idea #1: classifier based on similarity across different algorithms

[yarikoptic]

a rough idea

goal: to provide improved unit detection by combining results from multiple spike sorting algorithms. It closely relates to the idea of a metric to provide upper bound of accuracy for any spike sorting across multiple algorithms: https://github.com/flatironinstitute/spikeforest/issues/78#issuecomment-549647808

high level algorithm: given results (units) from multiple algorithms,

• establish cliques of "closest" units across algorithms
• use matrix (upper triangualr) of similarities within clique across algorithms as input to a classifier
• target label (binary 0 or 1) for a classifier would be decided based on thresholded max accuracy in respect to GT unit across classifiers
• eventually additional features (e.g. indicator variables for different neuron types) would be added
• trained classifier would establish "weighting" for similarities across algorithms to make a decision either a given clique of units correspond to GT unit
• cross-validate across datasets
• TODO: figure out also decision making on how final output unit from the spike sorter would be assembled from a clique (could be the unit from an algorithm with maximal overall "weighting" from the classifier)
• results of such metasort would be subject to spikeforest comparisons on a held out (for cross-validation) dataset

[satra]

this almost seems like ensembles of classifiers. you could create a bayesian formulation for this.

but the key is going to be validation: what is your validation dataset?

[yarikoptic]

this almost seems like ensembles of classifiers.

yeap

but the key is going to be validation: what is your validation dataset?

real (extracellular + intracellular recordings) and simulated - the ones https://spikeforest.flatironinstitute.org/ already uses . But eventually -- we would also train on curated data to learn professional idiosyncracies ;)