Modeling with Artificial Neural Networks

[Andy_20]

Hello there,
I'm to using ANN but I have read some studies and show that ANN is useful for Forecasting...

My concern is that, "What other area can ANN be useful?"
Can we use for estimating relationship among variables? What are the parameters and how do we interpret them?

Many Thanks!

[sbxkoenk]

Hello,

The term "Artificial neural network" refers to a biologically inspired sub-field of artificial intelligence modeled after the brain.
All the neural networks you use in SAS (and other software) are "artificial".

NN is a vast field.

• You can use NN for supervised learning, but also for unsupervised learning (like auto-encoders)
• You can use NN for forecasting indeed. But not the classical feed-forward neural networks. You need recurrent neural networks (RNN) for that. A recurrent neural network (RNN) is a class of artificial neural networks where connections between nodes can create a cycle, allowing output from some nodes to affect subsequent input to the same nodes. Long short-term memory (LSTM) networks are popular RNN.  
• Neural networks are used a lot in computer vision (CV) as well. See f.e. Convolutional neural networks!
• Neural Networks are also used in generative AI. See f.e. generative adversarial networks (GAN).
  You can synthetically generate artwork
  or you can mimic (sensitive) data with AI to generate synthetic data twins.

Good luck,
Koen

[Andy_20]

Hello Koen,
Did you mean that the Classical Feed forward NN is not for forecasting but rather the RNN or LSTM?

My study is based in Agriculture field and want examine relationship between crop yields among selected factors by using NN... Is there any class NN appropriate for this? Again, after this, I'm planning to predict yields in a near future..!

Thank you.

[sbxkoenk · Accepted solution]

Hello,

If you want to predict crop yields based on (selected) factors

, then you are doing classical predictive modelling on cross-sectional data.

It's NOT forecasting ... Forecasting is a sub-field of predictive modelling where you work on time series data.

( If you have time-series data for several cross-sections then you are dealing with panel data )

Because you are thinking about using NN , I suppose you are working on observational data that were collected in the past in the context of normal agricultural activities.
If the data still has to be collected, it's better to set up an experiment, like here :
     SAS/QC 15.2 User's Guide
     The OPTEX Procedure
     Example 15.9 Optimal Design in the Presence of Covariance

     (this is about comparing the effects of seven different fertilizers on crop yield)
https://go.documentation.sas.com/doc/en/pgmsascdc/9.4_3.5/qcug/qcug_optex_examples09.htm

But OK, let's assume you already have your data in a data set.

For predictive modelling, basic NN architectures will do.
Try multilayer perceptron first.

See also here :
https://www.sas.com/en_be/industry/agriculture-analytics.html

And one more thing :
due to the fact lots of weights (parameters) have to be estimated in NN, you need a lot of data (a lot of rows) to get acceptable confidence limits around your parameter estimates.

If you do not have a lot of data (let alone having room for data splitting)

, it might be better to turn to classical statistical techniques like mixed modeling or regression (without random effects).

Kind regards,

Koen

[sbxkoenk]

Something extra ...

Neural networks demystified
By Leo Sadovy on SAS Voices March 23, 2016
https://blogs.sas.com/content/sascom/2016/03/23/neural-networks-demystified/

Koen