Dear experts,
I am now working on multi-output regression through deep learning.
The output is the vector of image pixel.
In this application, the smoothness of image is really important, but unsolved yet.
Some techniques to make the output vector to be more smooth is really needed, without post-processing through spatial filter.
Is there any method to solve this problem?
Thank you so much for your help
P.S. Currently, I am using L2 loss function with regularization, and another type of loss function can be a good solution.
Generally, to do a simple regression problem you can use a feed-forward network with M input pairs of (X,y) where X is a vector of parameters. y is a scalar, which is approximated by a single node on the output layer.
First you need to choose your activation functions for the hidden and output layer(s) (linear, sigmoid etc.).
Finally, train your network on labelled x,y pairs, test against your test set to check accuracy, and then run against unlabeled data to find approximate y values.
Very similar to a classification problem, except your y is scalar instead of a vector and values are not constrained to be 0
Hi,
you can use cross-entropy as loss function, also the training dataset should consist of enough observation to build robust model
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Regression