Quantizing SoTA Models for Wastewater Prediction

In this paper, the authors explore the use of quantization techniques for compressing neural networks while minimizing performance loss in wastewater forecasting models. The study focuses on two state-of-the-art (SoTA) models, N-HiTS and TFT, which have shown promising results in predicting normalized wastewater volumes.
The authors evaluate the performance of these models using different quantization techniques, including N-HiTS MASE Loss and TFT MASE Loss, as well as N-HiTS Quantile Loss and TFT Quantile Loss. These metrics provide a comprehensive assessment of the models’ accuracy and robustness under various quantization scenarios.
The results show that both N-HiTS and TFT exhibit good performance across different quantization settings, with minor losses in accuracy. However, N-HiTS outperforms TFT in terms of model size and inference speed, making it a more attractive choice for edge devices.
To further analyze the performance of these models, the authors provide feature importances for both versions of TFT on the Duisburg dataset. These importances reveal that TFT’s attention scores are concentrated in specific parts of the data, highlighting its ability to focus on relevant features.
Overall, this study demonstrates the effectiveness of quantization techniques in compressing SoTA models for wastewater forecasting while maintaining their accuracy. The findings provide valuable insights for practitioners and researchers working in this field, particularly those aiming to deploy these models on edge devices.
By using quantization, the authors were able to reduce the model size and inference time of SoTA models without sacrificing too much accuracy. This is crucial for applications like wastewater forecasting, where fast and accurate predictions are essential for decision-making processes.
To put it simply, think of these SoTA models as recipes for cooking delicious meals. Just like a recipe needs to be concise and easy to follow to produce great food, these models need to be compact and efficient to make accurate predictions in wastewater forecasting. Quantization is the secret ingredient that helps these models stay small yet powerful, allowing them to cook up accurate predictions with minimal effort.
In summary, this study uncovers the potential of quantization techniques for compressing SoTA models in wastewater forecasting while maintaining their accuracy. The findings pave the way for more efficient and effective model deployment in real-world applications.