incremental training (large data set)

Is the following approach suitable for NN incremental learning in general?

for bn = 1:num_batches   
    inputs = <get batch bn inputs>;
    targets = <get batch bn targets>;
    net = train(net, inputs, targets);
end

Is there any recomanded method how to split large data set on a separate batches and how to perform validation and testing?

I will be happy for any relevant reference and/or matlab examples (tutorial).

Answer 1

[Sai Bhargav Avula]

Hi,

You can try using tall array. Datastore object can also be used for this purpose. You can split the data using cvpartition for the validation and testing

Hope this helps!

[Michal]

My question is more abut general methodology of inceremental training. I am still not sure if the incremental learning is really correct and reliable method how to train NN for large date sets.

[Sai Bhargav Avula]

Iam bit confused what you meant by correct. To the best of my knowledge it depends on the application rather than the data. If you want the NN to adapt to individual environment you need to use incremental learning. If you want to continuously integrate new information then regular batch ML algorithms won't work. These algorithms reconstruct new models from scratch. This is the application of incremental learning.

[Michal]

I am sorry for a bit confusing description of my problem. Yes, I want to continuously integrate new information. Where new information is represented by new batches of data.

Is there any suitable approach how to solve this kind of problem?

[Sai Bhargav Avula]

if you are looking for the learning functions, the learngdm and learngd are the two functions that can be used for incremental learning.

If you are using ADAPT to train, BLF can be LEARNGD or LEARNGDM. PF can be MSE, MSEREG, or SSE. BTF is not used when using ADAPT to perform incremental training.

trains can be used for Sequential order incremental training with learning functions.

Hope this what you are looking for!

[Michal]

I am bit confused now. So, all these above mentioned methods are suitable for learning process which realize continuous integration of new information? The incremental learning is something else?

Could you provide any example or tutorial how to integrate new information represented by separate batch file to NN (shallow or deep) in a case of function approximation (regression) problem?

I need to train one NN by several batches of data represented by separate data files. So final NN should integrate information from all separate data batches.

[Sai Bhargav Avula]

I think this example explains things better

https://www.mathworks.com/help/deeplearning/ug/neural-network-training-concepts.html

[Michal]

OK ... and my type of problem correspond to the "Batch Training" or "Incremental Training with adapt"?

I am still not sure, which method is suitable for integration of new information to one NN. Is the proper method "Batch Training" in my case?

May be, should be better to communicate directly via email.