Is the chart on:
https://nxtforum.org/index.php?topic=4139.msg137513#msg137513literal?
If so, how did you generate enough training data for a 2-5-1 NN that will have the desired output characteristics?
If not, and the number of hidden layers increases as you get more main nodes, what NN do you use? especially if there arent any that are precomputed?
I am assuming you are using the neural network to make routing decisions to find the fastest path. This is a cool idea, but I am not sure a 2-5-1 NN is anywhere near the capacity to handle all the real world scenarios.
James
James,
good question. Partly we've already answered it.
In fact, there is no any 2x5x1, the diagram only demonstrates the idea of the relationships within the kernel. Unfortunately, it does not reflect the interchangeability of nodes.
Today, according to our estimates with the current formula of the compound (with the current experimental coefficient) Node technology allows 10,000 leading nodes to keep a stable connection between themselves and with the 10,000,000 ordinary nodes at 10% load of bandwidth and fivefold excess of network stability level.
We are not yet ready to reveal some peculiarities, but as I said before we have moved away from conventional "linear" formula N*N.
We conclude this architecture is not suitable for MegaNet from a mathematical point of view. Yes, we are still far to a critical level which would jeopardize the system (which frankly none of the existed cryptocurrency has not yet reached and is unlikely to reach in the near future), but we are not willing to sacrifice transactions speed like other cryptocurrencies.
At the initial stage of the project we spent a few months on a number of experiments with different approaches and mathematical models, and eventually came to the concept of a neural network, as it solves our main tasks optimally.
Average human brain has approximately 86 billion neurons, wherein each neuron has up to 20,000 connections with other neurons. This allows the nerve impulse to find optimal chain and carry nerve impulse within millionths of a second. Furthermore, neurons within the site are interchangeable, which makes the network is almost perfect in terms of sustainability.
Unfortunately, the flat scheme does not able to demonstrate volumetric connections, so you're not yet able to see the beauty of the solution. So be patient
But I think you should realize that Node.js is only a transit technology, we may at any time to replace it with a more productive if it is more consistent with our tasks. 
