grero/hmmsort.py

Various methods to facilitate spike sorting using an HMM algorithm

hmmsort.py's Issues

Use double estimation of standard deviation to establish noise threshold

Add constraint that templates need to have points with amplitudes that exceed the noise threshold

Check that the superposition waveform exceed the noise threshold

Recognise templates with a large positive component

Current, the template finding algorithm seems to not identify templates with large positive components. This is illustrated in the following figure

Here, there is no template to separate the positive going signals, and they get grouped with the template showing a more 'normal' negative going spike. For this particular data set, even the initial 8 templates, after training, do not contain a signal with the positive component.

One possibility is to change the template initialisation to include a few with positive going spikes as well.

Copy cinv to hmmsort

The cinv variable, which is an estimate of the inverse covariance matrix, should be copied it hmmsort.mat so that we don't have to rely on the spike_templates.hdf5 file.

Possible transpose error

Sometimes the spike forms are truncated, indicating a problem with transposing the waveform matrix.

Multiple merges

Currently, I believe we can only merge waveforms into a single unit. We should allow merging into multiple templates as well.

Create a simple test case

Right now there are no tests for the algorithm. We should have something that at the minimum makes sure that the code runs.

Cleaning of templates

It struck me as I was looking at some of the templates return by hmm that perhaps we should do an additional cleanup step of the templates. We could fit each template to a damped sinusoid and use that fit as the templates fed to hmm_decode.

Small templates

It seems like templates are selected based on the energy of a noise patch rather than the peak exceeding some multiple of the estimated standard deviation. We could change this, so that the selection is more in line with way we expect from a visual inspection of the templates.

Shifted templates

Some templates appear to have a very long pre-excitation phase, which causes the template to be clipped.