Class Discussion Summary (Feb 04)

Put the citations to papers discussed here:

Image-to-Image Translation with Conditional Adversarial Networks
Phillip Isola Jun-Yan Zhu Tinghui Zhou Alexei A. Efros 

Scribes:

Jiahao Xue

Oasys Okubo

Melanie Wong

Negin Majidi

Write the Class Discussion Summary Here:

* When might the ML or non-ML version be better? 

ML works better on data with high level features; if ML is trained with a large amount of data, it works better than non-ML version.

* If you played with the ML demo, does it work well?

Yes it works well, though I draw badly in Image-to-Image Demo, it can still recognize my paint and turn it into what I want

The accuracy could be improved with larger training data sets.

* Consider applications we have considered (low light: noisy<->clear) and (blur: blur<->clear). Do you think these methods can be applied for the domain mapping? Either direction? Only in one direction?

We think ML version can work in either direction, because ML is good at handling with multi-dimensional data. But the non-ML version might not be able to map in either direction. If there is only 1 direction it may work.

Write the Class Discussion Summary Here:

* When might the ML or non-ML version be better? 

For non-ML versions, it is better for deterministic problems. Problems that require high computational power may not be too optimal for using ML algorithms. For example, if we want to calculate a transformation of clouds from a weather forecast, it is not optimal for an ML algorithm to compute these things everytime. 

* If you played with the ML demo, does it work well?

We believe it works, but not well because it has a good outline of where eyes are, uses texture fillings to fill missing attributes well. A downside is that you have to have a very good outline of the cat in order for the model to output something close to a cat itself. Although, once this is overcome, we notice a very noticeable transformation of a cat from a pixel drawing.

* Consider applications we have considered (low light: noisy<->clear) and (blur: blur<->clear). Do you think these methods can be applied for the domain mapping? Either direction? Only in one direction?

For the first paper, we believe for blurring, going both directions is possible because we can infer a higher resolution photo from the blurred direction. However, for lowlight, it is not possible because features that lack light don’t offer much accurate predictions to estimate what might be at a spot that is not distinguishable.

For the second paper, we can do either direction. Since we have a loss function, we can perhaps inverse the loss function of the generative adversarial network to input an image forward and input the output image backwards to go either direction of the modeling architecture.