From Predictive To Generative: A Brief Introduction to GenAI
Those two pictures above are neither real nor photoshopped; they were generated using Stable Diffusion, a deep learning, text-to-image model released in 2022 based on diffusion techniques, which is you can call it as an artificial intelligence (AI).
It’s so interesting to me how AI has been developing, from just an email spam checker to a generative program that can produce images like those. It’s fascinating, yet also horrifying. Now, we find ourselves wondering: what exactly is AI? And why is this article talking about something called predictive and generative? What do these terms mean in the world of AI?
Disclaimer: This post isn’t intended to explain the entire complex process of generative AI, but rather to provide a simplified/brief introduction for those who may be confused about this technology.
Predictive AI
Most machine learning tasks involve making predictions using statistical formula based on training data. We’re talking about some tasks like classification, clustering, regression, or even dimensionality reduction. If you’re still finding this a bit confusing, you can check out my article on understanding the Naive Bayes classifier here: Naive Bayes Classifier, Explained | by Jericho Siahaya | MLearning.ai | Jul, 2023 | Medium. This will give you a solid grasp of how machine learning operates.
Machine learning is like teaching computers to learn from examples and make their own decisions. It’s a part of AI that helps computers predict things or make choices without being told exactly what to do.
In essence, machine learning and predictive AI go hand in hand. Machine learning techniques are commonly used to build predictive models. These models analyze historical data and learn patterns to make predictions about future events or outcomes. For example, predicting stock prices, weather forecasts, disease outbreaks, or customer behavior in various industries.
So, when you talk about generating a new image or even when you use the LLM like ChatGPT or Google Bard, you’re using AI but not machine learning anymore, instead you’re using another subset of machine learning called deep learning which involves neural networks with many layers. This deep learning approach serves as the foundation for exciting advancements in Generative AI.
Generative AI
After the machine learning, we have the deep learning which is a neural network with many layers (hence “deep”) to learn complex patterns and representations from data. These neural networks, modeled after the human brain’s structure, are particularly powerful in learning intricate features from large and diverse datasets.
Generative AI, as an advanced field, builds upon deep learning techniques and focuses on creating original, novel content.
Models like Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) are examples of how deep learning is foundational to generative AI. GANs, for instance, use two deep neural networks (a generator and a discriminator) that compete and improve to generate highly realistic data, such as images.
To simplify, think of a painter creating a beautiful mountain scene that doesn’t exist. How do they do it? They use their imagination, drawing on memories of real scenes they’ve seen before. Humans can create something fake based on something real because our brains process information in intricate ways.
When we see something, our brains perceive it and process it. That’s similar to what neural networks aim to do: process input in a detailed way, mimicking the human brain. Generative AI uses this complex process to create new fake things.
Gen AI needs Pre AI too
While there are some distinctions between generative AI and predictive AI, the truth is that generative AI wouldn’t exist without predictive AI in the first place. This is because within generative AI lies a more intricate form of predictive AI, operating with a much more complex mechanism.
Let’s consider ChatGPT, currently the most widely used LLM. ChatGPT doesn’t actually create new sentences; it predicts the next one. It might seem like it’s generating new sentences, but it’s really just predicting the most fitting output for the next word or sentence. Generating a new photo follows a similar process. The machine predicts the most suitable embedding for the next pixel until the entire fake photo, which doesn’t exist in reality, is complete.
Gen AI solve things by creating
We tend to view generative AI as a distinct subset of machine learning in terms of usage and purpose. It’s often perceived as not primarily focused on problem-solving or task automation, but rather as a tool for generating content and aiding human creativity.
However, generative AI encompasses more than just that. It also addresses problems and tasks in higher domains, look into deeper and providing more resourceful solutions for humans.
TL;DR
- Predictive AI focuses on solving problems, while Generative AI concentrates on creating.
- Generative AI creates things to address issues in more complex domains with much more complex mechanisms.
- It’s like you have a subset of numerical data: 1, 4, 6, 9, 2. Then, you ask AI to come up with the next number, but instead of using the previous data, it generates a new, unseen number that doesn’t exist in the dataset. For example, it might produce the number 5, even though 5 was not originally in the dataset.
