That's an interesting way to think about hierarchy, trying to determine a
formulaic way to structure the hierarchy.  This kind of thought experiment
can be helpful.

 

However, there is a lot of evidence to suggest real cortex isn't working on
such a principle.  One piece of evidence is that different mammals such as
dogs, cats, monkeys, and humans have a similar set of sensors but different
number of regions and levels in the hierarchy.  Another piece of evidence is
that most mammals have primary and secondary sensory regions, such as V1 and
V2, where the secondary region is receiving input almost exclusively from
the primary one.  The convergence onto V2 is coming from parts of V1, not
two different regions.  A monkey has dozens of cortical regions.

 

I don't think of the retina as a single sense.  I think of it as an array of
1M sensors, that is the number of axons in the optic nerve.  The auditory
nerve is an array of 30K sensors.   There are about 1M somatic sensors.  The
input to the cortex is not just a handful of senses but millions of sensors
arranged in topological arrays.  With this view a simple calculus doesn't
work.

 

The goal of the cortex is to form representations of its inputs.  We want
those representations to be independent of what sensors are sensing the
object(s) and where on the sensor arrays the object is being sensed.

 

In theory, if we had one region of almost unlimited capacity (very large
number of mini-columns) and we had an almost unlimited amount of training
data then a  single region without hierarchy could learn everything.
Hierarchy is needed to make a practical system.

 

Consider V1.  It receives the input from the optic nerve.  The spatial
pooler wants to form representations of the spatial patterns coming from the
retina.  There is a limited capacity in V1 and a lot of patterns on the
optic nerve so each V1 SP bit learns to represent only a small part of the
retina.  As the number of spatial patterns seen by the animal declines then
each bit in the SP will represent larger patterns, representing larger parts
of the input area.  Similarly, if we increased the number of SP bits (number
of minicolumns in cortex) then each SP bit would learn to recognize larger
input patterns.

 

There is an equilibrium point where the SP forms good representations over
some part of the input space.  For humans and monkeys V1 neurons represent
about 1 degree of visual field.  If we severely restricted the number of
patterns that the eye could see then neurons in V1 would represent larger
areas of the retina.  The same sort of give and take between resources and
input pattern complexity exists in the sequence memory part of the CLA.

 

Hierarchy allows the system to get to representations that are independent
of sensor location in practical time with practical resources.  Each level
in the hierarchy builds on the previous.  How many levels in the hierarchy
are required?  It depends on how many columns and cells exist at each level
and how complex are the patterns in the input.

 

There is a tradeoff of using a hierarchy.  The cortex gives up the ability
to learn many possible patterns such as a visual pattern where pixels on the
left side of an image are tightly correlated with pixels on the right side
of the image.  However, practically these kinds of patterns don't occur.

 

I hope that helped and didn't make things worse!

Jeff

 

 

 

From: nupic [mailto:[email protected]] On Behalf Of
mariolakakis .
Sent: Sunday, February 16, 2014 2:01 PM
To: [email protected]
Subject: [nupic-discuss] What dictates the exact formation of the cortical
hierarchy?

 

I know that the goal is efficiency in training and storage but how is the
hierarchy in the neocortex done exactly? Is it a result of a mathematical
equation? Or is it the throw it in the wall and see if it sticks process of
evolution?

 

My mathematical theory is very simple and it's based on binomial
coefficients. Let's say that the human body consist of a K number of
sensors. The number of regions X in the hierarchy should be equal to X = !K
/ (2! * (K - 2)!) + K (for each sensor). That's the number of all possible
duplets of sensors with no repetitions. This equation creates a pool of the
highest variety but least density that we can use for representations. And
it also explains the tree like shape of the hierarchy and why it converges
and diverges and you up and down.

 

For example, if the human body consisted of just three sensors:

S1 = Optic, S2 = Acoustic, S3 = Touch

 

The number of regions would be X = 3 * 2 / (2 * 1) + 3 = 3 + 3 = 6

1. R1 = S1 (Optic)

2. R2 = S2 (Acoustic)

3. R3 = S3 (Touch)

4. R4 = R1, R2 (Optic + Acoustic)

5. R5 = R2, R3 (Acoustic + Touch)

6. R6 = R4, R5 (Optic + Acoustic + Touch)

 

Let's consider the part of the neocortex that handles language. The number
of characters in the alphabet is much smaller than the number of words and
the number of words is tiny compared to the number of phrases. This simple
observation makes me assume that the hierarchy in the brain is like this:

 

1. Letters (Highest level concepts)

2. Words

3. Phrases

 

Using a single region we would have to assign columns to letters and
sequences of cells to words.

 

For example, the words "god" and "dog" would share the same spatial pattern
but different temporal patterns. Since, the higher regions get only spatial
patterns from below how does the distinction of those two gets communicated
above? What happens if the word has multiple identical letters? Do the cells
in a column connect to other cells in the same column? For example, the word
"good" has two "o"s.

 

To summarise, if one region wasn't enough and I wanted to reconstruct the
human neocortex based on a K number of sensors, how would I know how many
regions I would need, and in what way should I connect them to make it all
work? Thats a problem you will face in the future. One day, one region won't
be enough.

 

I 've implemented a huge part of the CLA in Xcode and got it running on an
iPhone, I've seen a dozen videos of Jeff Hawkins' presentations and I 've
also bought the book On Intelligence but I haven't found any answers to
these questions.

 

I'm counting on you guys. :)

_______________________________________________
nupic mailing list
[email protected]
http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org

Reply via email to