Look-ahead Meta Learning for Continual Learning

Authors: Gunshi Gupta, Karmesh Yadav, Liam Paull

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

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we evaluate La-MAML in settings where the model has to learn a set of sequentially streaming classification tasks.
Researcher Affiliation Academia Gunshi Gupta Mila, Ude M guptagun@mila.quebec Karmesh Yadav * Carnegie Mellon University karmeshy@andrew.cmu.edu Liam Paull Mila, Ude M paulll@iro.umontreal.ca
Pseudocode Yes Algorithm 1 La-MAML : Look-ahead MAML
Open Source Code Yes The code for our algorithm can be found at: https://github.com/montrealrobotics/La-MAML
Open Datasets Yes MNIST Rotations, introduced in [17], comprises tasks to classify MNIST digits rotated by a different common angle in [0, 180] degrees in each task. In MNIST Permutations, tasks are generated by shuffling the image pixels by a fixed random permutation... We conduct experiments on the CIFAR-100 dataset... We also evaluate on the Tiny Imagenet-200 dataset...
Dataset Splits Yes Each task consists of 1000 images for MNIST Permutations and Rotations and 200 images for Many Permutations for training and validation and 2000 images for testing.
Hardware Specification No No specific hardware details (like GPU or CPU models, or cloud instance types) used for experiments were mentioned in the paper.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., Python 3.x, PyTorch 1.x) needed for replication.
Experiment Setup Yes We use a fully connected network with three hidden layers (256 units per layer) and ReLU non-linearities for the MNIST experiments. For Tiny Imagenet and CIFAR-100, we use a ResNet-18... For the CIFAR-100 and Tiny Imagenet experiments, we use a batch size of 64 and train for 50 epochs. Each method is allowed a replay-buffer, containing upto 200 and 400 samples for CIFAR-100 and Tiny Imagenet respectively.