Interpreting denoising autoencoders with complex perturbation approach

The goal of this study is to interpret denoising autoencoders by quantifying the importance of input pixel features for image reconstruction. The importance of pixel features is evaluated using the attributions of the pixel features to the latent variables of a denoising autoencoder used for image reconstruction. Pixel attributions are computed using a highly accurate and automatable perturbation approach and are plotted as saliency maps. Saliency maps highlight the contribution of the pixels for image reconstruction. The proposed approach produces more meaningful and understandable explanations than guided backpropagation and layer wise propagation methods. Three sanity checks are introduced to verify the fidelity of the generated saliency maps and also to elucidate the influence of inputs on the latent variables. The classification accuracy of images is significantly lowered when the most important pixel regions highlighted by the saliency maps are corrupted validating the proposed approach.