0
$\begingroup$

I have been working with neural networks for generation of a predictive model using a multivariate approach. I have come across Deep Learning (or Deep neural networks) as a tool to enhance the success of these models.

From what I have read so far though, it seems the most Deep Learning methods are catered towards image recognition/computer vision, for example ConvNet.

Anyone knows if there are any approaches in Deep Learning meant for model building (e.g. regression types) using several predictor variables, similar to a multilayer NN? I have read up on DBN, but am unclear if it can be similarly applied for my purpose.

Thank you for the advice.

$\endgroup$
0
$\begingroup$

I am working on a similar problem. I have a multivariate input that is not an image, and I am trying to predict a single valued output (1 output, so a regression problem). The way you can do this is to make the output layer have size=1, and a linear activation.

You are right that most success is in the image recognition realm, but you can apply the same techniques to your problem. For example, you can use a 1D CNN layer to recognize temporal patterns in input data. 2D CNN is useful more for spatial patterns, just like images. It is possible you can transform your data into an "image," using techniques such as STFT.

If in your problem you have input features that don't have these two relationships, then it is possible you wouldn't benefit from applying a CNN. You would be best off developing a multi-layer neural network with relu activations and a regularizer such as dropout or L1/L2 to help the model generalize. Just be sure the data is normalized properly so each feature has equal impact on the training.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.