Dark Experience for General Continual Learning: a Strong, Simple Baseline

Authors: Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, SIMONE CALDERARA

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental By conducting an extensive analysis on both standard benchmarks and a novel GCL evaluation setting (MNIST-360), we show that such a seemingly simple baseline outperforms consolidated approaches and leverages limited resources.
Researcher Affiliation Academia AImage Lab University of Modena and Reggio Emilia, Modena, Italy name.surname@unimore.it
Pseudocode Yes Algorithm 1 Dark Experience Replay Input: dataset D, parameters θ, scalar α, learning rate λ M {} for (x, y) in D do (x , z , y ) sample(M) xt augment(x) x t augment(x ) z hθ(xt) reg α z hθ(x t) 2 2 θ θ + λ θ[ℓ(y, fθ(xt)) + reg] M reservoir(M, (x, z)) end for
Open Source Code Yes Code is available at https://github.com/aimagelab/mammoth.
Open Datasets Yes In practice, we follow [10, 42] by splitting CIFAR-10 [21] and Tiny Image Net [38] in 5 and 10 tasks, each of which introduces 2 and 20 classes respectively. We show all the classes in the same fixed order across different runs. For this setting, we leverage two common protocols built upon the MNIST dataset [23], namely Permuted MNIST [20] and Rotated MNIST [27].
Dataset Splits Yes We select hyperparameters by performing a grid-search on a validation set, the latter obtained by sampling 10% of the training set.
Hardware Specification Yes We conduct all tests under the same conditions, running each benchmark on a Desktop Computer equipped with an NVIDIA Titan X GPU and an Intel i7-6850K CPU.
Software Dependencies No The paper mentions using a 'Stochastic Gradient Descent (SGD) optimizer' and 'ResNet18' but does not specify programming language versions or library versions with numbers, such as Python 3.x or PyTorch 1.x.
Experiment Setup Yes For MNIST-based settings, one epoch per task is sufficient. Conversely, we increase the number of epochs to 50 for Sequential CIFAR-10 and 100 for Sequential Tiny Image Net respectively... We select hyperparameters by performing a grid-search on a validation set... setting them both to 0.5 yields stable performance.