DBS-Adam Optimiser Improves Accident Severity Prediction
A new deep learning optimiser, Dynamic Batch-Sensitive Adam (DBS-Adam), dynamically scales learning rates based on batch difficulty, improving training stability and convergence speed. It is evaluated on Bi-Directional LSTM networks for vehicular accident injury severity prediction, addressing class imbalance with SMOTE-ENN and Focal Loss.
Key facts
- DBS-Adam uses a batch difficulty score from exponential moving averages of gradient norms and batch loss.
- It increases updates for difficult batches and reduces them for easier ones.
- Integrated with Bi-Directional LSTM networks for accident injury severity prediction.
- Class imbalance handled via SMOTE-ENN resampling and Focal Loss.
- Four experimental configurations compare baseline Bi-LSTM models and alternative architectures.
- Published on arXiv with ID 2605.15083.
- The optimiser is designed for imbalanced and sequential datasets.
- Aims to improve model efficiency and convergence speed.
Entities
Institutions
- arXiv