PCA

PCA = principal components analysis which is just computing the eigenvectors of the covariance or correlation matrix of values over time -- i.e., how different state values (e.g., neural network activations) change in relation to all the other state variables over time. The first principal component captures the greatest amount of variance in these changes. Hebbian learning in a neural network does much the same thing, actually.

• Docs: pca

Example code

	pc := &pca.PCA{}
	pc.TableCol(rels, "Hidden", metric.Covariance64) // does the PCA, stores results
	prj := &etable.Table{}
	// next project data onto first two eigenvalues (0,1):
	pc.ProjectColToTable(prj, rels, "Hidden", "TrialName", []int{0, 1})
	ss.ConfigPCAPlot(&PCAlot, prj)

and here's how you can configure a plot:

func (ss *Sim) ConfigPCAPlot(plt *eplot.Plot2D, dt *etable.Table) {
	nm, _ := dt.MetaData["name"]
	plt.Params.Title = "Family Trees PCA Plot: " + nm
	plt.Params.XAxisCol = "Prjn0"
	plt.SetTable(dt)
	plt.Params.Lines = false
	plt.Params.Points = true
	// order of params: on, fixMin, min, fixMax, max
	plt.SetColParams("TrialName", true, true, 0, false, 0)
	plt.SetColParams("Prjn0", false, true, 0, false, 0)
	plt.SetColParams("Prjn1", true, true, 0, false, 0)
}