Why not using NN for online training?

In the course, it explained that " One model that easily supports online learning and has the ability to update it using stochastic gradient descent using mini-batches is logistic regression .", and then used NN to generate non-linear features.

My question is why we cannot directly use NN for online training?


This might be due to the dynamic nature of the learning environment. Something like RNNs or LSTMs might be more suitable instead of a simple NN for such predictions. But since this covers Machine learning, not Deep Learning, the author might have presented how to best incorporate NNs into a predictive model.