With the plentiful information available in the Gaia BP/RP spectra, there is significant scope for applying discriminative models to extract stellar atmospheric parameters and abundances. We describe an approach to leverage an `Uncertain Neural Network' model trained on APOGEE data to provide high-quality predictions with robust estimates for per-prediction uncertainty. We report median formal uncertainties of 0.068 dex, 69.1K, 0.14 dex, 0.031 dex, 0.040 dex, and 0.029 dex for [Fe/H], $T_\mathrm{eff}$, $\log g$, [C/Fe], [N/Fe], and [$\alpha$/M] respectively. We validate these predictions against our APOGEE training data, LAMOST, and Gaia GSP-Phot stellar parameters, and see a strong correlation between our predicted parameters and those derived from these surveys. We investigate the information content of the spectra by considering the `attention' our model pays to different spectral features compared to expectations from synthetic spectra calculations. Our model's predictions are applied to the Gaia dataset, and we produce a publicly available catalogue of our model's predictions.
Comment: 22 pages, 14 figures, 5 tables. Accepted by MNRAS