Yes, that's right.

Yuwei

On Mon, Apr 4, 2016 at 1:33 PM, Sebastián Narváez <[email protected]>
wrote:

> Hi Yuwei, Thanks for your response, things are much clearer now. I was
> refering to this "if":
>
> if permanence > 0 and self.predictedSegmentDecrement > 0:
>
> Now, if I understand what you said, the connected synapses must be taken
> into account as much as the non connected, but active ones, for the
> formation of the matchingSegments and matchingCells variables. Their
> decrement will only be made when the next element of the sequence arrives
> and the matchingCells~Segments do not match with the current active cells.
> Is that right?
>
> On Mon, Apr 4, 2016 at 12:50 PM, Yuwei Cui <[email protected]> wrote:
>
>> Hello Sebastián,
>>
>> Please see my answers below:
>>
>>>
>>> 1) What do the matchingSegments and matchingCells represent?
>>>
>>
>> I think we recently include the logic here to model "long-term
>> depression". That is if the segment has sufficient activity at time t, but
>> does not become active at time t+1, it represents a potential false
>> prediction and should be punished. "sufficient activity" here means number
>> of active inputs is above minThreshold. matchingSegments and matchingCells
>> are used to determine predicted but inactive cells at the next time step
>> (see line 376 of learnOnSegments).
>>
>> This logic speeds up the forgetting of false predictions, but it should
>> be used with caution. If your problems has multiple correct predictions,
>> then it is OK to have some false predictions. Generally speaking, the ratio
>> between permanenceIncrement and predictedSegmentDecrement determines how
>> many multiple predictions can the model make at any time.
>>
>>
>>> 2) minThreshold is supposed to be the minimum number of synapses a
>>> segment must have in order to be considered for bursting, what does it do
>>> here?
>>>
>>
>> activationThreshold is the threshold for activation of a segment: if a
>> segment has more than activationThreshold number of active synapses, it
>> will fire a dendritic spike and depolarize the cell body.
>>
>> minThreshold is typically lower than activationThreshold and is only used
>> in learning phase (not in inference phase). It are used in two places as
>> far as I know.
>>
>> 1. If the number of synapses active on a segment is at least this
>> threshold, it is selected as the best matching cell in a bursting column.
>> (see function bestMatchingSegment)
>>
>> 2. It is used to determine predicted but inactive cells and segments as
>> described above.
>>
>>
>>> 3) As I see it, the if also grabs the permanneces above the connected
>>> threshold, why is that?
>>>
>>
>> I am not sure which "if" you are referring to here. Could you clarify
>> your question?
>>
>> Yuwei
>>
>>
>

Reply via email to