Skip to content

Cross-validation

The CV Matrix tab shows how well the model tells its classes apart. In cross-validation the training colonies are split into parts; each part in turn is classified by an SVM trained on all the others. So every colony is classified by a model that has not seen it.

The CV Matrix tab

Baclan runs this at the end of every training. To run it again with other settings, change them and press Recalc.

Control Does
% Ticked, each row is in per cent of that row's colonies. Unticked, it shows numbers of colonies
Export Opens a menu: Features writes the features of all the training colonies to a text file, Crossvalidation Table the matrix as shown
n-Fold Into how many parts the colonies are split
Reps How many times the whole thing is repeated, with a new split each time. The numbers are the averages
Recalc Runs it again, with a progress window. Abort there leaves the old matrix

Reading the matrix

Each row is a class as you defined it. Each column is what the cross-validation called those colonies. So the diagonal is what it got right, and in the row of class A a large number under class B means that colonies of A are being taken for B.

What you see What to look at
A low diagonal for one class Its plates: mistaken colonies, artefacts left ticked, a different medium or incubation time
Two classes mixed with each other Whether the organisms can be told apart at all. Other features, or more plates
All close to 100 % with few colonies Whether colonies of one plate are too alike. Test it on plates it has not seen

Note

This measures how well the model separates the colonies it was trained on. Plates grown on another day are the harder test: put them in the Test Set.

The matrix and its n-Fold and Reps are saved with the model. Recalculating marks the model as changed.