2
$\begingroup$

In object recognition, translating an image by a few pixels in some direction should not affect the category recognized. Suppose we consider images with an object in the foreground on top of a uniform background. Suppose also that the objects of interest are always at least 10 pixels away from the image's borders. Is the following neural network invariant to translations of at most 10 pixels in some direction?

a) Neural network with two hidden layers consisting of convolutions (5x5 patches with a stride of 1 in each direction) followed by max pooling (4x4 patches with a stride of 4 in each direction) and a softmax output layer.

Here the translation is applied only to the foreground object while keeping the background fixed.

$\endgroup$

1 Answer 1

2
$\begingroup$

What kind of neural network are you using?

It could be invariant if the initial weights are always the same, but some NN can also have stochastic inner mechanisms (ex, noise and minibatches).

In all cases, you can use the same random root number to always find the same result.

If you use CNNs, you can run them in a deterministic way. It requires some setups.

For example:

https://pytorch.org/docs/stable/notes/randomness.html

$\endgroup$
2
  • $\begingroup$ I am Using Convolutional Neural Network $\endgroup$ Nov 18, 2023 at 16:45
  • $\begingroup$ I have updated my answer. $\endgroup$ Nov 20, 2023 at 10:14

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.