i am fairly new to machine learning, so this may be a silly question. if that is the case, I apologise in advance.

i am training a convolutional neural network on oceanographic images, which include both positive and negative anomalies. The implementation i have tested employs a number of 2D-convolutional layers, followed by max-pooling.

Max-pooling naturally has a bias towards positive values. Does it mean that negative anomalies in my images are going to be given less wight in the training? or is the sign difference accounted for in the convolutional filters?

instinctively, i would like to implement some sort of 'absolute-value max pooling', but since i have found no mention of this in the literature, should i assume this is not needed?


1 Answer 1


If you need to implement such an absolute value max pooling you can convert the output of the convolutional layer to an absolute value and then apply a maxpooling.

Moreover such thought is not necessary as maxpooling works with the output of the convolutional layers and is not influenced by the images directly. By which I mean that the convolutonal layers extract the most important features from the image be it positive or negative that plays a role in predicting a correct output

So for example if a low intensity pixel affects the identification more then it will be given a positive large weight compared to a pixel value that is less influential.

  • 1
    $\begingroup$ Thank you Anoop for taking the time to answer my question. Your reply makes sense. But there is something i still don't understand. The convolutional neural network I am using (which comes from a paper) uses tanh as activation. I understand that the purpose of using tanh is not to lose information about negative weights. But this seems to contradict the purpose of max-pooling. Maybe i am still missing something fundamental? $\endgroup$
    – Fabio
    Nov 17, 2020 at 2:17

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.