Generative AI and GANs
Author(s): Aarya Brahmane
The term Artificial Intelligence was coined for the first time way back in 1956 by John Mcarthy. But it took us nearly 60 years to experience AI in daily chores. Generative AI is a branch of AI that creates new data instances from data it is trained on and gives real-life data instances. The 3d Avatars or Memojis are used in daily communication. Software used in creating them creates so realistic images and videos that a normal eye would consider it as a real image.
Generative AI is the branch of AI, which uses available text, audio files, images, videos to create a whole new set of the same which seems to be true and perfect in its own senses. The algorithms understand the pattern in the data fed to it and create a new version from it. Generative AI models are mainly Unsupervised AI models.
The models understand the pattern in the data fed to it and generate a new “item” out of it. Until now most of the Machine Learning/ Deep Learning models were based on the Discriminative Model of doing things. Discriminative techniques/models predict what is next on conditional probabilities. These models learn about the boundary within the classes in a dataset to make the decision. Unlike it, the Generative Model works on finding the actual distribution of the dataset. They often use the Bayes theorem to predict the joint probability.
Since Generative models create new data instances, they are Computational expensive. Laying in laymen language, the Discriminative Model discriminates between the data and answers it, for e.g if the image is of a car or bike. While Generative Model generates new data. For e.g, if given features of a car, it would generate an image of the car. For these work, they work on correlation and the distribution is complex.
General Adversarial Networks (GANs)
GANs create new output from the pattern found in the training dataset. In it two neurons are pitched against each other and creating a new output. These two neurons are generators and discriminators. To further understand this concept we will use the analogy of a soccer team and a coach. When the eleven players of the team meet for the first time. They don’t play in a proper way, there are many mistakes because each player has his own style of playing, his own tactic in mind. So the coach trains the players and gives them a tactic a strategy and the team plays like a charm. For all the soccer/football lovers, remember the Tiki-Taka style of FC Barcelona! What does the coach does, observes the team, marks the strengths and weakness of players, the gaps in gameplay, give them feedback, and improvises the gameplay of a team?
Now compare the above analogy with GANs. The team which plays the game, that is a Generator. Because it creates a new football play and works the action. And the coach is a Discriminator as he finds the strengths and weaknesses and gives back feedback to improvise. A Discriminator has two important tasks, to discriminate within the data and give feedback for the same. Hence Generator can be defined as the neuron which creates new data resembling the data on which it was trained(after finding the pattern underlying). And Discriminator can be defined as that neuron that discriminates between good and bad data and gives feedback.
Generative AI is used extensively in a few of the AMAZON products, one of the famous products being AMAZON DeepComposer. It is a device that uses Generative AI to create music. One even without an ounce of knowledge of music can create melodious tunes. For more information on DeepComposer, one can visit the Amazon page.
Feel free to connect with me at: https://www.linkedin.com/in/aarya-brahmane-4b6986128/
A must-watch series on AI, which is easy to access on YouTube. One can learn about the lengths at which AI is being used to make the world a better place:https://www.youtube.com/playlist?list=PLjq6DwYksrzz_fsWIpPcf6V7p2RNAneKc
Published via Towards AI