5 ms·

My understanding is that it's not quite that simple. GANs have stability problems (and as a result somewhat out of favor atm) and if the fake detection mechanis

My understanding is that it's not quite that simple. GANs have stability problems (and as a result somewhat out of favor atm) and if the fake detection mechanism isn't a differentiable function itself no training can happen.

sigmoid10 4y ago
The fake detection mechanism (aka discriminator) is usually just another neural network and I bet that's the case here as well. So it must be differentiable and thus, if anyone ever gets a hold of it, it could be easily used to train a generator that will eventually fool the discriminator.
- cudgy 4y ago
  So, basically a NOT operator on a neural network. Does this even require a differentiation?
  - nl 4y ago
    A NN is trained to make something indistinguishable from reality as possible - so it can't tell the difference. The inverse of that will just claim reality is fake too. What you need a a deep fake and a real video of the same thing, then train on the difference. Clearly this is impossible - which is what makes the problem hard.
    - sigmoid10 4y ago
      You actually don't need that. You only need a set of real videos and a generator for fake ones. Then train the discriminator to tell these two classes apart and make use of its differentiability to update the generator in tandem with the discriminator.
      - anankaie 4y ago
        Would you run into class balance issues if you do not have a growing pool of initial videos?
        sigmoid10 4y ago
        Not if you have a balanced dataset. If not, you're already getting into trouble with GANs, which is why they were superseded by diffusers on image generation tasks.
  - rusticpenn 4y ago
    Not really a not operator. It’s more like tuning the generated models until the fake detector cannot detect it.
  - anankaie 4y ago
    It depends on how you set it up. You can use a very non-expressive valuation that assigns a score, but then you need to use reinforcement learning techniques to transform that score to a model update. Alternatively, if your valuation represents a differentiable metric function from your target you have a way of going directly from your output to a model update. The second way requires dramatically fewer update steps (usually) than the first. Thus - having your adversarial target be differentiable definitely helps, though it is possible to do even absent such a criterion.
jasonjmcghee 4y ago
It is accurate that GANs have stability problems, but they are absolutely being used today for solving problems similar to this (an output needs to be "improved"). Stable Diffusion produces faces- especially eyes that are malformed. You'll often see "restore faces" in online services which feeds the end result into a GFPGAN which is used to restore faces.
- anticensor 4y ago
  Diffusors are not GANs.
  - homarp 4y ago
    indeed. that is why you use GAN to fix the face generated https://www.reddit.com/r/StableDiffusion/comments/x33rs4/how_do_i_improve_faces/ https://www.reddit.com/r/StableDiffusion/comments/x33rs4/how...
snowpid 4y ago
While Gradient Descent needs differentiable functions, there are evolutionary algorithms that do not need this and can also train neural networks.

sigmoid10 4y ago

The fake detection mechanism (aka discriminator) is usually just another neural network and I bet that's the case here as well. So it must be differentiable and thus, if anyone ever gets a hold of it, it could be easily used to train a generator that will eventually fool the discriminator.

cudgy 4y ago
So, basically a NOT operator on a neural network. Does this even require a differentiation?
- nl 4y ago
  A NN is trained to make something indistinguishable from reality as possible - so it can't tell the difference. The inverse of that will just claim reality is fake too. What you need a a deep fake and a real video of the same thing, then train on the difference. Clearly this is impossible - which is what makes the problem hard.
  - sigmoid10 4y ago
    You actually don't need that. You only need a set of real videos and a generator for fake ones. Then train the discriminator to tell these two classes apart and make use of its differentiability to update the generator in tandem with the discriminator.
    - anankaie 4y ago
      Would you run into class balance issues if you do not have a growing pool of initial videos?
      - sigmoid10 4y ago
        Not if you have a balanced dataset. If not, you're already getting into trouble with GANs, which is why they were superseded by diffusers on image generation tasks.
- rusticpenn 4y ago
  Not really a not operator. It’s more like tuning the generated models until the fake detector cannot detect it.
- anankaie 4y ago
  It depends on how you set it up. You can use a very non-expressive valuation that assigns a score, but then you need to use reinforcement learning techniques to transform that score to a model update. Alternatively, if your valuation represents a differentiable metric function from your target you have a way of going directly from your output to a model update. The second way requires dramatically fewer update steps (usually) than the first. Thus - having your adversarial target be differentiable definitely helps, though it is possible to do even absent such a criterion.

jasonjmcghee 4y ago

It is accurate that GANs have stability problems, but they are absolutely being used today for solving problems similar to this (an output needs to be "improved"). Stable Diffusion produces faces- especially eyes that are malformed. You'll often see "restore faces" in online services which feeds the end result into a GFPGAN which is used to restore faces.

anticensor 4y ago
Diffusors are not GANs.
- homarp 4y ago
  indeed. that is why you use GAN to fix the face generated https://www.reddit.com/r/StableDiffusion/comments/x33rs4/how_do_i_improve_faces/ https://www.reddit.com/r/StableDiffusion/comments/x33rs4/how...

snowpid 4y ago

While Gradient Descent needs differentiable functions, there are evolutionary algorithms that do not need this and can also train neural networks.