Metrics

Metrics Evaluation

NeuralEmbedding computes quality metrics on the embedded trajectories stored in obj.E. Use computeMetrics to run a metric and store the result on the object.

Quick start

% After computing an embedding
NE.findEmbedding("PCA");

% Compute an embedding quality metric
flag = NE.computeMetrics("arclength");  % returns true on success

% Results live in NE.M (with metadata), plus NE.M_ internally
disp(NE.M);

computeMetrics(type) accepts several aliases (e.g., "arc" → arclength). Parameters are pulled from object properties when present (e.g., baseline_idx, signal_idx for DPrime).

Supported metrics (via computeMetrics)

Metric	Call	What it measures	Key parameters
Arclength	NE.computeMetrics("arc")	Median trajectory length across trials using linear/spline/pchip interpolation.	method (defaults to spline).
Alignment	NE.computeMetrics("alignment")	Dispersion of trajectories after time normalization.	type (default / equidist / index), nsec (segments).
DPrime	NE.computeMetrics("dprime")	Modulation index between baseline and signal windows per dimension.	baseline_idx, signal_idx (defaults inside metric).

All metrics expect a cell array of size nTrials × nAreas where each cell holds an nDims × T matrix (the format produced by findEmbedding).

Additional metric utilities

The +metrics package also includes standalone functions that operate on embedded trajectories: Deviation, Tangling, Frechet, and Poincare. They accept the same cell-array input ({nDims × T} per trial) and return scalars or per-dimension values, enabling custom analyses even when they are not wired into computeMetrics yet.

Best practices

• Run metrics after computing an embedding so obj.E is populated.
• Set NE.appendM = true; before calling computeMetrics if you want to store several metric results instead of replacing the most recent one.
• The metrics operate on smoothed, standardized data; adjust preprocessing parameters on the object before embedding if you need different smoothing or binning for the evaluation.