Outliers and Anomalies
An outlier, in Tensormorph, is a value or region whose statistics don't match the distribution around it - a single scalar far outside a tensor's typical range, a tile whose absmax dwarfs its neighbors, a channel whose norm is an order of magnitude off from every other channel in the same layer. Unlike NaN/Inf detection, which flags an unambiguous invalid value, outlier detection is a statistical judgment - "unusual relative to what's normal here" - so it depends on comparing a candidate region's tensor statistics against a surrounding baseline rather than a fixed threshold.
Where anomalies get surfaced
The Problems panel is where flagged anomalies accumulate across a whole model - NaN/Inf occurrences, statistical outliers, and other checks - as a single navigable list rather than something you have to notice yourself while browsing. F8 steps to the next flagged anomaly from anywhere in the workbench, jumping the active view directly to it.
What counts as unusual
Because this is a judgment relative to a baseline, Tensormorph supports more than one comparison scope:
- Within-tensor - a value flagged relative to the rest of the same tensor's distribution (a classic statistical outlier: several standard deviations from the mean, or beyond a percentile threshold).
- Within-layer - a channel or tile flagged relative to its sibling channels/tiles in the same module, which is often more informative than a whole-tensor baseline for catching a single miscalibrated channel.
- Across revisions - a region flagged because it moved further than expected relative to a baseline checkpoint, which overlaps with weight diff territory but is framed here as "is this specific change unusual," not just "what changed."
Visualizing outliers
In the Matrix/Heatmap view, an outlier-highlight shading mode colors elements by how far they sit from the local baseline rather than by raw magnitude - useful specifically because a raw-magnitude heatmap can bury a genuinely anomalous small-magnitude value inside a sea of larger, but individually unremarkable, values.
Related resources
- Tensor Statistics - the statistics an outlier judgment is computed against.
- NaN and Inf Detection - the unambiguous special case of an invalid, rather than merely unusual, value.
- Expert Similarity - a related but distinct question: not "is this value unusual" but "are these two regions unusually similar or dissimilar."