Convolutional autoencoders and LSTMs - using deep learning to overcome Kolmogorov-width limitations and accurately model errors in nonlinear model reduction