They released GoogleAI researchers AutoBNN to address the challenge of effectively modeling time series data for forecasting purposes. Traditional Bayesian approaches such as Gaussian processes (GPs) and structural time series have not been able to overcome limitations in scalability, interpretability and computational efficiency. Neural network-based approaches lack interpretation and may not provide reliable uncertainty estimates. These issues create the need for a method that combines the interpretability of traditional approaches with the scalability and flexibility of neural networks.

Current methods for time series forecasting often involve either traditional Bayesian approaches such as GPs or methods based on neural networks. The proposed solution, AutoBNN, addresses these limitations by automating the discovery of interpretable time series forecasting models. It disables GPs for Bayesian Neural Networks (BNNs) while preserving the structure of the synthesis kernel. This makes it possible to combine the ease of understanding of traditional methods with the scalability and adaptability of neural networks.

AutoBNN is based on the concept of learned GP kernels, where the kernel function is defined complexly using basis kernels and operators such as Addition, Multiplication or Change Point. It translates this approach to BNNs by exploiting the correspondence between BNNs of unbounded width and popular GP kernels. AutoBNN introduces new kernels and operators such as the OneLayer kernel, ChangePoint, LearnableChangePoint, and WeightedSum, which enable the modeling of complex time series patterns. These elements enable structure discovery in a scalable manner, providing high-quality uncertainty estimates and improving the computational efficiency of traditional approaches.

In terms of performance, AutoBNN shows promising results in terms of prediction accuracy and scalability. AutoBNN is an effective tool for understanding and forecasting complex time series data because it automates the discovery of interpretable patterns and provides high-quality uncertainty estimates. Its ability to efficiently handle large data sets makes it suitable for a wide range of applications, from predicting economic trends to understanding traffic patterns and weather forecasts.

In conclusion, the paper introduces AutoBNN, a new framework for time series forecasting that combines the interpretability of traditional Bayesian approaches with the scalability and flexibility of neural networks. AutoBNN offers a powerful tool for understanding and forecasting complex time series data. With its promising performance and ability to efficiently handle large datasets, AutoBNN has the potential to significantly advance the field of time series analysis and forecasting.