SP fix global inhibition returns less than numDesired columns#542
SP fix global inhibition returns less than numDesired columns#542
Conversation
the global inhibition should return numDesired active cols, but it returns much less once sub threshold cols are removed
globa inh does not return nDesired columns how it should, even if it could have.
src/htm/algorithms/SpatialPooler.cpp
Outdated
| { | ||
| if (overlaps_[a] == overlaps_[b]) { | ||
| same_overlap++; | ||
| return a > b; //but we also need this for deterministic results |
There was a problem hiding this comment.
problem 1: we are removing columns with the same overlap from the top n selection.
src/htm/algorithms/SpatialPooler.cpp
Outdated
| activeColumns.end(), | ||
| compare); | ||
| // Remove the columns which lost the competition. | ||
| activeColumns.resize(numDesired); |
There was a problem hiding this comment.
Problem 2: here we crop to numDesired, but too early..
| // Remove sub-threshold winners | ||
| while( !activeColumns.empty() && | ||
| overlaps[activeColumns.back()] < stimulusThreshold_) | ||
| activeColumns.pop_back(); |
There was a problem hiding this comment.
..since this loop further discards weak columns.
|
I don't think that this is really a problem... Global Inhibition should only return fewer than num-desired active mini-columns if there are fewer than num-desired mini-columns which meet the activation threshold. |
well, then 2 is not a problem (just update the doc)
this still is. there can be "same (overlap)" columns > threshold, which we discard in the comparison fn. |
|
Actually disagree. Well, agree with the explanation, but that is a problem now:
|
2 participants
Inhibition reduces the number of active columns.
Global inhibition should return constant sparsity (translated to only numDesired columns survives the inhibition).
This is almost never the case, even if we could, g inh returns fewer columns.
This PR is just WIP now that debugs the problem, I'm looking for best way to fix it.
Cropping to nth_element (to numDesired) might be premature optimization, as after that we remove sub threshold, zero, etc columns.
Some items to look into in this PR:
computeActivityreturns only sparse array (removes zero overlaps) -- this might be nice optimization, as everywhere instead of iterating through 0..numColumns, we'd go << SDR.sizeif overlaps[a]==overlaps[b]: return a< bremoves duplicit entries. That is wrong, as 2 (or more) same overlaps are valid, and might be in the top n columns to keep.