Environmental Impact
Machine learning models offer a valuable approach to solving many challenges in the real-word, from solving complex engineering problems, to designing new innovative medical procedures, to aiding key business functions, to self-driving cars, to the construction of smart cities [1]. However, there are also environmental costs associated with machine learning that are becoming more prominent. Training and applying deep learning models requires electricity to power the CPU, GPU, and RAM electronics that are needed to perform the huge calculations that form the basis of training and using a neural network. This means such models have a carbon footprint that can be non-trivial. For example, to train the popular GPT-3 LLM it has been estimated that 1,287 MWh of electricity were required, with an associated carbon footprint of 552 CO2eq tonnes [2]. This is purely for training; inference use (day to day use of the models by users) will almost certainly be dwarfing this power consumption, although I could not find reliable estimates for this.
Given these concerns, I felt it worthwhile to generate carbon footprint estimates for deepTerra, which are presented below. These address a typical use case of training a model based on a 1 km sq satellite image patch, which maps onto 1296 224x224 individual images, then performing a prediction/test run on this model again using 1 km squared area (1296 images). The ResNet-50 model was used, with and without early stopping, and an architecture with 1,049,601 trainable parameters (weights). The Python plugin “codecarbon” was used to monitor power usage of these tasks. The experiments were run on a fairly modest “gaming” laptop, i.e. one with a dedicated GPU.
| Experiment | Power Usage (KWh) |
|---|---|
| 1296 training dataset, ResNet-50, early stopping | 0.00244 |
| 1296 training dataset, ResNet-50, no early stopping | 0.008686 |
| 1296 prediction/test run | 0.001537 |
Clearly, this represents negligible power usage. However, in the name of non-complacency I am building in power usage monitoring options in the tool for every major function, and for entire sessions.
[2] Luccioni, A. S., Viguier, S., & Ligozat, A. L. (2023). Estimating the carbon footprint of bloom, a 176b parameter language model. Journal of Machine Learning Research, 24(253), 1-15
[1] Nozari, H., Ghahremani-Nahr, J., & Szmelter-Jarosz, A. (2024). AI and machine learning for real-world problems. In Advances In Computers (Vol. 134, pp. 1-12). Elsevier.