Reducing catastrophic forgetting in neural networks via gaussian mixture approximation