Inside the model
We set this feature's activity to zero for this prediction. The input and model weights stay the same.
Explore the dataset ↗Research Lab
What happens when we turn off one feature?
CONTEXT LAB
What happens when the input changes?
Change six authored factors on one 64-state grid, or browse recorded input transformations.
CONTROLLED INPUTS
Selected configuration00 / 63
Feature activation
0 / 64 measurements32 paired differences for the selected factor
Mean on-minus-off difference for each factor
Experiment details
INPUT TRANSFORMATIONS
These are proposed transformations. A wrapper change does not establish that meaning stayed the same.
Recorded redistribution
Proposed transformationSOURCE CATEGORY
TRANSFORMATION
Measured summaries
Experiment details
COMPARE MODELS
Where do the recorded sites differ?
Explore the tradeoffs between active features, reconstruction, and downstream measurements. Compare matched protocols before interpreting differences. There is no single ranking.
Quality tradeoffs
No measurementsA point appears only when both axes have real measurements for the same site and selected protocol. Different units and metrics are not collapsed into one score.
RESEARCH
Do feature effects add up?
Construct a measured coalition. Compare the sum of individual effects with the actual combined response on the same downstream quantity.
MEASURED COALITION GROUP
SELECT FEATURES TO INCLUDE
Coalition∅
Predicted vs measured
Recorded responsesAdditive prediction—
Actual measurement—
Interaction residual—
Subset measurements