Fast-Inf: Ultra-Fast Embedded Intelligence on the Batteryless Edge

This is the repo associated to the paper Fast-Inf: Ultra-Fast Embedded Intelligence on the Batteryless Edge, to be published at ACM SenSys 2024.

To reproduce the experiments, please run:

python fff_experiment_mnist.py <leaf_width> <depth> <epochs> <norm_weight>

python fff_experiment_har.py <leaf_width> <depth> <epochs> <norm_weight>

python fff_experiment_speech_mfcc.py <leaf_width> <depth> <epochs> <norm_weight>

Note that, when norm_weight is set to 0, the model is equivalent to a fast-feedforward network, while setting it to a value > 0 makes the model will be trained with the L2 penalty as in the Fast-Inf paper.

Code, data, and results are available at: https://dagshub.com/leocus4/TinyFFF.

To cite this code:

@inproceedings{custode2024fast,
  title={Fast-Inf: Ultra-Fast Embedded Intelligence on the Batteryless Edge},
  author={Custode, Leonardo Lucio and Farina, Pietro and Yıldız, Eren and Kılıç, Renan Beran and Yıldırım, Kasım Sinan and Iacca, Giovanni},
  year={2024},
  publisher = {Association for Computing Machinery},
  booktitle = {Proceedings of the 22nd ACM Conference on Embedded Networked Sensor Systems},
  doi = {10.1145/3666025.3699335}
}