breznak

TM.anomaly: fix score for likelihood anomaly modes

5

we used to return likelihood (of the score), but for anomaly we must return `1 - likelihood` Fixes #665 TODO: - [ ] optionally integrate sine wave tests for anomaly...

Connections: apply synapsesCompetition() in adaptSegment()

1

- [x] SP uses synapse pruning, same as TM has been doing - [ ] make pruning default to true in adaptSegment (unless it breaks too many tests)? - or...

Provide bindings for BaseEncoder for python encoders

16

So py-only encoders can derive from that class (same as c++). Needed for: #691 , #259 - [ ] GridCellEncoder - [ ] RetinaEncoder TBD #691

followup to #745 At least with random input, average time per iteration (ms/iter) slows down after many steps of running a TM. Some is expected due to book-keeping and cleaning,...

Add new NLP encoder Google's Universal Sentence Encoder

2

See https://towardsdatascience.com/use-cases-of-googles-universal-sentence-encoder-in-production-dd5aaab4fc15 https://www.learnopencv.com/universal-sentence-encoder/ For #259 , #634

SP return const random representation on empty input & SP anomaly

1

#### Problem: empty SP output on low overlap Currently, when the overlap with input (unexpected input, unlearned, ..?) is too low, SP returns 0 active columns. This causes problems in...

Encoding binary categories & empty input

4

In DateTime encoder #462 I noticed a field weekend is encoded as a Category (with scalar encoder). Weekend is a binary field : T/F. We use scalar encoder for categories,...

Support 32bit Linux builds

2

This feature was removed, but as we got rid of all ASM, it might be feasible to turn it on easily again. - [ ] see of 32bit linux can...

Improved Conda install, dependencies

4

@dkeeney I've recently been toying with our conda python install. Turns out I could be able to get most of the deps through conda (eigen, cereal, pybind11,..) But our CMake...

Improve speed of SP local inhibition, 70x slower than global

1

CC CUDA SP #50, SP faster LocalInho #86 , #92 From https://github.com/htm-community/nupic.cpp/pull/119#pullrequestreview-178811874 : I'm evaluation the performance impact of ND vs 1D, where algorithmically the 1D is really simpler. -...