Generative AI & LLMsGANs, VAEs, Transformers & Ethics
Unleash the science behind the generative revolution. Learn how Generative AI (classified as Weak AI) synthesizes novel text, images, video, and audio. Differentiate Generative vs. Discriminative models, master GANs & VAEs, explore Large Language Models (LLMs & Transformers), build a Python Gemini API chatbot, and address critical ethical concerns like deepfakes and APA citations.
7.1 & 7.2 Working of Generative AI: GANs vs VAEs
Generative AI creates fresh digital artifacts—images, code, video, audio, or prose—that resemble its training distribution. In the official CBSE handbook, Generative AI is categorized as Weak AI because it operates strictly on statistical probabilistic associations without subjective self-awareness.
1. Generative Adversarial Networks (GANs)
Two-Player GameComprises two neural networks trained simultaneously in an adversarial min-max competition:
- · Generator: Creates fake candidate samples (e.g. images) starting from random noise.
- · Discriminator: Inspects samples and tries to distinguish real data from generated fakes.
- · Through competition, the generator masters producing photorealistic outputs indistinguishable from real data.
2. Variational Autoencoders (VAEs)
Latent MappingFocuses on probabilistic data compression and smooth reconstruction:
- · Encoder: Compresses complex input data into a lower-dimensional representation called the latent space.
- · Decoder: Decompresses and reconstructs the data back from this latent space into its original form.
- · Excels at structured data generation, missing value imputation, and anomaly detection.
Next: Unit 8 · Data Storytelling
Combine Data + Narrative + Visuals using Freytag's Pyramid and real-world case studies.