1: \begin{abstract} 

2: Humans possess a remarkable capability to make fast, intuitive decisions, but also to self-reflect, i.e., to explain to oneself, and to efficiently learn from explanations by others. This work provides the first steps toward mimicking this process by capitalizing on the explanations generated based on existing explanation methods, i.e. Grad-CAM. Learning from explanations combined with conventional labeled data yields significant improvements for classification in terms of accuracy and training time. %Furthermore, learning concurrently from both input/output pairs and explanations leads to faster convergence.

3: %We provide first steps to transcend the backpropagation learning paradigm. For classification, we perform a two step process by first classifying the object, computing activation weighted gradients and performing a second pass with this information to obtain the ultimate classification. This leads to remarkable performance improvements on multiple benchmarks.

4: \end{abstract}

5: