Hyperspectral images are corrupted by a combination of Gaussian-impulse noise. On one hand, the traditional approach of handling the denoising problem using maximum a posteriori criterion is often restricted by the time-consuming iterative optimization process and design of hand-crafted priors to obtain an optimal result. On the other hand, the discriminative learning-based approaches offer fast inference speed over a trained model; but are highly sensitive to the noise level used for training. A discriminative model trained with a loss function which does not accord with the Bayesian degradation process often leads to sub-optimal results. In this paper, we design the training paradigm emphasizing the role of loss functions; similar to as observed in model-based optimization methods. As a result; loss functions derived in Bayesian setting and employed in neural network training boosts the denoising performance. Extensive analysis and experimental results on synthetically corrupted and real hyperspectral dataset suggest the potential applicability of the proposed technique under a wide range of homogeneous and heterogeneous noisy settings. INDEX TERMS Bayesian estimation, discriminative learning, Gaussian-impulse noise, hyperspectral imaging, residual network.