Re: I think that relational DBs are dead. See link to my article inside

Hi Josip,

> But how about weighted graphs/synapses?
> Pointers/references don't have weights...

You are absolutelly right. Good question.

I have described logical model of OODB. It is not neural model. Neural model should be emulated. Of course it can be done by many ways.

If needed to emulate whole bilogical neuron I prefer this: Split one neuron into atomic pieces. Synapses, axons, dendrites, ... Pieces depends from used model. Then make some class for each piece. And then assemble all pieces into one neuron. So weght it is attribute of Synapse object not attribute of link. Synapses are linked to a neuron. And as i have restriction to make many links from one node to another with one color i use inverted identification. Synapses are linked not to soma, but to dendrite or axone. Dendrite or axone it is just anoter node. But from all axones all links are colored by target synapse ID. As each synapse should be unique it also should have an unique OID. And as synapses connected not to soma directly i can scan only input synapses (dendrite) or only output (axon)

It is some performance issue with this behaviour. So if i needed to emulate something simplier i use another way. OODB supports solid serialization. So it is possible to make some structure,

struct MyLink
{

    NativeHandle linkedNeuronID;
    float Weight;
    string SomethingUseful;
}

and have a collection of theses structures as field of SomeNeuron class. Then i implement IPersistent interface and save or load collection when instance is serialized.

If some neuron has about 1000 links to another neurones i will suggest second method. if neuron has 100000 or above links to anoter neurones i will wuggest first method. It is because solid serialization must serialize all links at once. And all of them required RAM when neuron is instantiated. Instead of this built in links collections can load only needed part of collection. Received on Thu Jul 06 2006 - 14:52:15 CEST