Andrew Hu February 2016

How different does each output unit have to be?

I'll use the example of classifying pumpkins. Take the example of the Cinderella pumpkin

enter image description here

Versus the gourd pumpkin

enter image description here

Intuitively, it may seem wise to classify these images as two different outputs, cinderella-pumpkin and gourd-pumpkin, due to how different they look.

My question is, if I take a training set of images that includes both cinderella pumpkins and gourd pumpkins and classify both of them under the category of pumpkin, will the performance of the network be worse than if I instead separated them into two categories? What is about the threshold for when two objects are so different that they should be put into separate categories?

Or to take a more extreme example for the sake of clarity, if I took pictures of cats and pictures of pineapples and classified them under the same category, how would the ability of the network be affected in classifying each respective object in comparison to if one created a cat output and a pineapple output?


Prune February 2016

It depends on the inherent similarity of your training observations. I don't set up a threshold: I use power iteration clustering (or other unsupervised classification) to guide me on where there are significant divisions in the training data. k-means is also a popular choice, since it's a common implementation, and relatively easy to comprehend.

The other consideration is the similarity of "non-pumpkin" data, such as a basketball (compared to your Cinderella). Again, I take the unsupervised learning approach. In this case, I expect that a basketball would plot closer to the Cinderella than either would to the gourd. This suggests separate classes for pumpkin types -- or perhaps more feature detection in the image processing, to find the similarities across pumpkin varieties.

Does that help?

Post Status

Asked in February 2016
Viewed 2,968 times
Voted 11
Answered 1 times


Leave an answer