikaži ovu nit
  10. 16. sij

    In the feedforward model, the quality of a letter’s representation depends only on the the line segments comprising that letter But in the top-down model, information from the word-level is sent back to the letter level—allowing context to enhance perception ‘from the top-down’

    Prikaži ovu nit
  11. 16. sij

    We can formalise our hypotheses by simulating the experiment with a variant of the famous neural net architecture by Rumelhart & McClelland The model first integrates lines to recognise letters, and then integrates letters to recognise words…

    Prikaži ovu nit
  12. 16. sij

    To get a feel of the experiment, this is what participants would see when lying in the MRI scanner (if the stimuli would have been English rather than Dutch…)

    Prikaži ovu nit
  13. 16. sij

    We presented participants streams of 5 letter words or nonwords (unpronounceable strings) embedded in Gaussian noise In each stream, the middle letter was fixed (a U or an N) while the outer letters varied, forming either word or nonword contexts

    Prikaži ovu nit
  14. 16. sij

    … namely: If the behavioural advantage indeed reflects a perceptual enhancement of letter stimuli, then it should be accompanied by an enhancement of sensory information in visual cortex (which processes these stimuli) already -- which is precisely what we set out to test 😎

    Prikaži ovu nit
  15. 16. sij

    Alternatively, top-down models propose that linguistic knowledge can enhance perception in a top-down fashion So under this account, word contexts don’t just help you guess the letters you’re seeing, but can also make you *see them better*

    Prikaži ovu nit
  16. 16. sij

    Historically, there have been 2 accounts for this effect Bottom-up models claim that the word advantage is only a post-perceptual advantage in guessing the correct letters So under this account, you *see* letters just as well/poorly—words only help you guess what you’re seeing

    Prikaži ovu nit
  17. 16. sij

    Our starting point is that letters are more easily recognised when embedded in a word We've all experienced this effect, for instance when navigating in bad weather — it's easer to read a word or name (like a road sign) than a random string (like a licence plate) But why?

    Prikaži ovu nit
  18. 6. pro 2019.

    It’s pretty slow (samplers will be samplers) but it does the job and the output is quite hilarious Here’s a positive and negative continuation of the same sentence — first try, not cherry picked

    Prikaži ovu nit
  19. 14. ruj 2019.

    Stoked to present my latest PhD project at today! Using a deep language model (GPT2) we find feature-specific linguistic predictions during passive listening to natural narrative Catch me @ poster 1A-4 or check it out below

    Prikaži ovu nit
  20. 13. ruj 2019.

    looking for ‘theory’ among all the neural networks in the papers...

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:

    ·