Inverting the Feature Visualization Process for Feedforward Neural Networks

This work sheds light on the invertibility of feature visualization in neural\nnetworks. Since the input that is generated by feature visualization using\nactivation maximization does, in general, not yield the feature objective it\nwas optimized for, we investigate optimizing for the feature objective that\nyields this input. Given the objective function used in activation maximization\nthat measures how closely a given input resembles the feature objective, we\nexploit that the gradient of this function w.r.t. inputs is---up to a scaling\nfactor---linear in the objective. This observation is used to find the optimal\nfeature objective via computing a closed form solution that minimizes the\ngradient. By means of Inverse Feature Visualization, we intend to provide an\nalternative view on a networks sensitivity to certain inputs that considers\nfeature objectives rather than activations.\n