Exploring Per-Input Filter Selection And Approximation Techniques For Deep Neural Networks