rgutzen / validationtools Goto Github PK

• Plot and compare the distributions of bin heights
• Quantify the difference in bin height distribution (compare to random activity?)
• Find significance in the distribution (via surrogates?)
• Also look at distribution of max bin heights and maybe the distribution of the corresponding time lags

• color image contur plot of histogram difference
• summation of one dimension (CC or lag) respectively -> marginals
• How to quantify multidimensional distribution difference? (t-test?)

The marginal of the collection of all CCHs where the time lag dimension is reduced can be represented as a generalized correlation matrix.
The matrix entries could be:

• The summed bin heights of the pairs CCH
• The maximum value of the pairs CCH
• The number or sum of the CCH bins larger than a certain threshold

• Review the test implementations and namespaces
• Wrap the tests in SciUnit classes
• Write documentation
• Link tests to HBP Validation Framework

Extending the view on correlation via the Pearson correlation coefficient (which only regards bin-wise synchrony) the correlation should be analyzable with both

• a fixed time lag
• by adding a time delay dimension

Should the function be collected separately or should the functionality be included in the non-time-resolved function via a parameter?

• Reduce the CCHs to the dimension of the first neuron of the pairs (CCH_i = Sum over j of CCH_i<j)
• Visualize the temporal correlation spread with X=Neuron i, Y=Time-Lag tau, Color=Correlation
• Quantify the temporal spread

For a network of N neurons, there are N*(N-1)/2 CCHs. The resulting data is 4 dimensional (Neuron#1, Neuron#2, Time-Lag, Correlation). Significant features in this data have to extracted, visualized and quantified.

• Plot 3D graph with X=Neuron-Axis-1, Y=Neuron-Axis-2, Z=Time-Lag-Axis, CCH bin-height in direction Y and color coded.
• Rescaling of CCH values to [0 .. 1] to apply visual aids (color, alpha)
• Correction for the firing rate of the neuron (cross_corr_coef)
• Thresholding of the CCH values to declutter the graph
• Quantify and interpret the dominant features

Consistent namespacing

Clustering of CC matrix

• Ordering by eigenvalue & -vectors
• Hierarchical clustering (scipy)

Distance Measure

• Either on CC matrix or on clustered matrix depending on whether neuron ids are consistent
• Angle between eigenvectors
• Spatial distance (scipy.spatial.distance.pdist)

• Should the single CCHs be corrected for firing rates, border effects?
• Significance testing for the bin entries (surrogates?)
• Include multiple testing corrections