1
$\begingroup$

I am very new in Machine learning, I recently implemented the spherical k-means, but finally I found a interesting point from the result. I used four datasets, they are minst, cifar-10, fashion-minst, and svhn. I was following the paper Learning Feature Representations with K-means (Coates & Ng, 2012) https://www-cs.stanford.edu/~acoates/papers/coatesng_nntot2012.pdf. I did not finish the deep learning yet.

See the following image, I tried the max pooling and average pooling to see if there any difference, but I found an interesting point, the dataset with the clear feature (edge, color, etc. of the object in the image ) has the high performance, the dataset without clear feature has low performance.

See the four pictures on the bottom, I extracted one image from each dataset. I ordered them according the performance. I tried search online, but I did not get any useful information about my question.

I am not sure whether the point is correct or not. Could any one please explain this to me. Thanks!

enter image description here enter image description here

$\endgroup$

1 Answer 1

0
$\begingroup$

Yes it is true that in these images follow the trend you pointed, but that is not the only reason for the drop in accuracy:

  • Image 1 is pretty simple, and binary so it is easier samples, not because of clarity but of simplicity.

  • Image 2 is a bit more complex, but it is still binary and can be learned easily.

  • Image 3 is simple, but it is noisy and (as you pointed out) does not have strong features, since you're not dealing with CNNs you should pre-process this kind of dataset by binarization or image sharpening after noise removal (if you do it before you will amplify noise). This probably was taken with bad light conditions.

  • Image 4 this is really complex image, it is on color and until now I couldn't figure out what it is lol

Well, you are not wrong in your idea, but taking conclusions that way may mislead you on the future. There are other reasons (as pointed above) for your findings. So:

  • Create a thesis (like you did)
  • Test your thesis by trying to change the observed phenomena (e.g. pre-process images to find stronger features)
  • Check if that improved your results, create a new thesis and repeat until satisfied with results
$\endgroup$

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.