No, the quality of visualization and the amount of overfitting are uncorrelated.
No, the quality of visualization and the amount of overfitting are uncorrelated. Both an overfitted and a perfectly generalizing model will return excellent faces of dogs. The method is however a way to tell whether the model is underfitted (because then you won't see dog faces). In the field of generative models, the way to tell whether a model is 'working', is by calculating the chance that an image in the validation set could have been generated by the model. The higher the total chance of your validation set, the better your generative model. These 'deepdreaming' methods (and I know, because I created one of them before Google's code was out [1]) are nice and pretty and stuff, but it's literally impossible to tell whether the stuff they generate is a copy from the train set or is truly original. And as long as there is no way to evaluate them based on a validation set, there is no basis to assume that any of them is actually working and not copying and melding images it knows. But then again, if they are copying and melding, but we can't tell, maybe it's good enough to pass a Turing test? After all, aren't we humans just copying and melding the stuff we know neither? And, the stuff these networks create are definitely good enough to be 'interesting'. [1] http://317070.github.io/Dream/
Posted by 317070 on Jul 30, 2015, 11:21 AM
Tags
Related issues
- That isn't really relevant for the tool starting the container.
- Why is there no link to the source?
- Something I wonder about in this domain is if this is the fault of the language.
- Namespace pollution is an annoying problem in Rust.
- Currently, no language ports of Timefold Solver are planned.
- I agree with you, almost all .NET code I write is within the .NET framework, yet when I look at C# repos, it's disheartening to see so ma...
Raise real-world problems, find people facing the same issue, and discuss solutions.5,000 issues0 comments
hackernews (Software)
That isn't really relevant for the tool starting the container.
Why is there no link to the source?
Something I wonder about in this domain is if this is the fault of the language.
Namespace pollution is an annoying problem in Rust.
Currently, no language ports of Timefold Solver are planned.
I agree with you, almost all .NET code I write is within the .NET framework, yet when I look at C# repos, it's disheartening to see so ma...
Why is there no simple workaround for this?
I have been using https://github.com/jacktasia/dumb-jump to quickly jump to definitions.
No, the quality of visualization and the amount of overfitting are uncorrelated.
No, the quality of visualization and the amount of overfitting are uncorrelated. Both an overfitted and a perfectly generalizing model will return excellent faces of dogs. The method is however a way to tell whether the model is underfitted (because then you won't see dog faces). In the field of generative models, the way to tell whether a model is 'working', is by calculating the chance that an image in the validation set could have been generated by the model. The higher the total chance of your validation set, the better your generative model. These 'deepdreaming' methods (and I know, because I created one of them before Google's code was out [1]) are nice and pretty and stuff, but it's literally impossible to tell whether the stuff they generate is a copy from the train set or is truly original. And as long as there is no way to evaluate them based on a validation set, there is no basis to assume that any of them is actually working and not copying and melding images it knows. But then again, if they are copying and melding, but we can't tell, maybe it's good enough to pass a Turing test? After all, aren't we humans just copying and melding the stuff we know neither? And, the stuff these networks create are definitely good enough to be 'interesting'. [1] http://317070.github.io/Dream/
Comments (0)
No comments yet. Start the discussion.