Online/temporal learning

Is it possible to update a model fit one data point at a time? More importantly, in what capacity is real-time time series analysis supported? Most/all of the examples are focused on static datasets with batch fitting, but I'm not sure if that's the only way to do it with MXNet.

Some topics that I'm interested in:
- LSTM-g
- HTM (Heirarchical Temporal Memory)
- Spiking RNNs
- Reservoir Computing (Liquid State Machines)

with applications in:
- Video sequence analysis
- Auditory processing
- Semantic analysis
- Data stream sequencing, non-linear econometrics

To what extent are any of these possible through MXNet? (Even if it's not possible today, would the framework support them naturally?) I would consider contributing some algorithms here, or creating standalone packages that could be used by MXNet or my experimental package: [OnlineAI.jl](https://github.com/tbreloff/OnlineAI.jl). If the MXNet framework does not support these topics naturally, I'll probably just continue to develop my OnlineAI framework independently.