Over the years, generative models have become more popular. These models are used to enable the handling of information that is missing, as well as when dealing with sequences that are of variable length. Essentially, generative models are concerned with models of distributions. These are then defined over points of data in a space that is high dimensional.
These are the seven types of generative models:
1. Bayesian Network
Also known as the Bayes Network, the Bayesian Network is a graphical modal of generative probability. It allows for a better representation of joint probability. Such a distribution is set over variables that are random. There are two main components to the Bayesian Network. These are structure, as well as parameters.
This structure is also known as a directed acyclic graph. The parameters include each node’s conditional probability distributions. The Bayesian Network can be used in various ways, such as in time-series predictions, anomaly detection, and more.
2. Gaussian Mixture Model
The Gaussian Mixture model is a model that is generative probabilistic. This model assumes that every single data point is generated from a combination of Gaussian distributions of finite numbers that have parameters that are unknown. Usually, this model is used as a probability model that is parametric. It can be used in a biometric system, which includes features such as spectral features related to the vocal tract, as well as a speaker recognition system.
3. Autoregressive Models
Also known as the AR model, this is when a bye that is from a time series regresses back on prior values within the same time series. The order present in an autoregressive model is the total number of values that were immediately before in the same series. These are then used to calculate the present time value. To put it simply, the AR model enables the prediction of future values based on values from the past. AR models have the capacity to handle a flexible range of patterns of time series.
4. Generative Adversarial Networks
Also known as GANs, Generative Adversarial Networks are a kind of generative model that has two components. The first of these are the generators, and the second, the discriminators. In this model, generative models are estimated using a process that is adversarial. The generative model is responsible for gathering the distribution of data. The discriminatory process, on the other hand, is responsible for estimating the probability pertaining to whether a sample was obtained from training data over the generative model.
Generative Adversarial Networks are popular among generative models that have already been used in the creation of images of people who don’t exist.
5. Hidden Markov Model
Also known as an HMM model, this is a statistical model that can be used to better understand how certain observable events evolve. These events should depend on factors that are internal, even if they can’t be observed directly. The Hidden Markov Model has found popular use in modeling correlations related to symbols that are adjacent, events, or even domains. The fields of digital communication and speech recognition are where the HMM model has found popular use.
There are only two processes that are stochastic in an HMM model. The first is a process that is invisible, pertaining to hidden states. The second is a process that is visible, pertaining to symbols that can be observed.
6. Variational Autoencoders
Also known as VAEs, Variational Autoencoders are most popularly used when it comes to approaches related to complicated distributions and learning that is unsupervised. A VAE is usually built atop approximators of standard function. These are networks that are neural, on which a gradient descent that is stochastic can be applied.
Through VAEs, users can generate all kinds of complicated data, such as those related to faces, or images that are CIFAR, and more. This model can predict the future using images that are static, and even more.
7. Latent Dirichlet Allocation
Also known as LDAs, this is a probability model that is generative. It contains discreet data, an example of which would be text corpora. An LDA is essentially a Bayesian model with three tiers of hierarchy. Each item in an LDA is a collection that is then modeled in the form of a finite mixture. These are pertaining to a set of topics underlying each item. There are various applications of this model, such as collaborative filtering, retrieving images based on content, and more.
These were the various generative models having an impact on artificial intelligence, as well as machine learning.
—
By Redwood Creative, Inc., a leader in online marketing services.
