Hi , i try to implement HTM system.I have my first classification results for rec-center data set. pred_actual_values_HTM|666x500
These results came , with system parameters :
CellNumber : 8 per column
Columns : 512
Y axis represents value for energy consumption , an X axis timestep.
Is there a similar test in nupic git , or a way i can check if i step in thge right direction?
I would appreciate any ideas.
This is a prediction problem, not a classification problem. You might want to update the title of this thread to reflect that.
Did you start off with any example code? How did you get your model parameters? Where did you come up with 8 cells per column and 512 mini-columns? Usually we use 32 cells per column and 2048 mini-columns.
I have started with reference to BAMI’s algortihm, and nupic’s implementation(python).
The use of 512 mini-column and 8 cells per column is because of my system’s memory size.The other parameters are according to nupic’s suggestions.Could this difference in parameter size critically affect the system’s response?
@scott Take a look at:
Do you think decreasing the size of the mini-columns and number of mini-columns would cause predictions like this? Seems odd that they are always consistently higher (and lower in the troughs) than actual values.
My gut feeling is that something else is wrong.
@watchdog Can you tell us what version of nupic you are running?
I have built my implementation in MATLAB. I tried to stay really close to nupic in algorithm details.
I could send you my code to review, in case you have the time.
Ah, ok. I moved this from #nupic:developers into #htm-hackers. I hope you understand that we can’t help debug community HTM implementations. I’m not sure what is wrong, but there are so many moving parts that it could be any number of things. You’ll have to dig into it to figure it out.
(@scott never mind about my reference above)
That’s great that you wrote your own version!
You can lower the number of cells per column down to 8 and everything should still work fine. But 32 minicolumns is far too few to get the SDR properties. If you really need to reduce the memory usage I’d recommend lowering cells per column and segments per cell but keeping the number of minicolumns high (we default to 2048 but 1000 will probably work fine, lower than that risks issues).
Excuse me for the typo, i use 512 mini-columns.
Has anyone faced the same issue? I use SDR classifier ,with softmax function.In cell level i get accurate predictions,with some false positives but with really low false negatives.