0
Jeff Hawkin's Model of the Cortex
Posted by
thughes
,
29 April 2008
·
907 views
Life Extension and Transhumanism
I've been reading this:
http://www.amazon.co...a...7890&sr=8-1
He presents an interesting hypothesis based on the cortex running a uniform pattern matching and prediction algorithm that can be adapted to many purposes.
It works something like this:
This is a vastly simplified cortical column. There are of course many horizontal attachments, and several layers in the column have various types of neurons. The cortex is a hierarchy of regions of these columns. Note that we know such columns and regions exist from experiment and observation, so here at least the model matches reality.
Information can travel both up and down this hierarchy. When information travels up the hierarchy, the cortex is pattern matching on input from the outside world. When information travels down the hierarchy, the cortex is predicting what pattern it should match next.
A Cortical Column
Information from the next lower cortical region enters a column via synapses in layer 4. Multiple columns from the lower region will synapse here. The column will only turn on for certain combinations of these inputs. Note how this neatly handles ambiguity: the match doesn't have to be exact, it just has to cross a certain threshold. Columns also inhibit neighboring columns, reinforcing this ambiguity resolution.
When layer 4 fires, it lights up the rest of the column, essentially signaling a certain input has been recognized. This match then propagates around the hierarchy in 3 main ways:
* Up the Hierarchy: layers 2 and 3 synapse with layer 4 in various columns in the next higher cortical region, signaling the pattern match.
* Recursively: Layer 5 sends its information to the thalamus, which sends it back to layer 1 of the same cortical region. Since this is time delayed, it allows the region to know what the last input it matched was. This is important for temporal pattern matching.
* Down the Hierarchy: Layer 6 sends its information to layer 1 of the next lower cortical region.
Patterns
A pattern would be a certain set of columns firing together, which would be recognized by a column (or columns) in the next higher cortical region. A pattern can be temporal, since we have incoming information from the thalamus about the last group of columns that fired in this region. Or in other words, a certain sequence of column firing in the current region could cause a column in the next higher region to continuously fire, recognizing this sequence.
Note that this column in the higher cortical region also sends input back to the cortical region, via layer 1. This is the pattern "name".
Prediction
Input from layer 6 of the above region represents the pattern "name". Input from the thalamus represents where we are in the pattern, temporally. Both of these go to layer 1 of the cortical region.
Layers 2, 3, and 5 of the column have synapses in layer 1. If the synapses in layer 1 are active when the column turns on from below (via input to layer 4), these connections grow stronger. Eventually, the column can fire without layer 4 if the correct input is in layer 1. This is a prediction.
How does a column recognize when a prediction is met or not? The proposed hypothesis is as follows:
Layer 2 of all the columns involved in a pattern will stay lit (via those synapses from layer 1) when the cortical region above believes its recognizing the pattern. There are 2 actual layers in layer 3 (there is apparently evidence for this at least): 3a and 3b. When a pattern is in layer 1, layer 3a of the columns in the pattern will inhibit layer 3b. Layer 3b of a column will only turn on if the column fires unexpectedly, via input from below.
Unexpected patterns will propagate up the hierarchy until some level can handle them. Those that propagate all the way up will enter our conscious awareness as an anomaly...
Invariant Patterns
This model helps explain how the brain is so flexible in its pattern matching. A pattern may have variant parts, its the sum of these parts that cause the cortical region to recognize the pattern.
The model also nicely explains what we know about brain plasticity (eg. co-opting of various brain regions for different functionality in injured brains), since it implies the cortex does not start out specialized, it learns to be specialized. Thus, cortical regions could be co-opted for different input if the original input is damaged (eg. loss of a sense, or loss of a cortical region)
For a much better overview of the entire model, see the original book.
Doesn't this just make you want to run out and model it?
http://www.amazon.co...a...7890&sr=8-1
He presents an interesting hypothesis based on the cortex running a uniform pattern matching and prediction algorithm that can be adapted to many purposes.
It works something like this:
This is a vastly simplified cortical column. There are of course many horizontal attachments, and several layers in the column have various types of neurons. The cortex is a hierarchy of regions of these columns. Note that we know such columns and regions exist from experiment and observation, so here at least the model matches reality.
Information can travel both up and down this hierarchy. When information travels up the hierarchy, the cortex is pattern matching on input from the outside world. When information travels down the hierarchy, the cortex is predicting what pattern it should match next.
A Cortical Column
Information from the next lower cortical region enters a column via synapses in layer 4. Multiple columns from the lower region will synapse here. The column will only turn on for certain combinations of these inputs. Note how this neatly handles ambiguity: the match doesn't have to be exact, it just has to cross a certain threshold. Columns also inhibit neighboring columns, reinforcing this ambiguity resolution.
When layer 4 fires, it lights up the rest of the column, essentially signaling a certain input has been recognized. This match then propagates around the hierarchy in 3 main ways:
* Up the Hierarchy: layers 2 and 3 synapse with layer 4 in various columns in the next higher cortical region, signaling the pattern match.
* Recursively: Layer 5 sends its information to the thalamus, which sends it back to layer 1 of the same cortical region. Since this is time delayed, it allows the region to know what the last input it matched was. This is important for temporal pattern matching.
* Down the Hierarchy: Layer 6 sends its information to layer 1 of the next lower cortical region.
Patterns
A pattern would be a certain set of columns firing together, which would be recognized by a column (or columns) in the next higher cortical region. A pattern can be temporal, since we have incoming information from the thalamus about the last group of columns that fired in this region. Or in other words, a certain sequence of column firing in the current region could cause a column in the next higher region to continuously fire, recognizing this sequence.
Note that this column in the higher cortical region also sends input back to the cortical region, via layer 1. This is the pattern "name".
Prediction
Input from layer 6 of the above region represents the pattern "name". Input from the thalamus represents where we are in the pattern, temporally. Both of these go to layer 1 of the cortical region.
Layers 2, 3, and 5 of the column have synapses in layer 1. If the synapses in layer 1 are active when the column turns on from below (via input to layer 4), these connections grow stronger. Eventually, the column can fire without layer 4 if the correct input is in layer 1. This is a prediction.
How does a column recognize when a prediction is met or not? The proposed hypothesis is as follows:
Layer 2 of all the columns involved in a pattern will stay lit (via those synapses from layer 1) when the cortical region above believes its recognizing the pattern. There are 2 actual layers in layer 3 (there is apparently evidence for this at least): 3a and 3b. When a pattern is in layer 1, layer 3a of the columns in the pattern will inhibit layer 3b. Layer 3b of a column will only turn on if the column fires unexpectedly, via input from below.
Unexpected patterns will propagate up the hierarchy until some level can handle them. Those that propagate all the way up will enter our conscious awareness as an anomaly...
Invariant Patterns
This model helps explain how the brain is so flexible in its pattern matching. A pattern may have variant parts, its the sum of these parts that cause the cortical region to recognize the pattern.
The model also nicely explains what we know about brain plasticity (eg. co-opting of various brain regions for different functionality in injured brains), since it implies the cortex does not start out specialized, it learns to be specialized. Thus, cortical regions could be co-opted for different input if the original input is damaged (eg. loss of a sense, or loss of a cortical region)
For a much better overview of the entire model, see the original book.
Doesn't this just make you want to run out and model it?
