Generative AI Models: Unlocking Creativity and Innovation

Discover the world of generative AI models, their core principles, applications in industries like entertainment and healthcare, and the future of AI-driven creativity.
By
Rahul Rego
September 6, 2024
5 min read

Introduction

In recent years, generative AI models and AI algorithms have emerged as a revolutionary force in artificial intelligence, reshaping industries and redefining what machines can create. Unlike traditional AI, which focuses on classification, prediction, and decision-making, generative AI is capable of producing new, original content—from text and images to music and video—by learning patterns from vast datasets.

This article delves into the core principles behind generative AI models, including how they function, the types of models that exist (such as Generative Adversarial Networks, Variational Autoencoders, and Transformers), and their groundbreaking applications across sectors like entertainment, healthcare, and design. We will also explore the ethical considerations, challenges, and future potential of this transformative technology, which is not only automating creativity but also empowering humans to push the boundaries of imagination.

PRINCIPLES OF GENERATIVE AI

Generative AI operates on the premise of learning the underlying distribution of training data to generate new, similar instances. The core idea is to model the probability distribution of the data and sample from this distribution to create new data points. This involves:

Learning the Data Distribution: Generative models estimate the probability distribution of the training data. This can be explicit, where the distribution is directly modelled, or implicit, where the model learns to generate data that resembles the training set without explicitly defining the distribution.

Generating New Data: Once the model has learned the distribution, it can generate new data points by sampling from this distribution. The quality and diversity of the generated data depend on how well the model has captured the training data distribution.

Before we get into the Generative AI models, let us understand the types of learning

The first stage of Generative AI is Machine Learning (ML). Machines learn in 4 ways:
- Supervised Learning
- Unsupervised Learning
- Reinforcement Learning

A model is defined simply as a program designed to perform a certain function or set of functions.

SUPERVISED LEARNING

Supervised Learning is when a model gets trained on a data set, provided by the programmer, also known as an AI Engineer or AI Developer. The model is provided with a dataset, known as a training dataset, which is labelled, which means that it has both input and output parameters. The algorithms learn to make a connection between the input and output.

The programmer then validates the learning by providing unlabelled data in the dataset, known as testing data which the algorithm then provides the labels based on what it has learned.

Supervised learning is further divided into 2 categories:

Classification – The model is trained to output a discrete category or class.  For example, classifying emails as spam or not spam, grading students based on their marks
- Regression – The model is trained to output a continuous variable which represents numerical values. For example, predicting the price of a house based on size, location and amenities.

UNSUPERVISED LEARNING

Here, there is no answer key to the data inputted into the machine. The machine must find patterns and organize the information on its own. As it analyses more data, its decision-making ability gradually improves and becomes more precise.

Unsupervised learning uses several different techniques, such as:

1. Clustering - Clustering involves grouping a set of objects in such a way that objects in the same group (or cluster) are more like each other than to those in other groups.

2. Association - Association techniques find interesting relationships or patterns between variables in large datasets.

3. Dimensionality reduction - Dimensionality reduction involves reducing the number of random variables under consideration, simplifying the dataset while retaining its essential information.

4. Anomaly Detection - Anomaly detection identifies rare items, events, or observations that raise suspicions by differing significantly from most of the data.

REINFORCEMENT LEARNING

Supervised and Unsupervised Learning (an Intuitive Approach) | by Metehan  Kozan | Medium

Reinforcement learning (RL) is a technique that trains software to make decisions for optimal results. It mimics the trial-and-error learning process humans use to achieve their goals, reinforcing actions that contribute to the goal and ignoring those that do not.

Key Concepts

  1. Agent: The learner or decision-maker, like a robot or a software program.
  2. Environment: The world in which the agent operates and makes decisions.
  3. State: A specific situation in the environment at a given time.
  4. Action: A decision or move the agent can make.
  5. Reward: Feedback from the environment in response to the agent’s action. It can be positive or negative.
  6. Policy: A strategy that the agent uses to decide which actions to take based on the current state.
  7. Value Function: A function that estimates the expected cumulative reward from a given state, helping the agent to make better decisions.


How It Works
Imagine you have a robot in a maze. The robot gets a reward for reaching the end of the maze and is penalized for hitting a wall or taking too many steps.
Over time, the robot learns the quickest and least obstructive way to reach its goal.
Reinforcement learning is classified into 2 types -

Model-based RL – The machine builds a model of the environment and uses it to plan its actions. This model could be transition based, where it predicts the next state, or reward based, where it predicts the reward.

Model-free RL – The machine does not build a model of the environment. Instead, it learns a policy which maps states to actions, or a value function – which estimates the expected rewards for each combination of states and actions.

In short, model-based RL relies on planning before acting, while model-free RL uses a simpler but potentially less efficient trial and error method of achieving the same goal.

SELF-SUPERVISED LEARNING MODELS

Self-supervised learning (SSL) is a machine learning approach where a model learns from the data itself, generating its own supervisory signals, instead of relying on external human-provided labels.Imagine you have a puzzle with some missing pieces. You learn to guess what the missing pieces look like by looking at the rest of the puzzle. Over time, you get better and better at filling in missing parts of any puzzle.

A screenshot of a computerDescription automatically generated

The above image is an illustration of a Self-Supervised NLP (Natural Language Processing) model which performs a sentence auto-complete task. The labelling of the dataset is done by the model itself.

A more complex example would be Convolutional Neural Networks (CNN) which deal with images. This uses tens or hundreds of variables.

Imagine you have a big picture, like a photo of your favourite pet. When you look at the photo, your eyes don’t look at the whole picture all at once. Instead, they focus on small parts of the picture. Your brain then puts all these together to form the big picture.

A Self Supervised Learning Model in this case filters out each “part” of the picture individually to create a set of labels, hence creating a labelled data-set which “pre-trains” the model.

The model then learns the relationship between the “parts” – such as colours, textures, brightness, edges etc and the bigger picture and can replicate it in the future with unlabelled data.

DEEP LEARNING

Deep learning is a branch of machine learning that harnesses deep neural networks—complex layers of interconnected nodes inspired by the human brain—to replicate intricate decision-making processes. Many of today's AI applications rely on some variant of deep learning for their functionality.

Deep Learning Basics

  1. Neural Networks: Deep learning is a type of machine learning that uses neural networks, which are inspired by the way the human brain works. A neural network consists of layers of nodes (neurons) that process information.
  2. Layers: In a neural network, there are multiple layers:some text
    • Input Layer: This layer takes in the raw data, like an image or text.
    • Hidden Layers: These layers process the data by applying mathematical transformations. Each hidden layer extracts increasingly complex features from the data.
    • Output Layer: This layer produces the final result, such as a classification (e.g., identifying a cat in a photo) or a prediction (e.g., predicting house prices).
  3. Deep Learning: When a neural network has many hidden layers (hence "deep"), it is called a deep neural network. These deep networks can learn to recognize very complex patterns in the data.

How Deep Learning Works

  1. Training: Deep learning models are trained using a large amount of labeled data. During training, the model adjusts its internal parameters (weights) to minimize the difference between its predictions and the actual labels.
  2. Backpropagation: This is the process of updating the weights in the network. The model calculates the error of its prediction and then propagates this error backward through the network to adjust the weights, making the model better over time.
  3. Activation Functions: These functions decide whether a neuron should be activated or not. They introduce non-linearity into the network, allowing it to learn complex patterns. Common activation functions include ReLU (Rectified Linear Unit), sigmoid, and tanh.
  4. Loss Function: This function measures how well the model's predictions match the actual labels. The goal of training is to minimize the loss function.

GENERATIVE MODELS

Discriminative algorithms try to classify input data given some set of features and predict a label or a class to which a certain data example belongs. Generative algorithms do the complete opposite — instead of predicting a label given to some features, they try to predict features given a certain label. Generative Models are a type of machine learning model that can generate new data that looks similar to the data they were trained on. Generative models are broadly classified as Explicit Generative Models and Implicit Generative Models

EXPLICIT GENERATIVE MODELS

Explicit Generative Models assume a probability distribution for the given data. Based on this probability distribution, the model performs a Maximum Likelihood Estimation function, which predicts the most probable output based on the probability density function (PDF) computed by the model.

Explicit Generative Models are further divided into:

1. Tractable density models, where the probability density function can be easily and efficiently be captured in a parametric function. E.g. Flow-based models, autoregressive models 

2. Approximate density models, where the PDF is approximated and there is no parametric function.

IMPLICIT GENERATIVE MODELS

An implicit generative model does not explicitly define a probability distribution for the data it generates. Instead, it learns to produce samples from an implicit distribution by transforming samples from a simple, known distribution (e.g., Gaussian noise) through a complex, learned function.

The best-known example of an Implicit Generative Model is Generative Adversarial Networks (GANs). Variational Autoencoders (VAEs), can also be implicit when they operate without defining the probability.

Variational Autoencoders (VAEs)

Think of VAEs like a smart drawing robot that can see a picture, then recreate it from memory.
First, it compresses the picture into a simple sketch (imagine drawing with fewer details). Then, it tries to redraw the picture from this simple sketch.

VAEs have 2 components:
Encoders, which compresses the data, and decoders, which tries to reconstruct the data as close to the original form as possible.

Dimensionality reduction is used to lower the dimensionality (the number of random variables)

9. Dimensionality Reduction — Single-cell best practices


The picture illustrates a very simple example – where a 3 dimensional cube is reduced to a 1 dimensional line.

Oxford Dictionary defines latency as “the state of existing but not yet being developed.” In Machine Learning terms, latent space is the term used to represent the most compressed form of the data.

Training process:
The encoder and decoder are trained together to minimize the reconstruction error and regularize the latent space to follow a known distribution (usually Gaussian).

Applications:
Generating new data samples similar to training data
Dimensionality reduction
Image and text generation

Autoregressive Models

Autoregressive models generate data one step at a time, with each step conditioned on the previous steps.

Training process:
These models are typically trained using maximum likelihood estimation, learning to predict the next data point given the previous ones.

For example, imagine writing a story one word at a time, but each new word depends on the previous words. If you start with “Once upon a time,” the next word might be “there,” and the next “was,” and so on. Each word is chosen based on the words that came before it.

Images are generated pixel by pixel, audio is generated wave by wave and text is generated word
by word, also known as GPT (Generative Pre-trained Transformer).

Temperature, Top-p, and Repetition Penalty? The secret parameters they  don't want you to know


The illustrated example shows how a GPT works. “Logprob” or log probability is simply the logarithm of probability, which ranges from negative infinity to 0. The closer to 0, the more probable the outcome.

After calculating the probability of the next token using a probability density function, the GPT outputs the most probable token. Flow-Based ModelsFlow-based models use invertible neural networks to transform simple probability distributions into complex ones that resemble the training data.
Imagine you have a simple shape, like a ball of clay. With flow-based models, you can stretch, twist, and mould this ball of clay into a new shape, like a star. The cool part is that you can also reverse the process exactly (invertible), turning the star back into the ball of clay without losing any details.

In machine learning, a flow-based model starts with a simple, easy-to-understand distribution of data (like the ball of clay). It then uses a series of transformations (the stretching and moulding) to turn this simple data into complex, real-world data (like the star). Because these transformations are reversible, the model can easily go back and forth between the simple and complex data.

This ability to transform and reverse accurately helps the model learn and generate new, realistic data, like pictures or text.

Generative Adversarial Networks (GANs)GANs consist of two neural networks, a generator and a discriminator, that are trained simultaneously through adversarial processes.

Training process:
As the name suggests, the generator tries to produce data that can deceive the discriminator.
The discriminator tries to correctly differentiate real data from fake data.

Both networks improve through this adversarial training until the generator produces highly realistic data.

Applications:
Image generation – e.g., generating realistic human faces
Creating art and enhancing images (e.g., StyleGAN for generating high-quality images)

GAN vs VAE: Differences, Similarities, Examples - Analytics Yogi

The above image illustrates the difference between VAEs and GANs

Diffusion Models

Diffusion models generate data by reversing a gradual noising process.

Noise” in machine learning is the part of the data set that serves no purpose or relevance and can cause inaccuracies in the output.

The model learns to differentiate noise from the “signal”, which is meaningful, relevant information in the data – the opposite of noise. So, while noise makes it harder for the model to learn useful patterns, the signal is what the model needs to understand the data and make good decisions.

Diffusion Model for Generative Image Synthesis

In diffusion models, the main idea is to teach a neural network how to undo the diffusion process. As the model trains, it figures out how to guess the noise that was added at each step of the process going forward. To do this, it uses a loss function that checks how close its guesses are to the real noise values.

Boltzmann Machine

Think of a Boltzmann Machine model as a neural network where every node has a relationship with each other. Each node has a state (on or off) and connections to other nodes with varying strengths (weights).

Boltzmann Machine Structure: In a Boltzmann Machine, the nodes are organized into two layers:

  • Visible Layer: Nodes that represent the input data.
  • Hidden Layer: Nodes that help discover patterns or features in the data.

Learning Process: The network learns by adjusting the weights between nodes. It tries to find a configuration where the overall "energy" of the system is minimized. Lower energy means a better fit for the data.

Energy and Probability: The Boltzmann Machine uses concepts from physics. Each configuration of the network has an energy level. The network explores different configurations, with lower-energy configurations being more likely.

Training: During training, the network adjusts its weights to reduce energy and better represent the input data. This is done using a method called "stochastic gradient descent."

Applications: Boltzmann Machines can be used for tasks like pattern recognition, classification, and data generation. They are the foundation for more advanced models like Restricted Boltzmann Machines (RBMs) and Deep Belief Networks (DBNs).


Let’s say you’re organizing a party.

You have called a number of people you know (visible layer) but many of them bring in friends you don’t know (invisible layer)

Mood states: These guests can be in a good mood (on) or in a bad mood (off)
Influences: The moods of these guests are influenced by other guests. Guests who do not get along can make both their moods worse. There are however some people who can make many people’s moods better, and they get along with everyone. These influences represent the connections (weight) between the nodes of the machine.

The goal of the party is to have the best mood possible. In this case, the mood of the party can be described as the summation of everyone’s mood.

Making the mood better: Guests keep mingling and adjusting their moods based on their conversations. If someone sees that their friend is happy, they might become happier too. This is like the Boltzmann Machine adjusting its states.

Finding the Best Vibe: The party tries different setups (different configurations of happy and not-so-happy guests) to find the one where the most people are happy. This process is similar to the network exploring different configurations to find the one with the lowest energy.

Learning from the Party

During the learning phase, the host (the training algorithm) keeps an eye on how the guests are interacting and which setups make the party the best. Over time, the host learns which guests should be talking to each other more (adjusting the weights) to keep the party vibe positive.

In the end, a Boltzmann Machine is like a party where guests (nodes) keep chatting and adjusting their moods to find the happiest, most enjoyable configuration (the best solution to a problem). It’s important to note that the “mood” mentioned in the above analogy can be described as inversely proportional to energy. So, the highest possible mood would be the lowest possible energy, which is the goal of the Boltzmann Machine.

APPLICATIONS OF GEN AI

FOUNDATION MODELS

The term “Foundation Models” is defined by by The Stanford Institute for Human-Centred Artificial Intelligence's (HAI) Centre for Research on Foundation Models (CRFM) as “any model that is trained on broad data (generally using self-supervision at scale) that can be adapted (e.g., fine-tuned) to a wide range of downstream tasks.”

PRINCIPLES OF FOUNDATION MODELS

  • Foundation models are trained on vast amounts of data. GPT-3 for example, was trained on 500,000 million words, about 10 human lifetimes of normal reading. It also includes 175 billion parameters. The development of these models involves enormous computational processing power, data and resources.
  • Due to the massive amounts of data involved, foundation models operate on self-supervised learning principles.
  • Foundational models are generalized in nature. This is so they are capable of being fine-tuned in as many different ways as possible.
  • Due to foundation models being generalized, they can be adapted in countless ways.

LANGUAGE MODELS

The most widely known application of Gen AI are Language Models. Language models are useful for a variety of tasks such as:

  • Text Generation – Language models can generate coherent and contextually appropriate text based on a prompt. E.g. chatbots, virtual assistant, content creation
  • Language Understanding – Language models try to interpret human language by predicting the next word in a sentence, completing sentences and filling in missing words. This function is crucial for tasks like machine translation, speech recognition, and sentiment analysis.
  • Contextual Prediction - Language models predict the likelihood of a sequence of words given the preceding context. This helps in tasks such as autocomplete suggestions in search engines, grammar correction in text editors, and predictive typing on smartphones.
  • Information Retrieval - They assist in retrieving relevant information from large volumes of text data by ranking and scoring documents based on their relevance to a query. This function is used in search engines, question answering systems, and recommendation systems.
  • Summarization - Language models can summarize lengthy documents or texts by identifying the most important information and generating concise summaries. This function is valuable for quickly extracting key points from articles, reports, or documents.
  • Labelling - Language models classify sentences into categories (e.g., spam detection in emails) or label entities (e.g., named entity recognition in text) based on the semantic meaning and context.
  • Language Model Evaluation - Language models are used to evaluate the quality and coherence of generated text, assess language model performance on benchmarks, and compare different models' capabilities.

Large Language Models (LLMs) are foundational language models, trained on vast amounts of data with billions of parameters.

Some notable LLMs are:

  • OpenAI’s GPT series – used by ChatGPT and Microsoft  used for generating natural sounding text
  • BERT – Developed by Google, BERT is primarily use for tasks where understanding the context of words and sentences is crucial, such as question answering, sentiment analysis, and language translation.

TEXT TO IMAGE MODELS

These models take a natural language input and produce an image that matches the description. These models use a language model to transform input text to latent representation i.e a condensed, essential description of data that captures its meaningful features for efficient processing and understanding. It then uses a generative image model, which produces an image according to that representation.

Notable Text to Image Models are: 

  • DALL-E series by OpenAI 
  • Adobe Firefly
  • Imagen and Imagen 2 by Google
  • Midjourney
  • Stable Diffusion
  • RunwayML 

AUDIO

Generative audio refers to the creation or synthesis of audio content using artificial intelligence and machine learning techniques. This field focuses on generating realistic and coherent audio waveforms that mimic natural sounds, musical compositions, speech, or other auditory experiences.

Applications of generative audio include:

  • Speech Synthesis and Voice Cloning: Generating synthetic speech for use in interactive voice response systems, audiobooks, and personalized virtual assistants.
  • Music Composition and Sound Design: Creating new musical pieces, soundtracks, and sound effects for movies, video games, and multimedia projects.
  • Interactive Audio Experiences: Designing interactive environments where audio responses adapt to user inputs or environmental changes.
  • Audio Augmentation: Enhancing audio recordings by removing noise, improving clarity, or adding effects using generative models.
  • Adaptive Soundtracks: Generating dynamic music tracks that adapt to the pace, mood, or context of a user’s experience in virtual environments or interactive media.

VIDEO

Generative AI can generate new video sequences, modify existing footage and enhance video content through automated processes.

Applications of Generative Video include:

  • Speech Synthesis and Voice Cloning: Generating synthetic speech for use in interactive voice response systems, audiobooks, and personalized virtual assistants.
  • Music Composition and Sound Design: Creating new musical pieces, soundtracks, and sound effects for movies, video games, and multimedia projects.
  • Interactive Audio Experiences: Designing interactive environments where audio responses adapt to user inputs or environmental changes.
  • Audio Augmentation: Enhancing audio recordings by removing noise, improving clarity, or adding effects using generative models.
  • Adaptive Soundtracks: Generating dynamic music tracks that adapt to the pace, mood, or context of a user’s experience in virtual environments or interactive media.

ROBOTICS

Generative AI is increasingly being integrated with robotics to enhance their capabilities, improve their efficiency, and enable more sophisticated interactions with humans and environments.

  • Autonomous navigation and path planning, through simulating the environment
  • Gen AI can simulate images and scenarios to train robotic arms to recognize objects and handle them correctly
  • Generative models like GPT-3 can be used to enable robots to understand and generate human-like language. In advanced cases, it can also simulate emotions and behaviours.
  • Generative models can predict potential failures or maintenance needs of robotic systems

COMPUTER-AIDED DESIGN

Computer-Aided Design (CAD) is the use of Generative AI to assist in the creation, modification, analysis, and optimization of a design. CAD software enables engineers, architects, and designers to create precise drawings and technical illustrations.

Applications of CAD:

  1. Engineering:some text
    • Mechanical Engineering: Design of mechanical components, assemblies, and systems, including detailed engineering drawings.
    • Civil Engineering: Planning and design of infrastructure projects such as bridges, roads, and buildings.
    • Electrical Engineering: Creation of electrical schematics and layouts for circuit boards and wiring systems.
  2. Architecture:some text
    • Building Design: Creation of architectural plans, elevations, and sections for residential, commercial, and industrial buildings.
    • Interior Design: Design of interior spaces, including furniture layouts, lighting plans, and finishes.
  3. Manufacturing:some text
    • Product Design: Design of consumer products, industrial machinery, and tools, including detailed specifications for manufacturing.
    • Tool and Die Design: Creation of tools, molds, and dies used in manufacturing processes.
  4. Automotive and Aerospace:some text
    • Vehicle Design: Design of cars, airplanes, and other vehicles, including detailed component and assembly drawings.
    • Aerospace Engineering: Design of aircraft, spacecraft, and related systems with precise aerodynamic and structural specifications.
  5. Animation and Game Design:some text
    • Character and Environment Modelling: Creation of detailed 3D models for characters, environments, and assets used in animation and video games.

FINETUNING VS PROMPT ENGINEERING

Traditionally, Foundation Models were finetuned to suit specific use cases. The Foundation Model is used as a base to create a more refined model with suitable parameters.

However in recent years, foundational models such as GPT-3 have enabled users to achieve their use cases through simply adjusting their prompts by adding instructions and context.

GEN AI – LEGAL AND ETHICAL CONCERNS

Generative AI has certainly had its fair share of controversies, such as 

  • Deepfakes are synthetic media in which a person's likeness is digitally manipulated to appear in a video, image, or audio clip, often making it seem like they are doing or saying something they never actually did or said. This opens the door to a lot of potential misuse, such as fake news, political manipulation, pornography and identity theft.
  • Copyrights – With the training of generative AI models on huge amounts of data, the nature of the data means it could infringe on the copyrights and intellectual property rights of other companies and people.
  • Data privacy concerns – The data used for training machines could expose sensitive information which could be misused. It is imperative for developers of pre-trained models and companies which fine-tune it to ensure that personally identifiable information is removed.
  • Sensitive information disclosure – While using platforms like ChatGPT, some users could carelessly upload sensitive details such as legal contracts, source codes, proprietary information etc 
  • Bias – AI models are only as good (or bad) as the data they’re trained on. This data is very much susceptible to societal biases. The differences in society are evident across different socio-economic groups and geographical locations. The programmers must try to account for as much diversity as possible in the dataset.
  • Lack of transparency – Given that AI systems are black-box in nature, the workings of such systems are difficult to figure out, even to the developers. There is inherently an unpredictable aspect of AI, which can lead to devastating consequences.
  • Jobs – Automation has reduced the workload of a huge amount of employees. There are justifiable concerns that AI will eventually lead to job losses. While AI will definitely open the door to a number of tech jobs, the rule of efficiency dictates that even more jobs would be lost to AI.

Conclusion

Generative AI represents a groundbreaking advancement in the field of artificial intelligence, offering transformative potential across a myriad of domains. From creating lifelike images and videos to enhancing human-computer interactions through sophisticated language models, generative AI is pushing the boundaries of what machines can create and achieve.

As we continue to harness the power of generative AI, it is crucial to navigate its challenges and ethical implications with care. The technology's ability to generate highly realistic synthetic content can lead to misinformation, privacy concerns, and security threats if misused. Therefore, responsible development, robust regulation, and ethical considerations must guide our journey forward.

Looking ahead, the promise of generative AI lies in its capacity to augment human creativity, improve efficiencies, and solve complex problems. By leveraging its capabilities for positive impact—whether in healthcare, education, entertainment, or beyond—we can unlock new possibilities and enhance the quality of life. As with any powerful tool, the key lies in our approach to its application: with innovation, integrity, and a commitment to the greater good, generative AI can be a force for remarkable progress in our world. If you’re interested in learning more comprehensively about Generative AI, check out our Full Stack Generative AI Course For Beginners

Share this post
Rahul Rego
Data Science & Machine Learning

Track your Learnings easily

Our platform makes it easy for you to track your progress. Stay on top of your learning journey with clear and simple progress updates!
Ready to become a Data Scientist that industry loves to hire? Apply Now. 
Explore Courses