metodj/eenn-avcs

Code for "Anytime-Valid Confidence Sequences for Consistent Uncertainty Estimation in Early-Exit Neural Networks" paper.

EENN-AVCS

Code for paper Anytime-Valid Confidence Sequences for Consistent Uncertainty Estimation in Early-Exit Neural Networks.

Setup

1. Clone or download this repo. cd yourself to it's root directory.
2. Create and activate python conda enviromnent: conda create --name eenn-avcs python=3.8
3. Activate conda environment: conda activate eenn-avcs
4. Install dependencies, using pip install -r requirements.txt

Code

• For sythetic data experiment, see regression_synthetic.ipynb
• For ALBERT experiment, first download precomputed logits from here and store them in data/ folder. Then see regression_albert.ipynb
• For MSDNet experiment, first download precomputed logits from here and store them in data/ folder. Then see classification_msdnet.ipynb

Acknowledgements

The Robert Bosch GmbH is acknowledged for financial support.