3
$\begingroup$

I am trying to get to the logical intuition of differences between stacking multiple features and passing it via a final block (which could comprise multiple layers and lets say a final classification layer) vis-a-vis passing these features via different layers and concatenate them before passing it via a final classification layer.

To give you a concrete example, if I am classifying images of different types,

  • I might pass them through a pre trained model(weights) and get a set of features, extract other features of image like maybe some blobs, histogram etc , ensure they have the same dimensionality and stack them and pass them through 1 model

  • OR i might pass the extracted features through 1 FC layer and then , say, the histogram features through another FC layer and concatenate them and finally pass this concatened output through a final FC layer to get the classification.

i would presume, 2 diff models + concatenation should be applied when the features are co-related but have different characteristics. For e.g. stock prices AND monthly employment numbers BUT if they are coming from the same image, they should be stacked ?? quite confused and any pointers would be manna from the heavens

$\endgroup$
0

1 Answer 1

0
$\begingroup$

In general, Deep learning works when the data is based through as many layers as possible so the model can learn which non-linear combinations of features are most associated with the task.

In your image classification example, it would be best to have a large number of images and pass them unprocessed into the model. Then the model can automatically learn the most informative features.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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