Parameter optimization with htm.core

Engineering NuPIC Community Fork
1

I am new to HTM and I have a few newbie getting started questions. I am using python.

I am getting very different results with the hotgym example in python from htm.core and putting it in HTM Studio. HTM Studio detects a lot fewer anomolies and I am interested in how to optimize the parameters I am using for the encoder, SP and TM.

With htm.core, how should I do parameter optimization? Should I use swarming?

I am struggling with figuring out how to use Network API with htm.core? Specifically, I am not sure what regions are available or how to use them without reading the source. Is there documentation or an example for this?

I am having problems using the example hotgym example with other csv files using larger numbers. For example, I took the nyc_taxi.csv file from HTM Studio and it starts to consume all my memory when it gets to large numbers around 30,000 or so and I have to kill the process.

Does anyone have any tips for me?

Thanks,
Brian

1 Like
2

If I remember correctly, the algorithm behind HTM Studio’s anomaly detection is the BacktrackingTM, which is an older version of sequence memory that employs some non-biological methods. I do not think this logic was ported into htm.core (@dmac @David_Keeney correct me if I’m wrong). So I would not expect you would get the same results from these two systems.

2 Likes
3

BacktrackingTM was ported to htm.core but was later removed because it was not biological.

I am not sure which TM was used in the original hotgym example but the TM in htm.core my use parameters slightly differently so that could cause differences as well. @dmac or @breznak could give more insight on this.

1 Like