The injection of domain knowledge in deep learning models can boost their performance, especially in the case of complicated functions to be learned and scarcity of training data. For instance, in the domain of transprecision computing an open issue is learning the relationship between the precision assigned to the Floating-Point (FP) variables composing a benchmark and the associated error (measured as the difference w.r.t. the execution of the benchmark with all FP variable at maximum precision). DL models can be used to learn this relationships, albeit it is a very complex function, non-linear, non-monotonic and with plenty of local minima. For this reason, injecting domain knowledge in these DL models can increase their performance.
This repository contains the source code of the experiments conducted to demonstrate the benefits of injecting domain knowledge in DL model for transprecision computing. The experimental analysis has been described in detail in the paper "Injective Domain Knowledge in Neural Networks for Transprecision Computing", Borghesi et al., 2020, presented at LOD2020 (arXiv version: https://arxiv.org/abs/2002.10214).
For more details on transprecision computing and how to use DL models in that settings we refer to "Combining Learning and Optimization for Transprecision Computing", Borghesi et al., 2020, presented at Computing Frontiers 2020 (arXiv version: https://arxiv.org/abs/2002.10890).
- Python > 3.6
- TensorFlow 1.x
- Keras
- numpy
- scikit-learn
- pandas
- Spektral (https://graphneural.network/getting-started/)