Reducing Catastrophic Forgetting in Neural Networks via Gaussian Mixture Approximation