Neural network-oriented graph database

[anssiko]

The Pipcook, a front-end oriented DL framework talk by @WenheLI makes the following point on model storage issues in the web context:

And finally, we are thinking about model storage.
As you know that, a deep learning model essentially is a graph with some weights.
And if we can come up with, it's kind of like, we call it neural network orientated database.
Basically, it's a graph orientated database that stores the information in a graph format.
And in this way, we can definitely reduce the serialization overhead.
We are just trying to put some, like deep learning models into the indexedDB or some other regular database.

Questions:

• Can we improve IndexedDB to make it more amenable to storing graphs such as ML/DL models? What are the specific requirements that should be satisfied?
• What can we learn from graph databases such as Neo4j that would be transferrable to the web?

[yorkie]

@WenheLI @anssiko IMO the storage stuff should contain the following works:

1. define an intermediate representation for the deep learning models, it might be a web-oriented edition of ONNX intermediate representation spec.
2. define the ops, it's working with the WebNN spec.
3. define the update strategy, it contains how to load a model from a remote URL, how to patch, and the cache control.

The purpose of model storage and spec might make the inferencing libraries like tfjs, tvm, and others are able to generate more performant code & model for the Web.

Thus I don't think starting from IndexedDB and Neo4j might be a good choice, which would change the behavior of most DL libraries to work with, and can't resolve the real problems that we met today.