How to deal with this?

Question

Neural networks

I have only recently started doing neural networks. For learning purposes, I want to write it from scratch, without frameworks. But I don’t understand the following points: The result of training should be recorded somewhere (it shouldn’t be in the RAM), but what type of data and type of record should this file have? And how then to access these results (not by running through the entire file each time, if there are more than 3000 lines)? And how should the network behave if data was entered that was not there during training?

Answer 1

but what data type and record type should this file have

It is necessary to write/load only the values ​​of weights between neurons - the result of training.
Well, the code of your program itself contains all the global parameters of the network: how neurons are located and behave, how many there are, and so on.
All scales and other parameters must be in RAM, and read and rewritten every iteration of training.
Here, for example, replication trained by Alexnet (breakthrough network model from 2012),
https://github.com/BVLC/caffe/tree/master/models/b...
233mb.
Alexnet has 62 million parameters, 650 thousand neurons, and to train it for the same number of epochs as in the original article, you need to spend 5-6 days on two GTX 580 3GB video cards (Although the current record on faster hardware and optimized code is 24 minutes).
A network is, roughly speaking, just a function that is changed by a learning algorithm based on certain data. She considers something. With each iteration of training, a ready-made function is obtained, which stupidly counts the output from the input data.