Speaker
Description
Convolutional Neural Networks (CNNs) are now widely deployed in astronomy, often claiming superior performance over classical linear estimators for an extensive range of data applications. Yet it remains unclear when a CNN can genuinely outperform the matched filter (MF) -- which attains the Cramér–Rao Lower Bound (CRLB) as the minimum-variance unbiased estimator (MVUE) under Gaussian noise, and remains the best linear unbiased estimator (BLUE) in general. We test a ResNet-based CNN against the MF using image noise properties as the sole performance discriminator, and ask: when does a CNN's advantage reflect genuine information beyond the MF, rather than the bias-variance tradeoff inherent in any trained regression model? We address this through a series of realistic noise scenarios encountered in mm/submm astronomy (e.g., CMB surveys), demonstrating the surprising robustness of matched filtering, and isolating two distinct regimes where a CNN offers genuine advantage. The first is the conditionally Gaussian regime: per-image covariance variation that opens a modest CNN edge through adaptive noise modeling. The second is the genuinely non-Gaussian case, where the CNN exploits higher-order noise statistics to achieve lower-variance estimation than any linear filter. We further show that the advantage is template-dependent: governed by the spectral overlap between the source profile and the noise departure from Gaussianity, with distinct scaling for compact versus extended sources.