Implement Column Pooler from htmresearch

[breznak]

Column Pooler, CP, is implemented in python in HTM research. Used for sensory-motor experiments, it extends functionality of TM.

[ctrl-z-9000-times]

Ive never heard of a column pooler, so what does it do? What features does the column pooler have which the SP / TM don't have?

[breznak]

there are papers on Column pooler and sensory-motor learning, "touch" experiments, invariant object representation. See eg https://numenta.com/assets/pdf/posters/CosynePoster_2017.pdf

[ctrl-z-9000-times]

We should incorporate the paper "Kropff & Treves 2008" into the column pooler. This paper describes simple modifications to the SP which make viewpoint invariance form naturally.

[Thanh-Binh]

@ctr-z-9000-times do you mean porting ColumnPooler from Python into c++? Or complete new ones?

[ctrl-z-9000-times]

I don't know, why do you ask? Im hoping to make a plan for this, but i dont have a plan yet.

[ctrl-z-9000-times]

@Thanh-Binh ,

do you mean porting ColumnPooler from Python into c++? Or complete new ones?

I'd like to write the column pooler in C++, because C++ is fast and its the common language that all other languages have bindings for. After done, it should be very similar to the current python version.

Conceptually, the column pooler is similar to the (SP + TM) system; it does many of the same things. I hope we can copy-paste a lot of code from SP&TM to Column Pooler.

Python Column Pooler: https://github.com/numenta/htmresearch/blob/master/htmresearch/algorithms/column_pooler.py

[Thanh-Binh]

@ctr-z-9000-times: understood. The c++ porting is straight forwards

[ctrl-z-9000-times]

FYI: The column pooler should also implement multiple proximal segments. Each cell's overlap is the maximum of its segment's overlaps.

[ctrl-z-9000-times]

I've started work on branch columnPooler2 https://github.com/htm-community/nupic.cpp/tree/columnPooler2

See file: src/nupic/algorithms/ColumnPooler.cpp So far it is an incomplete list of todo notes.

[ctrl-z-9000-times]

FYI everyone, the ColumnPooler should contain a complete and working SpatialPooler. To get just the SP functionality without any of the new stuff you'd use:

• No distal segments, effectively disabling them
• 1 proximal segment per neuron, each neuron is a mini-column in this context
• Call cp.reset() before every use, which effectively disables the temporal aspects of the CP.

@Thanh-Binh, IIRC the Temporal Pooler was an early incarnation of the Column Pooler.

[breznak]

the ColumnPooler should contain a complete and working SpatialPooler.

does it have to, do we want to mix the functionality? Mostly if stuff like inhibition, boosting etc need to be implemented in CP.

OT, but if CP does need to implement the above for its own functionality, should we scratch the SP implementation and just make the class create instance of CP that acts as a SP as you described .. ?

o get just the SP functionality without any of the new stuff you'd use:

[ctrl-z-9000-times]

does it have to [...] ?

Yes. The column pooler needs everything in the SP and a more. In branch columnPooler2 where I'm keeping the work, I have copied the MNIST example and modified it to use a CP instead of an SP, and it currently scores 93% using all of the same parameters as the SP which scored 95%.

should we scratch the SP implementation?

No, at least not until the CP is far more complete, and we prove that it works as I'm advertising. Even if we could replace everything with it, it might be nice to keep the old stuff around for all of the same reasons why we will probably keep the python implementations of the algorithms.

[ctrl-z-9000-times]

More info about Kropff & Treves 2008:

• Original Paper: https://onlinelibrary.wiley.com/doi/epdf/10.1002/hipo.20520
• I reproduced their grid cell model using NuPIC SP: https://github.com/ctrl-z-9000-times/KropffTreves2008_reproduction/blob/master/README.md
• I re-purposed it to do sensory-motor inference / view-point invariance: https://github.com/ctrl-z-9000-times/sdr_algorithms/blob/master/ascii_stability.pdf

[ctrl-z-9000-times]

I have a hypothesis about how the column pooler could function as a temporal memory. Im discussing it at https://discourse.numenta.org/t/column-pooler-as-a-temporal-memory/5384

[ctrl-z-9000-times]

Update: I have a prototype of the column pooler with the experimental changes described in the previous post. I'm testing it on sequence recognition where it currently scores 75% accuracy on a 50 category recognition task. I'm going to continue experimenting with the goal of making it work well enough to use in place of a temporal memory. Then I'm going to write a report about it.

[ctrl-z-9000-times]

I'd like to reproduce Numenta's Columns paper. They generate several artifical datasets to test their creation, which we should emulate. It does not need to be an exact reproduction, just close enough to verify the critical properties of the ColumnPooler. These properties include:

• Network converges on a single answer, more columns helps converge faster.
• Network recognizes objects, measure capacity.
• Measure noise robustness.

https://www.frontiersin.org/articles/10.3389/fncir.2017.00081/full

[dkeeney]

Sounds good to me. When you get the algorithm operational I will build a NetworkAPI plugin for it.