Most approaches to deep neural network compression via pruning either\nevaluate a filter's importance using its weights or optimize an alternative\nobjective function with sparsity constraints. While these methods offer a\nuseful way to approximate contributions from similar filters, they often either\nignore the dependency between layers or solve a more difficult optimization\nobjective than standard cross-entropy. Our method, Mutual Information-based\nNeuron Trimming (MINT), approaches deep compression via pruning by enforcing\nsparsity based on the strength of the relationship between filters of adjacent\nlayers, across every pair of layers. The relationship is calculated using\nconditional geometric mutual information which evaluates the amount of similar\ninformation exchanged between the filters using a graph-based criterion. When\npruning a network, we ensure that retained filters contribute the majority of\nthe information towards succeeding layers which ensures high performance. Our\nnovel approach outperforms existing state-of-the-art compression-via-pruning\nmethods on the standard benchmarks for this task: MNIST, CIFAR-10, and\nILSVRC2012, across a variety of network architectures. In addition, we discuss\nour observations of a common denominator between our pruning methodology's\nresponse to adversarial attacks and calibration statistics when compared to the\noriginal network.\n