We introduce CANOLA, a novel framework for robust and stable corrupted label correction. Motivated by the strong generalization capability of DNNs, CANOLA rethinks when and how label correction signals should be applied during training. CANOLA decouples the learning phase from the correction phase to ensure label updates are cautiously guided by a mature, converged model state.
- Overview of CANOLA: An iterative pipeline alternating between Noise Transition Matrix Construction (Phase 1) via asymmetric co-training and Corrupted Label Correction (Phase 2) using noise-aware learning and loss-triggered soft relabeling.
The prepared datasets can be downloaded from the following link:
# Install dependencies
pip install -r requirements.txtpython train.py \
--ground_truth_path /path/to/data.csv \
--features_path /path/to/features.feather \
--batch_size 256 \
--seed 42If you have any questions, comments, or suggestions, please do not hesitate to contact us.
- Email: 22024505@vnu.edu.vn