Generative AI in the machine learning landscape

1. Generative AI in the machine learning landscape

In the last video, we got a first look at generative AI models. But how do they fit in with other types of machine learning?

2. Models that analyze

Many machine learning, or ML, models are known as discriminative models because they discriminate between different types of inputs. These models can answer closed-ended questions, which have a limited, predefined set of answers. For instance, we might train a model to identify if an image is of a puppy or a bagel. These models need to learn from training data, after which, they can guess a correct answer based on inputs, or recommend categories to group data. But that's all.

3. Bagels and puppies

In this example, let's train our model to discriminate between puppies and bagels by sharing labeled puppy and bagel pictures. We only have four pictures here, but in a real training setting, we could need millions of images to teach the algorithm how to tell the difference.

4. Guessing with confidence

With enough data, when the model sees a new image, it can generally tell the difference. But that's all--it can only express how confident it is that a picture is a puppy versus a bagel.

5. Models that imagine

In contrast, another type of machine learning model called generative models flips this on its head. Generative models guess what the data would be for a given prediction. They still require training, just like discriminative models. But they can generate new content that is similar to their training data. If we ask a generative model for a puppy image, we would get just that. It would look similar to what our discriminative model guessed was a puppy.

6. Mixing for effect

Generative AI integrates discriminative models, generative models, and other statistical techniques. But we can't mash them together haphazardly. The models must work together like parts of a machine to produce high-quality responses. Ultimately, these mashups can produce beautiful creative works, like the examples we saw previously.

7. Generative adversarial networks (GANs)

Let's take the example of generative adversarial networks, or GANs. This is a type of generative AI that trains a generative model and a discriminative model together. They compete with one another, one trying to trick the other. Afterward, they share notes and each gets better over time.

8. Bagel Puppy GAN

The generator creates confusing images, attempting to fool the discriminator. Meanwhile, the discriminator tries to guess correctly. After every round, they compare notes and each model learns from the results. The two models compete with each other over many rounds until the generator gets very good at creating bagel pictures that are so puppy-like or vice versa that they fool the discriminator.

9. Artificial general intelligence (AGI)

But where is this all headed? A long-time goal of the Artificial Intelligence community is to create generative AIs that exhibit human-like intelligence. Beyond just generating new data similar to previous data, such Artificial General Intelligence, or AGI, would have a broad range of knowledge about the world, be able to reason across different domains, possess social skills for interacting with humans, and have the ability to think creatively and reason critically. Finally, AGI would have other human-like cognitive competencies, such as sight and language.

10. Use the right tool for the job

Now that we understand how discriminative models, generative AI, and AGI relate, we can consider which type applies in different situations. Discriminative models can be used to predict the weather, categorize a large collection of books, or classify puppy and bagel images. Generative AI can create code for a website, customer service responses, or pictures of aquatic felines. AGI, which may arrive in the coming decades, would be able to complete traditionally human jobs on its own.

11. Let's practice!

Let's complete some exercises.

Create Your Free Account

By continuing, you accept our Terms of Use, our Privacy Policy and that your data is stored in the USA.