Generative vs. Discriminative Models
Вставка
- Опубліковано 16 жов 2024
- Python code example for GAN:
machinelearnin...
Bayesian Parameter Estimation:
• Bayesian Parameter Est...
Maximum Likelihood Estimation:
• Maximum Likelihood Est...
Image source:
www.flickr.com...
#MachineLearning #GenerativeModels #DiscriminativeModels #ArtificialNeuralNetworks #DeepLearning #ANN #SupervisedLearning #DataScience #ExplainableAI #GenerativeAdversarialNetworks #GANs #Deepfake
Is it predetermined to the Discriminator that the Generator always produces fake data (is the Discriminator biased into believing that Generator produces fake data)?
The Discriminator has no clue where the images are coming from. It only gets an image as input and has to figure out whether its real or fake.
Does an equilibrium exist where the generator wins i.e. discriminator is unable to distinguish the two distributions on finite samples?
The generator definitely wins in many cases, but I am not sure if one could call that "equilibrium". Interesting terminology though!
Do autoencoders come under one of these categories?
Not really since autoencoders are just trying to learn the representation of given data. But yes, autoencoders can be used to create GANs.
What are the applications of GAN apart from image and text generation?
Not much since video and audio generation is also related to image and text generation.
sir, in the video at 2:04, have you said discriminative instead of Generative?
Or am I understanding it wrong?
Yes, thats right! I should have said Generative.
Is there a threshold that the generator aspires to achieve in defeating the discriminator? is it possible that such a threshold might be unachievable in certain cases?
The threshold is determined by the human programmer and of course, it can be unachievable depending on how high its set.
Is GANs only used in binary classification or it can also be used in multi class classification? Also won't the computation will increase if we do multi class Classification using GANs.
I don't see any reason why GAN cannot handle multi class classification. Of course, computational requirements will increase.
As we have seen that deep learning models are usually racially, ethinically, etc. biased if the dataset is biased. So if we train the discriminator on a biased dataset, will it be biased towards the fake images generated by the generator? And if it is biased, we can't rely on the generated dataset to be used for training some other model.
The bias creeps in when we totally rely on Deep Learning models and remove/reduce human intervention. However, in GANs, we are usually looking for a specific kind of result and so currently GANs are highly human moderated. In other words, Deep Learning can fool you only if you don't know what you are looking for!
@@EvolutionaryIntelligence Thank you sir!