The neocortex is likely to process sparse representations in a hierarchy with information close to the raw sensor input appearing in lower levels and abstract concepts being appearing in higher levels. However, I'm interested in the exact nature of information being passed upwards.

  1. There are some gating mechanisms in the neocortex. I could imagine that they prevent obvious information from being passed up. For example, since it was already reported and didn't changed. Much like workers within corporate hierarchies only report the same status once.
  2. Similarly, it would make sense to pass up information that a lower layer doesn't understand and can't handle itself. This would also be comparable to workers at a company who ask their managers for decision they can't make.
  3. If planning happens within the hierarchy of the neocortex, higher layers might hand down high levels options for the lower levels to unroll and simulate. In that case, the result might be passed up afterwards.
  4. Encoding continuous input streams, attractor patterns are likely to emerge within levels of the hierarchy. I could imagine that the type of information send upwards changes when a level settles in a stable pattern.

Is there any evidence for or against these guesses? What else know about the nature of information that's passed up within the neocortex?

  • $\begingroup$ Re: 1., do you mean something like predictions from agranular neocortex "explaining away" incoming sensory input (responsible for phenomena like repetition suppression)? $\endgroup$
    – mrt
    Jan 18, 2016 at 2:27
  • $\begingroup$ @mrt Is there a theory about that? I think it would make sense for a level to only send upward what cannot be explained from the top-down information. $\endgroup$
    – danijar
    Jan 18, 2016 at 9:50
  • $\begingroup$ I'm thinking about standard predictive coding models (e.g., Clark, 2013), where predictions flow up from agranular, deep cortex (low dimensional; top of hierarchy) and prediction error flows down from granular, superficial layers (high dimensional; bottom of hierarchy) (e.g., Chanes & Barrett, 2015). But I'm not sure if this is helpful or not! :) $\endgroup$
    – mrt
    Jan 18, 2016 at 17:50
  • $\begingroup$ @mrt That helps a lot. $\endgroup$
    – danijar
    Jan 18, 2016 at 17:58

1 Answer 1


Your guess 1 basically sounds like habituation:


Per your clarification in your comment, 2 sounds like you are generally talking about the role of prediction error in learning. There's a lot of work on this. Neural network models generally learn by modifying the connection strengths in response to error. The most well-known algorithm for this is known as backpropagation. Although this is an error signal propagating down a network, not up it, as you proposed. There's also the Rescorla-Wagner model of learning, another example of error-driven learning.

3 sounds like top-down processing:


Regarding 4, neural firing patterns definitely settle into attractors. This has been pretty extensively modeled:


  • 1
    $\begingroup$ Thanks for your answer. With the second point, I was referring to some kind of surprise or misprediction of next input. Is there anything known about those inputs or their prediction error terms being fed upwards? More generally, I was looking for an overview or review of what's known about the information passed upwards. Do you know a resource for that? $\endgroup$
    – danijar
    Jan 18, 2016 at 1:34
  • $\begingroup$ Ah, I'll edit my answer with something about that. $\endgroup$ Jan 18, 2016 at 1:35
  • 1
    $\begingroup$ I'm from machine learning but error backpropagation seems biologically implausible. I'm also aware of CHL, XCAL, etc. But before diving into learning rules, I'm interested in what information flows between the levels. Does the local prediction error at a level affect the information that gets passed upwards? $\endgroup$
    – danijar
    Jan 18, 2016 at 1:54
  • $\begingroup$ Yeah, I've also heard from one of my profs, who is himself a connectionist, that it's been a challenge to find a biological basis for backprop. Not sure about your question about error affecting how information flows. $\endgroup$ Jan 18, 2016 at 2:01
  • $\begingroup$ @danijar, by the way, if you like my answer, can you 'accept' it so I can get those sweet, sweet points? $\endgroup$ Jan 18, 2016 at 2:57

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.