search

Caleydo / confusionflow-ui

Visualize learning dynamics in multi-class image classifiers (e.g., convolutional neural networks) 💬 @gfrogat @thinkh
https://confusionflow.org
BSD 3-Clause "New" or "Revised" License
1 stars 0 forks source link

retrieval of class combination for nclasses >> 10 #198

Closed gfrogat closed 6 years ago

gfrogat commented 6 years ago

If we have more than 10 classes we require a way to select a "interesting" subsets of classes, where "interesting" could be for example abnormal error curves.

Error Curves

For the normal error curves this could be achieved by simple histograms (see #86, #149), which would return a set of classes that have curves that show this behaviour.

While "stacked histograms" (see #86) would work I think it would be even more interesting to create a horizon chart (rotated by 90 degrees) which provides a space filling layout of the histograms.

image

By brushing the individual horizon histograms for each epoch we could easily retrieve interesting classes.

Trajectory Plot

For the planned trajectory plots we could also utilize a histogram filtering approach. By projecting each "epoch error pair" onto the diagonal we would also end up with a histogram.

image

gfrogat commented 6 years ago

related #111

thinkh commented 6 years ago

Out of scope for paper.