SiteAmp
Uncertainty-aware frequency-dependent predictions for single site amplification
This in an ungoing collarboartion with Dr. Chuanbin Zhu from Univ of Canterbury, New Zealand. Stay tuned for more in-depth findings.
Currently Bayesian deep recurrent models are adopted for single-site amplification prediction. Specifically, both sources of uncertainties, aleatoric and epistemic, are considered and reflected in the predictive uncertainties. Refer to the post uncertainty decomposition for a discussion in differentiating aleatoric and epistemic uncertainty in constructing predictive intervals. Importanly, in this research, accurate predictions are obtained by the model and we show the frequency-dependent variances.
Various frequency patterns have been learnt and reasonable uncertainty bounds have been estimated. The ground truth has been well captured.
For uncertainty measures, Fig. 6 shows the width measure (i.e. width between upper and lower bounds) of the 95% predictive interval with respect to each frequency, as an indication of the magnitude of uncertainty. Apparently, high frequency ranges have shown comparatively larger widths, suggesting the model’s uncertainty when predicting at high frequency ranges. Also, it can be seen that only very few times, mostly in the middle frequency range, that the PI has missed the ground truth value.
The residuals with respect to different frequencies can be seen below.