Generative modeling learns the underlying distribution of training data so a model can produce new data points resembling it. That is a different goal from a discriminative model, which learns boundaries between categories to support prediction or classification. Generative AI puts this capability to work creating text, images, video, and other content. section 10.3 section 1.3
A solution may begin with unsupervised learning, then use supervised or reinforcement learning to specialize for a task. Model choices include generative adversarial networks (GANs) and autoregressive transformer-based models. In a GAN, a generator and discriminator improve the realism of generated data through adversarial training. section 1.5 section 10.2 section 10.9
Generating useful text
A text model trained on large datasets learns patterns among words and sentences, including grammar, syntax, and meaning. It can draft an email, summarize a document, translate a language, or respond through a chat interface. The result still needs human review: generated text can be inaccurate or reflect biases in its training data. section 7.17
For a business use case, turn the idea into a model-selection and validation workflow:
- Define the objective, such as customer support, personalization, or content generation. section 10.6
- Check data quality, quantity, representativeness, and supporting infrastructure; then choose a pretrained model or a custom build. section 10.6
- Prepare the data, train or adapt the model, and validate its performance on the intended real-world use case. section 10.6
Generative modeling also appears in deep belief networks: their layers are pretrained to capture progressively more abstract features of the data. This is a more specific training approach, not a required step for every generative solution. section 8.9.1
