Energy-efficient neural network training through runtime layer freezing, model quantization, and early stopping

Background: In the last years, neural networks have been massively adopted by industry and research in a wide variety of contexts. Neural network milestones are generally reached by scaling up computation, completely disregarding the carbon footprint required for the associated computations. This trend has become unsustainable given the ever-growing use of deep learning, and could cause irreversible damage to the environment of our planet if it is not addressed soon. Objective: In this study, we aim to analyze not only the effects of different energy saving methods for neural networks but also the effects of the moment of intervention, and what makes certain moments optimal. Method: We developed a novel dataset by training convolutional neural networks in 12 different computer vision datasets and applying runtime decisions regarding layer freezing, model quantization and early stopping at different epochs in each run. We then fit an auto-regressive prediction model on the data collected capable to predict the accuracy and energy consumption achieved on future epochs for different methods. The predictions on accuracy and energy are used to estimate the optimal training path. Results: Following the predictions of the model can save 56.5% of energy consumed while also increasing validation accuracy by 2.38% by avoiding overfitting.The prediction model developed can predict the validation accuracy with a 8.4% of error, the energy consumed with a 14.3% of error and the trade-off between both with a 8.9% of error. Conclusions: This prediction model could potentially be used by the training algorithm to decide which methods apply to the model and at what moment in order to maximize the accuracy-energy trade-off.