What are Generative Adversarial Networks?
Generative Adversarial Networks (GANs) are a type of AI that learns from a set of training data and then makes new data samples of its own. Each GAN has two neural networks, which are programs that learn patterns from examples. One is called the generator and the other the discriminator, and the two compete in a kind of game. The generator makes new samples. The discriminator checks each one and tries to tell whether it is real or fake.
Training a GAN is a game in which the two networks take turns. First, the generator makes a new sample and passes it to the discriminator. The discriminator then tries to label it as real or fake. If it gets the label right, the generator is said to have lost the game. If it gets it wrong, the generator is said to have won.
The aim is for the generator to learn, over time, to make samples that fool the discriminator into thinking they are real. Once the generator can do this, it can be used to make new samples that may look real and be useful. But looking real does not prove that a sample is true, or that it suits a given use.
In AI and machine learning, GANs have been used for many tasks, such as making images, video and text.
How do they work?
GANs are used to make lifelike synthetic data. That is data made by a computer rather than collected from the world. They have two parts: a generator network and a discriminator network.
The generator takes in noise, which is just random input. It tries to turn this into data real enough to fool the discriminator. The discriminator is shown both real data and fake data from the generator, and it tries to label each item as real or fake. This sets up a contest between the two. One tries to make data real enough to fool the other. The other tries to get better at spotting fakes. During training, each network tries to get better at its own task. But that does not mean both keep improving. Training can be unstable, as the limits below explain.
One of the great things about GANs is that they can make data that would be hard or even impossible to get in any other way. For example, they have been used to make lifelike pictures, videos and even 3D models. This can be a big help when training machine learning models. It lets you train them on far more data.
You can also try out other kinds of output by changing the training data, or by conditioning the GAN on traits you choose. Conditioning means giving the GAN extra input that steers what it makes. But a standard GAN aims to learn the overall pattern of its training data. You cannot rely on it to make data with any new pattern you ask for.
There are a few things to keep in mind when using GANs
First, GANs can be very sensitive to the settings used to train them, known as hyperparameters. During training, the two networks can also struggle to reach a stable balance.
Second, GANs can suffer from mode collapse. This means the generator learns to make only a small part of the range of possible data. You can ease this in a few ways. You could change the way the model is trained, or use a different loss function, which is the score the model tries to improve. But just adding more noise to the input does not ensure that the model will cover more of the possible outputs.
Finally, GANs can be hard to make sense of, because there is no clear way to map the noise that goes in to the data that comes out.
GANs are fairly new, and a lot of research is still going on. As more research is done, they may become more robust and easier to use.
What are their potential applications in advertising?
In marketing, GANs offer new ways to make image, text, video and audio ads. They are one kind of generative AI, the wider family of AI that makes new content. The text and dialogue uses discussed here need a model suited to them. A tool that writes copy is not always a GAN.
GANs have been used to make lifelike pictures, videos and even sound recordings. Marketers can use them to create new ads that may feel real and convincing to the people who see them. Whether they do this better than older methods needs to be tested.
For example, think of a firm that sells a product made to reduce wrinkles. It could use a GAN to make an image ad that shows a person’s face before and after using the product. The GAN would produce both the before and after pictures, based on real data samples. The result could look real, just as a picture made with older tools such as Photoshop could. But neither proves that the product caused the change shown.
In simple terms, the model is trained on before-and-after photos of wrinkles. It learns to mimic both the before and the after, and then uses what it has learned to make new before and after pictures. The people in them are not real, and the improvement in wrinkles is only a generated depiction of change. The fact that the model learned from real photos does not prove that this product gives the result shown. Any claim about what it does needs evidence of its own. And the ad must not suggest that a real customer got this result when none did.
GANs can also be used to make videos and audio. For example, a firm could use a GAN to make a video ad of a person using the product and seeing results within minutes. The GAN would produce the video, based on real data samples, and it could look convincing. But it would not show that a real person used the product or saw results that fast. Those claims need evidence of their own.
GANs can also make sound recordings. For example, a firm could use a GAN to make a clip of a person talking about how the product has helped them reduce wrinkles. The GAN would produce the audio, based on real data samples, and it could sound convincing. But a synthetic voice must not be passed off as a real customer telling their own story. A suitable audio model could even remake a clip with changes added. It could, say, make the clip sound as if it were recorded in a rain forest, not in a sound proofed studio booth.
If more firms take up GANs to make lifelike, believable ads, we are likely to see a surge in new and creative ads.
How could they be used to create more realistic and effective ads?
By training a GAN to make lifelike pictures of people or products, advertisers can try out ads that might catch the eye and work well. Whether they do needs to be tested against other creative.
GANs can also make ads feel more true to life, with new pictures close to what the target audience likes. For example, say an advertiser knows that a certain group of people likes pictures with blue skies and green trees. Then the advertiser can train a GAN to generate such images.
GANs can also be used to make more lifelike animations of people or products in ads. You can build an animation from a series of generated pictures that show a person or product in different poses. How lifelike it looks depends on the model and on the pictures used. So it is not by nature more lifelike than a normal video or picture.
There are many ways GANs could be used to make ads that feel more real and that work. As the tech moves on, we will likely see more lifelike ads made by GANs.
What ethical considerations are there with using Generative Adversarial Networks in advertising?
When it comes to the ethics of using GANs in marketing, there are a few key points to keep in mind. First and foremost, be aware that data can be twisted and misused when you use GANs. So marketers must take care to handle the data they gather for GANs ethically and responsibly.
It’s also not always clear who owns the data a GAN makes. The company that built the GAN could claim it. But so could the people whose photos were used to train it.
You also need to think about how marketing content made by GANs might affect consumers. GANs can make pictures, videos and other content that look real and are easy to believe. But there is always a chance that this content could mislead or misinform people. So marketers need to think hard about how they use GANs. They should only use them in ways that truly help consumers.
Finally, it’s worth thinking about what using GANs in marketing could mean in the long run. As GANs get better and more widely used, they could be used to make ever more lifelike and convincing fake news stories or other content. That could have a deep impact on society. So we need to think hard now about how we use GANs, to avoid harm later on.
What are the potential risks and benefits of using Generative Adversarial Networks in advertising?
GANs are a new type of AI that is being used more and more in marketing. Their use brings both possible risks and benefits. A few of each are listed below.
Some examples of the potential risks of using GANs in marketing include:
1. They could be used to make fake customer reviews or testimonials.
2. They could be used to make fake social media posts or comments.
3. They could be used to skew marketing research with false results, such as showing an effect that is not really there (e.g., by making fake survey answers).
4. They could be used to make targeted ads that are too good to be true (e.g., by making fake customer profiles).
5. They could be used to make fake news stories about a company or product.
Some examples of the potential benefits of using GANs in marketing include:
1. They could be used to make lifelike 3D images or videos of products (e.g., for product demos).
2. They could be used to make realistic simulations of how customers interact (e.g. for customer service training).
3. They could be used to make targeted ads that are more personal and relevant to each customer.
4. They could be used to help automate the making of marketing content, saving time on tasks that tend to take a lot of it (e.g. by making product descriptions or images in bulk).
Wrapping up
It’s exciting to think about all the ways GANs could be used in marketing and ads. The future here is largely unwritten. But some firms are early movers and have already begun to try GANs in their day-to-day work. Whether their campaigns work better still needs to be measured.
So before deciding whether GANs are right for its marketing, each company has to weigh the risks and benefits listed above.
I’m curious to see how their use grows over the coming months and years, and what they end up being used for. There are some risks in using GANs, but I’m hopeful that the tech will be used for good.
Skeptical about what generative AI can do? The first version of this article was generated with an AI tool. The exact model was not recorded here, so I cannot verify that it was a GAN.