Online Continual Learning on Class Incremental Blurry Task Configuration with Anytime Inference

Authors: Hyunseo Koh, Dahyun Kim, Jung-Woo Ha, Jonghyun Choi

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our empirical validation demonstrates that the proposed method outperforms prior arts by large margins. Code and data splits are available at https://github.com/naver-ai/i-Blurry.
Researcher Affiliation Collaboration 1GIST, South Korea 2Upstage AI Research 3NAVER AI Lab. 4Yonsei University
Pseudocode Yes Algorithm 1 Update Sample-wise Importance; Algorithm 2 Sample-wise Importance Based Memory Update; Algorithm 3 Adaptive Learning Rate Scheduler
Open Source Code Yes Code and data splits are available at https://github.com/naver-ai/i-Blurry. All code and i-Blurry-N-M splits (see Supp.) is at https://github.com/naver-ai/i-Blurry.
Open Datasets Yes We use the CIFAR10, CIFAR100, Tiny Image Net, and Image Net datasets for empirical validations.
Dataset Splits No The paper specifies the datasets used (CIFAR10, CIFAR100, Tiny Image Net, Image Net) and mentions using "i-Blurry-N-M splits" and averaging results over 3 runs. However, it does not provide explicit percentages or counts for training, validation, or test splits (e.g., 80/10/10 split or specific numbers of samples for each).
Hardware Specification No The paper mentions that "All experiments were performed based on NAVER Smart Machine Learning (NSML) platform," but it does not provide specific details about the hardware used, such as GPU or CPU models, memory, or number of processors.
Software Dependencies No The paper mentions software components like "Adam optimizer," "Auto Augment," and "Cut Mix," but it does not provide specific version numbers for these or any other software dependencies, such as the deep learning framework used (e.g., PyTorch 1.9, TensorFlow 2.x).
Experiment Setup Yes For CIFAR10, we use a batch size of 16 and 1 updates per streamed sample. For CIFAR100, we use a batch size of 16 and 3 updates per streamed sample. For Tiny Image Net, we use a batch size of 32 and 3 updates per streamed sample. For Image Net, we use a batch size of 256 and 1 update per every 4 streamed samples. ... Adam optimizer with initial LR of 0.0003 is used. Exponential with reset LR schedule is applied for all methods except ours and GDumb, with ">γ = 0.9999 for CIFAR datasets and γ = 0.99995 for Tiny Image Net and Image Net. Ours use adaptive LR with γ = 0.95, m = 10 for all datasets...