[info] Update icarl results.

Author: Arthur Douillard

README.md

@@ -11,11 +11,18 @@ Every model must inherit `inclearn.models.base.IncrementalLearner`.
 
 ## iCaRL
 
-I have not yet been able to reproduce the results of iCaRL as you can see on the
-following figure:
-
 ![icarl](figures/icarl.png)
 
 My experiments are in green, with their means & standard deviations plotted.
 They were runned 40 times, with seed going from 1 to 40, each producing a
 different classes ordering.
+
+The metric used is the `average incremental accuracy`:
+
+> The result of the evaluation are curves of the classification accuracies after
+> each batch of classes. If a single number is preferable, we report the average of
+> these accuracies, called average incremental accuracy.
+
+If I understood well, the accuracy at task i (computed on all seen tasks) is averaged
+with all previous accuracy. A bit weird, but doing so get me a curve very similar
+to what the papier displayed.