I agree, higher level expectations could be used to boost (random spark) lower level regions input when expectation of the higher level region is not met. Also, if input is very clear and new for higher level regions, then higher level region expectation shouldn't boost much of the lower level region input. we could boost the lower level regions input based on some condition 1. If the sparse activity is very low in lower level regions i.e 0.1% 2. If the higher level regions expectations much higher, might be the lower level regions very low sparse activity triggered the much expectations in higher regions and inturn the higher regions will boost (randomness) in lower regions .
On Mon, Sep 9, 2013 at 5:02 AM, Fergal Byrne <[email protected]>wrote: > My picture of the information travelling around in the neocortex and the > CLA is that the SDR's are all "best guesses" which encode the transmitter's > "most probable" interpretation of the state of the world (it's bit of it). > The SP's job is to constantly try and improve the guess by rewarding good > guessing columns and inhibiting the more spurious, but it's always likely > that columns will activate when the "real" data does not justify it, and > vice versa. > > In a hierarchy, higher level regions are using this guess to predict their > own higher-level summaries for their own purposes, but also sending > "encouraging" feedback to the lower-level regions, essentially saying "I > think you're onto something, keep it up." This manifests itself in > enhancement of the perception itself (ie it feels like you're actually > seeing or hearing better). > > I've been calling this "feed-round" because it envisages the feedback > neurons going round to the input level of the lower-level region, adding to > the feedforward data. > > As I'm typing this, I actually hear my words, so this mechanism is also > creating perceptions (or apparent perceptions). > > I believe that the appearance of extra randomness in perception is not > random at all, in that it is likely driven by erroneous hierarchical > feed-round in a situation of true input poverty. Feature detectors are > struggling to fire, and they'll sometimes over-eagerly use higher-level > expectations to create mirages at lower levels. Very low-level mirages > appear as apparently random sparks. > > Regards, > > Fergal Byrne > > > > > On Sun, Sep 8, 2013 at 11:35 PM, Ramesh Ganesan <[email protected]>wrote: > >> Cool :) Its worth to investigate this randomness behaviour. we might have >> two types of randomness >> 1. randomly increase the input to some upper limit to cortex to fill the >> rest of the things OR >> 2. hierarchically regulated randomness for the lower level region's input >> >> >> >> On Sun, Sep 8, 2013 at 9:43 AM, Fergal Byrne <[email protected] >> > wrote: >> >>> Very interesting. I'm pretty sure that we're feeding images/sounds into >>> a very low-level region in the hierarchy as a result of predictions at a >>> higher level, and that we perceive this "interference" as noise or >>> scintillation. We do this when the actual data does not present enough >>> information for us to lock on to a pattern or sequence with sufficient >>> confidence to see or hear what we actually see or hear. Blind people use >>> this to navigate, using their aural expectation/prediction of the echo >>> characteristics of a learned space to figure out where they must be in that >>> space. There's clearly a capability to compare predicted patterns (coming >>> from up the hierarchy as well as sequence memory at the current level) with >>> incoming data to coerce perception into a "lock-on" to a particular >>> candidate perception. As this is in progress, we see fireflies and sparks >>> as noise, but these fade out as they're incorporated into the predicted or >>> expected pattern. >>> — >>> Sent from Mailbox <https://www.dropbox.com/mailbox> for iPhone >>> >>> >>> On Sun, Sep 8, 2013 at 11:20 AM, Patrick Higgins >>> <[email protected]>wrote: >>> >>>> The addition of noise to push low level signals >>>> above the noise floor is a known technique for >>>> making these signals perceivable. It essentially >>>> boost some information just below the floor above >>>> it, and given the brain is adept at handling noise, >>>> it improves the overall dynamic range of the system. >>>> >>>> Scintillating vision at night with dark adapted eyes >>>> is another example of this phenomenon. Adding in >>>> a little familiarity with the environment makes it easy >>>> make to the bathroom at night without any lights. But >>>> the cost of this improvement is clearly a degraded >>>> image with missing edges and false objects hits as >>>> well as an undulating video lacking smooth continuity. >>>> >>>> I find it interesting to learn how this process is >>>> implemented at a synaptic level, fascinating. >>>> I see it as a refinement, something that could be >>>> added based on the application. Such as cleaning >>>> up weak noisy radio transmissions, audio or video >>>> to bring previously unattainable information out. >>>> >>>> (never fails to trip me up that one of those words >>>> ends in dio and the other deo even tho they sound >>>> the same) >>>> >>>> This blog explains how adding (good) noise before >>>> something is encoded improves dynamic range. This >>>> might be an argument for playing with this idea at the >>>> encoder level. >>>> >>>> http://blog.discmakers.com/2013/01/dithering-adding-good-noise-to-improve-your-home-recordings/ >>>> >>>> Taking this a little further, I use image stacking >>>> to reduce the noise and improve the dynamic >>>> range of my astrophotos. Its a common technique >>>> used by astrophotographers, taking many frames >>>> of the same object taken at different times and using >>>> photoshop or special image stacking software to align >>>> them and average each pixel. This has the effect of >>>> canceling the noise to some degree and adding to the >>>> signal. Contrast is greatly improved. Using this technique >>>> at the encoder level would also be a nice way to remove >>>> some unwanted noise (possible after it was added in for >>>> reasons described above). I'm wondering if these >>>> techniques in tandem would be worth investigating. >>>> >>>> Thanks TIm! >>>> >>>> >>>> >>>> Patrick >>>> >>>> >>>> >>>> >>>> On Sep 8, 2013, at 12:22 AM, Tim Boudreau wrote: >>>> >>>> > Thought this might be of interest on this list: >>>> > >>>> > "Release of neurotransmitter is an inherently random process, which >>>> could degrade the reliability of postsynaptic spiking, even at relatively >>>> large synapses. This is particularly important at auditory synapses, where >>>> the rate and precise timing of spikes carry information about sounds." >>>> > http://www.ncbi.nlm.nih.gov/pubmed/24005293 >>>> > >>>> > It suggests that the phenomenon that introducing dither noise into an >>>> audio signal can improve intelligibility of low amplitude signals (pushing >>>> enough samples above the noise floor that the brain will fill in the >>>> blanks) is also inherently present in the brain. >>>> > >>>> > -Tim >>>> > >>>> > -- >>>> > http://timboudreau.com >>>> > _______________________________________________ >>>> > nupic mailing list >>>> > [email protected] >>>> > http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org >>>> >>>> >>>> _______________________________________________ >>>> nupic mailing list >>>> [email protected] >>>> http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org >>>> >>> >>> >>> _______________________________________________ >>> nupic mailing list >>> [email protected] >>> http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org >>> >>> >> >> _______________________________________________ >> nupic mailing list >> [email protected] >> http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org >> >> > > > -- > > Fergal Byrne > > ExamSupport/StudyHub [email protected] http://www.examsupport.ie > Dublin in Bits [email protected] http://www.inbits.com +353 83 > 4214179 > Formerly of Adnet [email protected] http://www.adnet.ie > > _______________________________________________ > nupic mailing list > [email protected] > http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org > >
_______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
