Abstract
As sequential neural architectures become deeper and more complex, uncertainty estimation is more and more challenging. Efforts in quantifying uncertainty often rely on specific training procedures, and bear additional computational costs due to the dimensionality of such models. In this paper, we propose to decompose a classification or regression task in two steps: a representation learning stage to learn low-dimensional states, and a state space model for uncertainty estimation. This approach allows to separate representation learning and design of generative models. We demonstrate how predictive distributions can be estimated on top of an existing and trained neural network, by adding a state space-based last layer whose parameters are estimated with Sequential Monte Carlo methods. We apply our proposed methodology to the hourly estimation of Electricity Transformer Oil temperature, a publicly benchmarked dataset. Our model accounts for the noisy data structure, due to unknown or unavailable variables, and is able to provide confidence intervals on predictions.
Introduction
Deep learning models like LSTMs and Transformers have revolutionized time series forecasting, yet they often suffer from being "overconfident." In critical domains such as energy grid management or medical diagnostics, point predictions are insufficient; reliable uncertainty quantification is essential. Traditional approaches like Bayesian Neural Networks (BNNs) are computationally expensive, while simpler methods like MC Dropout often fail to provide well-calibrated confidence intervals. This paper introduces a decoupled architecture that separates high-level feature extraction from probabilistic modeling to address these challenges.
Methodology: The Two-Stage Approach
The core innovation lies in a Two-Stage Learning framework that bridges the gap between deep representation learning and rigorous probabilistic state space models (SSMs).
- Representation Learning Stage: A deep neural network (e.g., a 3-layer GRU) acts as an "input model," transforming raw input sequences $u_t$ into high-level features $\tilde{u}_t$.
- State Space Model (SSM) Stage: The standard output layer is replaced with a stochastic last layer. This layer models a hidden state $x_t$ that evolves over time and generates the final observation $y_t$, accounting for both process and observation noise.
To handle the mathematical intractability of the hidden states, the authors employ Sequential Monte Carlo (SMC), or particle filtering. By propagating a set of "particles" representing the state distribution, the model can estimate predictive distributions. Parameters are optimized using a Monte Carlo estimator of the score function, leveraging Fisher’s Identity to enable gradient-based learning.
Experiments and Results
The methodology was evaluated on the Electricity Transformer Temperature (ETT) dataset, specifically the ETTh1 subset. The goal was to predict hourly oil temperature based on various power load features.
Calibration and Accuracy
A key finding is the superior calibration of the Sequential Monte Carlo Layer (SMCL) compared to standard baselines. While deterministic models and MC Dropout often produce narrow, overconfident intervals, SMCL provides intervals that more accurately reflect the true data distribution.
Quantitative results demonstrate that SMCL achieves a Prediction Interval Coverage Probability (PICP) of 98%, significantly closer to the target than MC Dropout (59%). This means that when the model claims 95% confidence, the true value is actually within that range 98% of the time.
| Metric | SMCL (Ours) | MC Dropout (p=0.01) | HMM |
|---|---|---|---|
| RMSE (Lower is better) | 0.24 | 0.25 | 0.44 |
| PICP (Target: 95%) | 98% | 59% | 85% |
| Computation Time | 210ms | 193ms | 994ms |
Discussion and Conclusion
By decoupling representation learning from the probabilistic output layer, this approach avoids the "curse of dimensionality" associated with fully Bayesian networks. The SMCL adds negligible computational overhead compared to MC Dropout while providing vastly superior uncertainty quantification.
Future work may involve incorporating more complex transition structures (like LSTMs) within the state space layer itself or applying this method as a fine-tuning step for large pre-trained models. This research highlights that for sequential data, modeling latent noise is just as important as capturing complex temporal patterns.
Summary updated