Marcus Lewis comments

Results 19 comments of


                                            Marcus Lewis

C++ implementations of TM result in different NAB scores

`tm_py` results: ``` Final score for 'numenta' detector on 'reward_low_FP_rate' profile = 39.90 Final score for 'numenta' detector on 'reward_low_FN_rate' profile = 53.09 Final score for 'numenta' detector on 'standard'...

Connections: createSynapse() has non-optimal / unneeded if test

At that point in code we haven't instantiated the `Synapse` yet. We could create a `Synapse(-1, segment, None)` and then populate its `synapse.idx` and `synapse.data` in the if/else. Then we...

Fix encoder implementation to standards

Questions: - It sounds like you have a strong position against putting encoder Sensors next to their corresponding Encoders. Tentatively, I support the current approach. Otherwise we're polluting the codebase....

Use optimized lingebra math libraries

I commented on https://github.com/numenta/nupic/pull/3131#issuecomment-221298390.

Document better the new Extended TM parameters apical/basal input dim

Yeah I agree that the research code could be better separated from the public API. (FYI there is already a `experimental/README.md`, but maybe it should be longer.)

lazy-growing synapse graph

Thinking of a tall hierarchy, it's interesting to think about how this would change things. Starting with an untrained model, the first region would start activating. Then the second. Then...

Suggestion to try "substitution" or "context" pooling

I'm still catching up, at my own pace and in my own style (currently rereading On Intelligence). Your "replaying the original sequence" brought this passage to mind: > The next...

[Bug] `botorch.optim.utils.sample_all_priors` fails to sample priors that GPyTorch can sample

I'm running into this same issue. This issue occurs if you use `fit_gpytorch_mll` on a model that uses different priors parameters for different ARD dimensions (e.g. if you set lengthscale...

[Bug] load_state_dict doesn't invalidate cached transformed inputs

Cool, I'm looking forward to watching that PR. Maybe this is obvious at this point, but a quick fix to the current codebase would be to add to `botorch.models.Model`: ```python...