Continual learning strategies(EWC, GEM) for rotated MNIST dataset
Ruinan Zhang rz1109@nyu.edu Manlan Li ml6589@nyu.edu
In this projct, our group exlpored the rotated MNIST dataset with two continual learning strategies:
-
(1) Elastic Weights Consolidation (EWC) Strategy (code can be found both in repo or here)
-
(2) Gradient Episodic Memory (GEM) (code can be found both in repo or here) !!! To run GEM please also use google drive to download zip dataset: https://drive.google.com/open?id=1QSi3ua8e41kUDu364s4e_hs5lBJUNN31
Training Dataset: 60000 * 28 * 28
Test Dataset: 10000 * 28 * 28
Rotation: Randomized rotation
MNIST VS Rotated MNIST
- (1) EWC run on total 10 tasks with final acc 85.78%:
- (2) GEM run on total 20 tasks with final acc 90.21%
Besides accuracy, there are two additional metrics used to assess the ability of the algorithm to transfer knowledge - Backward transfer (BWT) and Forward transfer (FWT)
BWT: the influence that learning a task t has on the performance on a previous task k ≺ t
FWT: the influence that learning a task t has on the performance on a future task k ≻ t
