Generative and Classification Artificial Intelligence
There are various types of AI. The following two are very different:
Generative AI Models aim to learn the patterns and structures in the training data to create new, similar data points. I.e, Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs). They are often used in image synthesis and text generation.
Training Generative AI focuses on capturing the underlying distribution of the entire dataset via techniques like maximum likelihood estimation (MLE) or adversarial training and may include a latent space of a compressed and abstract representation of the data.
Proper training and quality data are crucial for generating diverse and high-quality samples. Proper interpretations of the learned representations in the latent space of generative models require expertise. These determine the trustworthiness of generated outputs.
Mode collapse may occur in Generative models, especially in the case of GANs (Generative Adversarial Networks). When this occurs, the generator fails to capture the full complexity of the data distribution and produces limited sample varieties.
Evaluating the performance of generative models is frequently via metrics like Frechet Inception Distance (FID) or Inception Score helps to evaluate the performance of generative model though they may not fully capture the quality, diversity, and relevance of generated samples.
Classification AI Models focus on learning the boundary between different classes or categories and the goal is to define similarities, contrasts and uniqueness of classes or labels. When the purpose is to rank classes, they may be called discriminative models. I.e., Support Vector Machines (SVMs), logistic regression, and neural networks for classification. They are often used in classification, regression, image and object detection, and sentiment analysis.
Classification models are trained on labeled datasets, where each input is associated with a specific output or class label, to maximize the probability of correct outputs via parameters. This involves understanding the conditional distribution of the output classes given the input data.
Classification models require effective feature representation and deep learning to distinguish between different classes. Techniques like gradient descent and back-propagations - with emphasis on minimizing the error between predicted and actual output - are commonly used to train Classification models.
