Marcus Lewis

Encoders' output is incompatible with the C++ SpatialPooler

1

Using the C++ SpatialPooler, calling `SpatialPooler.compute(encoder.encode(42))` will quietly do the very wrong thing because it's given a buffer of 8-bit integers and it thinks it's getting a buffer of 32-bit...

In this notebook, we should replace occurences of `dtype="int"` with `dtype="uint32`". https://github.com/numenta/nupic/blob/6924a904644bbb8cdba6ca7a62050cf8456614f8/examples/NuPIC%20Walkthrough.ipynb You can see we had some integer overflow issues in https://discourse.numenta.org/t/inconsistency-between-spatialpooler-in-nupic-and-nupic-bindings/1549/10?u=mrcslws Sometimes "int" is mapped to 32-bit, other...

Add a "checkInputs" parameter to TemporalMemory

1

Here's the current state of things: - Encoders output **a binary array**. - The SpatialPooler outputs **a binary array**. - The TemporalMemory receives **a list of active column indices**. -...

See discussion: https://github.com/numenta/nupic.core/pull/1164#discussion_r90347310 With local inhibition, the `inhibitionRadius` varies with the statistics of the input, so if you're using the `numActiveColumnsPerInhArea` parameter, for certain inputs you might end up with...

Proposal: SP potentialRadius should default to -1

1

Proposal: - In the SpatialPooler, support `potentialRadius=-1`. This value would disable any topology relationship between columns and inputs. This would save everyone from having to figure out their input width...

First discovered here https://github.com/numenta/nupic/pull/3362#issuecomment-249324499 If you change the SP_PARAMS seed in https://github.com/mrcslws/nupic/blob/6ad9c24c84d6d9a05606c8d63f63feb34995aa91/tests/integration/nupic/engine/network_creation_common.py to 1956, the `temporal_memory_compatibility_test.py` will fail. So there's a subtle difference between the Python and C++ TMs. It...

ETM suggestion: Use apical segments when determining the learning cell

1

Relevant code: https://github.com/numenta/nupic.core/blob/8a846d1d897440bc6ca1e46120520da166bc551a/src/nupic/experimental/ExtendedTemporalMemory.cpp#L616 I didn't change the algorithm at all, I just put all the logic for "learning cell" selection in one place. When a column is bursting, we choose...

The random number generator is slightly biased toward returning 0

4

This code looks wrong. https://github.com/numenta/nupic.core/blob/69b9c79cf34a9c7ab90c8c0f4616504da40edee7/src/nupic/utils/Random.cpp#L192 I think this line should be `while (sample >= smax);`. Here's a quick explanation of this code. It generates a random number between 0 and...

Activations in this model are ternary {-1,0,1}, not binary {-1,1}

3

This code uses `tensor.sign()` to binarize the activations and weights. https://github.com/itayhubara/BinaryNet.pytorch/blob/f5c3672dede608f568e073a583cadd7a8a88fa9d/models/binarized_modules.py#L13 The desired behavior is to always return -1 or 1, but `sign()` returns 0 for values that are 0....

[Bug] load_state_dict doesn't invalidate cached transformed inputs

4

# 🐛 Bug botorch models do not properly support `model.load_state_dict` when using `input_transform`s with trained parameters. After performing `model.load_state_dict`, the model continues using cached transformed inputs that were computed with...