On Intelligence Book Summary Hawkins
<p class=default>
02:43 PM, June 22, 2006
</p><p>
Chapter 6: How the Cortex Works
<br/>
p 106
<br/>
Trying to figure out how the brain works is like solving a giant
<b>jigsaw puzzle</b>
. (great analogy).
<br/>
You can approach it in one of two ways.
<br/>
Using the "top-down" approach, you start with the image of what the solzed puzzle should
look like,
<br/>
and use this to decide which pieces to ignore and which pieces to search for. The other
approach is "bottom-up", where
<br/>
you focus on the individual pieces themselves.
<br/>
(This is how he is proceeding to try to understand the brain; also it is an analogy to
how the brain actually works).
<br/>
107 Understanding the brain is like this: we have thousands of pieces and some are porrly
shaped and many won't even be in the solution.
<br/>
To undertstand the brain we need a top-down theory.
<br/>
THE MEMORY PREDICTION MODEL IS IT>
<br/>
TO MAKE PREDICTIONS YOUR CORTEX NEEDS A WAY TO MEMORIZE AND STORE KNOWLEDGE ABOUT
SEQUENCES OF EVENTS.
<br/>
TO MAKE PREDICTIONS OF NOVEL EVENTS< THE CORTEX MUST FORM INVARIANT REPRESENTATIONS.
<br/>
Your brain needs to create and store a model of the worls as it is, independent from how
you see it under changing circumstances.
</p><p>
Christof Koch at Caltech has found cells that fire whenever a person sees a picture of
BILL CLINTON.
</p><p>
Cortex is the size oa an unfolded dinner napkin as thich as 6 business cards: 6 layers.
</p><p>
Vison: V1 at bottom, then V2, V4, and IT. All with two way connections.
</p><p>
Eyes: new saccades three times per second: take in a face by fixating on the eyes, nose
mouth.
<br/>
Each cell in V1 is seeing edges at one specific location and orientation (a la Hubel and
Wiesel); IT are invariant representation.
<br/>
Hubel and Wiesel is the feedforward part (up from V1 to V2 etc); equally esential is the
FEEDBACK at each step with expectations
<br/>
at each step.
<br/>
For many years scientists ignored the feedback connections. But, when you realize that
the cortex's core function is to make prediction, then you
<br/>
have to put feedback into the model; the brain must send info flowing back to the regions
that first received the inputs.
<br/>
Prediction requires a comparison btwn what is happening and what you expect to happen.
<br/>
WHAT IS ACTUALLY HAPPENING IS FLOWING UP< WHAT YOU EXPECT TO HAPPEN IS FLOWING DOWN!!!!
<br/>
(the same think happens at all levels and in all the senses).
</p><p>
114: Fig 3: Bottom level:(touch, audition, vision: spatially specific, fast changing,
features/details
<br/>
Top level spatially invariant, slow changing "objects"
</p><p>
Auditiory cortex might have a group of cells that fire thruout the utterance "good morning
or thank you ...
<br/>
doesn't matter who says it; what voice, what pitch etc. (even though A1 - the primary
auditory cortex - is receiving totally different input)
<br/>
Three Blind Mice : same cell pattern fires whether played on piano or clarinet or sung by
a child.
<br/>
Touch: its the same pen clip that I touch whether with my fingers or lips.
</p><p>
MUST HAVE FLOW OF PATTERNS OVER TIME (can't do it with momentary input ... you need the
saccades)
<br/>
Researchers have mistakenly thought we could do recognition with brief fixations ... but a
flow of input is critical.
<br/>
NORMAL VISION, like reading this book, REQUIRES CONSTANT EYE MVMT.
</p><p>
117 INTEGRATING THE SENSES
<br/>
Association areas: I hear Keo our family cat with his bell jingling ... as he walks I
integrate all the sense input into a model of Keo.
<br/>
I'd be surprised if something other than Keo walked in. Same with touching the pen clip.
Or riding my bike (I ride my bike to work several days a week).
<br/>
My brain predicts what bike feels like, motion of pedals, chain rattle, wheel spinning.
</p><p>
119 Once you understand how interconnected the senses are, you must conclude that the
entire neocortex, all the sensory and assoc. areas, acts as one.
<br/>
Everything flow up and down a single multibranched hierarchy.
</p><p>
ALL PREDICTIONS ARE LEARNED BY EXPERIENCE. Even the motor cortex works by predictions ...
it looks just like sensory cortex.
</p><p>
122-125 Invariant representations are formed at every level of the hierarchy, not just at
the top (at IT).
</p><p>
A MODEL OF THE WORLD 125
<br/>
Why is neocortex bulit as a hierarchy ... cause the world itself is hierarchical.
<br/>
The real world's nested structure is mirrored by the nested structure of your cortex.
<br/>
Songs are compsoed by phrases, composed of intervals, composed of notes;
<br/>
letters combine to form syllables to form words to form clauses etc.
<br/>
Houses have rooms which have walls and floors etc.
<br/>
All objects are composed of subobjects.
</p><p>
Real worlk objects can be concrete, like a lizard, a face or a door, or abstract like a
word or theory.
<br/>
They are treated the same way.
<br/>
They are both just sequences of patterns that occur together over time in a predictable
fashion.
<br/>
Predictability is the very defintion of reality. A predicatable sequence of patterns must
be part of an object that really exists.
</p><p>
129: Patterns that you continually see together: he refers to as having a common NAME.
The name of the sequence of sequences is passed up higher in the hierarchy.
<br/>
Each region of cortex has a repertoire of sequences that it knows, analogous to a
repertoire of songs. ... the sound of surf crashing, mother's face, the path to
<br/>
the corner store, how to shuffle cards, etc.
<br/>
The NAME is the group of cells whose collective firing represents the set of objects in
the sequence. The cells are active as long as the sequence is palyed;
<br/>
this is the "name that is passed up higher.
</p><p>
130 By collapsing predictable sequences into "named objects" we achieve more stability
the higher we go.
</p><p>
The opposite happends when you move down the hierarchy. Recite the Gettysburg Address.
First its unfolded into a sequence of phrases, then each
<br/>
phrase is unfolded into words, then to the auditory cortex into phonemes and motor
commands.
<br/>
Invariance works to our advantage at it flows down the hierarchy ... say you want to type
it instead of saying it.
<br/>
When the words hit the motor cortex they can be re-routed into typing commands instead of
speaking commands.
<br/>
Four score and seven years ago our fathers brought forth. You don't have to learn it
twice.
<br/>
EFFICIENCY ... use one representation ar the top for both speaking and typing; also use
one rep. at the top for Martin Luther Kings talk and Lincolns.
</p><p>
132: Hierarchy of motor commands is like the general who tells his underlings "move the
troops to Florida for the winter" which percolates down the chain of command
<br/>
to more and more specific steps.
</p><p>
134: What is a SEQUENCE? Sorting colored bits of paper: first sort each paper into
buckets of red, green, yellow, blue;
<br/>
then make sequences of them rrgpog. Note that you can USE the SEQUENCE expectation to
help classify each object ...
<br/>
if that last swatch green or yellow ... well it should be green cuz I just saw rrgpo ....
...
<br/>
Whe you hear an ambiguous word in a sentence you fill in based on the rest of the setnence.
<br/>
You hear what you expect to hear and see what you expect to see.
<br/>
Notice the MEMORY OF SEQUENCES ALLOWS YOU TO RESOLVE AMBIGUITY AND ALSO TO PREDICT WHAT
WILL HAPPEN NEXT.
</p><p>
Note: unlike AI and machine vision ... there is no template that is matched ... there is
no picture of a cup stored in your brain.
<br/>
Unlike a camera's memory, your brain remembers the world as it is, not as it appears.
</p><p>
137 WHAT A REGION OF CORTEX LOOKS LIKE
<br/>
The largest primary sensory cortex is V1 the primary visual cortex ... the size of a
passport.
</p><p>
There are SIX LAYERS : like 6 business cards. see Fig 6.
<br/>
Layer 1 is at the TOP - primarily mats of axons running parallel to the surface.
<br/>
Layer 2 and 3 are v. similar with tightly packed pyramidal cells.
<br/>
Layer 5 has regular pyramidal cells and also extra big pyramid shaped cells.
<br/>
Layer 6 at bottom: pyramid cells and also several unique cells types.
</p><p>
Cells are loosely organized into columns (they activate together). Also embryologically a
single precursor cell forms a column.
<br/>
Imagine each column as follows: glue a thousand hairs together and cut them to the size
of an i. Thats layers 2 thru 6.
<br/>
Then lay long extra-thin hairs on top running horizontally ... that's layer 1.
<br/>
BUT 90% of the synapses onto the column come from elsewhere ... soem from neighboring
columns, some from halfway across the brain.
</p><p>
Why do you need this wiring? To predict the next note of a song, you need to know the
song's name, where you are in the son, how much time passed
<br/>
since the last note, and what it was.
</p><p>
141 Layer 4 is the main input layer.
<br/>
LAYER 1 is the essential layer that spreads out information and spreads the invariant
representation to other columns so that
<br/>
it can generate specific predictions.143 See Fig 8: Layer 6 at bottom is an output layer
which then connects to layer 1 of other (lower) columns
<br/>
to make specific predictions. Layer 1 provides the means to aloow the flow of patterns to
branch at each level.
<br/>
So, first input to Layer 1 is from Layer 6 axons from columns that are higher in the
hierarchy.
</p><p>
</p>