Efficient Deep Approximation of GMMs