In the task of White Dwarf+Main Sequence (WDMS) finding in massive spectral data release, convolution can significantly improve the classification accuracy by extracting hierarchical, translational-invariant features. In this paper, by designing one dimensional convolutional neural network (1-D CNN) which was further trained in a discriminative, supervised way, 12 kernels with stable numerical distributions were produced, helping to generate spectral feature maps of WDMS. To solve the problem brought by biased sampling in the WDMS training set, we proposed a learning principle called Anti-Bayesian Learning Paradigm (ALP) which was built on the basis of order statistics by implying a comparatively looser prior distribution of spectral types. And in the way of separating training spectra into several groups according to their signal-to-noise ratios (SNR), we analyzed the robustness of convolutional extraction process to spectral noise. Experimental results indicated that, (1) WDMS classification with 1-D CNN and ALP reached the accuracy of 99.0%±0.3%, which outperformed the classic PCA+SVM model. (2) Pooling after convolution operations relieved the negative impact of spectral noise by lowering resolution. (3) When the SNR was less than 3, more epochs were required to learn stable kernels; when the SNR was between 3 and 6, the spectral convolutional features was stable; when the SNR was greater than 6, the convolution process acquired higher stability to eliminate the negative impact of SNR on model performance.