DLwbm123 / FRCL

Functional Regularisation for Continual Learning with Gaussian Processes

FRCL

Functional Regularisation for Continual Learning with Gaussian Processes

by Pavel Andreev, Peter Mokrov and Alexander Kagan

This is an unofficial PyTorch implementation of the paper https://arxiv.org/abs/1901.11356 . The main goal of this project is to provide an independent reproduction of the results presented in the paper.

Project Proposal: pdf

Experiments launching

To launch our experiments use results_script.py The example of script run below:

> python .\results_script.py --device 'your device' --task 'permuted_mnist' --method 'baseline' --n_inducing 2

Available options for --task argument are split_mnist, permuted_mnist and omniglot. Available options for --method argument are baseline, frcl_random and frcl_trace.

Results of our experiments are presented in '.\results'. Besides, one can find notebooks with minimal working examples in '.\notebooks'.

Results

The presentation with the project main results is available here.

We results are also summarized in the table below.

Datset	Method	N points	Criteria	Accuracy (ours)	Accuracy (paper)
Split-MNIST	baseline	2	-	0.981	-
Split-MNIST	baseline	40	-	0.985	0.958
Split-MNIST	FRCL	2	Random	0.827	0.598
Split-MNIST	FRCL	2	Trace	0.82	0.82
Split-MNIST	FRCL	40	Random	0.986	0.971
Split-MNIST	FRCL	40	Trace	0.979	0.978
Permuted-MNIST	baseline	10	-	0.695	0.486
Permuted-MNIST	baseline	80	-	0.865	-
Permuted-MNIST	baseline	200	-	0.908	0.823
Permuted-MNIST	FRCL	10	Random	0.628/0.527*	0.482
Permuted-MNIST	FRCL	80	Random	0.838	-
Permuted-MNIST	FRCL	200	Random	0.942	0.943
Omniglot-10	baseline	60	-	0.381	-
Omniglot-10	FRCL	60	Random	0.376	-

Results of our experiments are presented in .\results

* the results appeared to significantly depend on initialization of parameters