Visual Information Processing
Stemming from a Conversation with Adam Reeves (NU Dept. Psychology, 4/21/06).
Adam is attending a conference this summer in Germany on visual masking. Of particular interest
is a Dutch fellow who records from cortex in monkeys as they plan a visual trajectory from a focus point to a remote visual
target. The monkeys are trained to follow different visuomotor routes to the target, depending e.g. on the color of
the cue. What is interesting is that there is topographic activation of cortical paths within lower level visual
areas during the preparatory phase, i.e. before the monkey is actually cued to make the eye movement path. This preparatory
activity can be detected as far "down" as V2 (the seond cortical visual area), but is not evident in V1, the "first" region
of cortex to receive visual information.
This discussion stemmed from my comment that prediction was a massive component of cortical activity (IAW J. Hawkins)
and that such predictions might play a role in masking. Adam said that this general idea has been around since at least
the 1960s in psychological circles (I forget the term used). Of course the details of cortical microarchitecture have
been determined in much more recent years and Hawkins' work attributes specific functions to connections within and between
One key thing that Adam noted was that much more progress has been made in looking at feedfoward processes (in vision
and to a lesser extent audition and olfaction), whereas the establishment of clear functional roles for feedback pathways
has been the achilles heal of the neocortex community. This is why the V2 priming work and work by Hawkins and
by Paul Adams are of such interest to me, because they are starting to get at what cortex really does. Accelerating developments
in this realm will lead, I suspect, to a deeper understanding of what intelligence is and why our brains are so powerful.
But even these explanations are inadequate and Adam also stated that visual information gets to prefrontal cortex
more quickly than it gets to lower level visual areas such as IT (within 60 msec, i.e. without enough time for the visual
information to traverse V1, V2, V3, etc. etc.). Assuming that these facts/my recollection are correct, this suggests
the presence of what might be called a metaoligarchical organization, based
on distributed routing and massive simultaneous interconnections. The problem is how you get all of these divergent
data streams to work in coordinated and hierarchical manners.
But I think that we are approaching a new vista of cortical operations, that we will undertand
these things and that such results will inform the AI community as to the kinds of operations that make brains smart.
Then it becomes a technical problem to incorporate such findings into new digital architectures, and once this happens, since
these architectures are freed from the unnecessary biological baggage that slow down biocomputing, these new machines will
be blindingly fast and powerful. See the what AI can do Now page, and then just try and extrapolate a litle bit.
Indeed, humanity is very much, on the threshold of a scream,
Information Content: Storage and Retrieval
Adam also described an experiment where a subject was exposed to something like 600 visual scenes at a rate of 6 per
second, and then later challenged with visual scenes and required to say novel or "saw already". [numbers may be off
some; should be in ballpark]. Subjects could identify priorly viewed scenes with 90% accuracy, when given the entire
scene, but if asked to identify smaller components of the screen, quickly fell to chance performance. This suggests
that for brief scene exposures, the cumulative information derived from each full scene was sufficient for effective
recognition memory. In contrast, the information acquired from small elements of individual scenes was insufficient
for recognition. This indicates that at some level in the visual hierarchy associated bits of information from
a briefly displayed scene are somehow binned and become sufficient to make comparisons with scenes shown after a
This is curious in several respects, most notably a seemingly large gap between our short term memory for things like
digits (7 to 10 for most people) vs. a tremendously larger ability to recognize things that we saw. This seems to be
a fundamental feature of our CNS basic design, where we have tremendous pattern recognition abilities and vastly inferior
conscious recall. Presumably this state reflects an optimization of general intelligence neural processing protocols,
otherwise evolution would have arrived at a different solution. It further suggests that this organization reflects something
fundamental in the design of intelligent beings, but I cannot say what this is. But even lacking such insight,
what I can say is that if we mimic the salient parts of this in silico, we will build smarter machines, even if we do not
know how or why they are smarter.