Getting anomaly likelihood

Engineering NuPIC
1

Modifying the hot gym anomaly code, is this the correct way to obtain anomaly likelihood?

result = model.run({
  "timestamp": timestamp,
  "kw_energy_consumption": consumption
})

likelihood = result.inferences['multiStepBucketLikelihoods'][1].values()[0]
2

Hi,

I think that Iā€™m struggling with the same issueā€¦ I also modified the hot gym code to try and get both an anomaly probability/score and a prediction. Did this approach proved to be working correctly for you?

3

No, the anomaly likelihoods are not part of the core HTM computation. It is a post-process that should be implemented outside the compute loop. There are instructions here.

4

Hi,

I seeā€¦ but in order to compute the anomaly likelihood the raw anomaly score is required, however I found out that when I set ā€œinferenceTypeā€ in my model params to ā€œTemporalMultiStepā€ I can only get predictions in the inferences object and the ā€œanomalyScoreā€ is None and when I set it to TemporalAnomaly the ā€œanomalyScoreā€ is available but the ā€œmultiStepBestPredictionsā€ is notā€¦ Is it possible to use one HTM model to get both? Do I need to run two models in parallel (one for predictions and one for anomaly detection) or is there a better way? Can you please share some example code for that?

1 Like
5

This was not true when I created this tutorial years ago. Iā€™m not sure what has changed. Maybe you can look over this and see what might be different in your model?

6

Hi, Thanks a lot for the quick reply. I saw in that tutorial that ā€œYou might be wondering how to convert your prediction model into an anomaly modelā€

So, a model can be used for prediction OR anomaly detection but not both?

7

I thought that the TemporalAnomaly model still returned predictions, but I may be wrong.

1 Like
8

In practice a ā€˜predictingā€™ model should also be just as capable at anomaly detection, tho not necessarily vice-versa. Both are totally based on the current TM state.

Since the anomaly score is just a proportion of bursted to activated columns its trivially calculated. The ā€˜predictionā€™ functionality however is non-trivial, since a given TM state (which can vary a LOT in sparsity) has to be mapped to predicted value(s) back in the raw data typeā€™s encoding space.

There have been different implementations of ā€˜Classifiersā€™, mapping schemes from TM state to raw data value. I think the latest is the SDR Classifier.