Can AI Dream?

The capabilities of artificial intelligence (AI) have been growing fast, and intelligent computer systems can now do things that were previously the preserve of human beings, such as learning. If AI can see, hear, and hold meaningful, intelligent conversations (think ChatGPT), what else can it do? Can AI dream?

AI can dream. It has been shown to generate concepts and visuals that didn’t exist before. Moreover, such dreams are useful to AI systems, just like to humans. However, AI dreams are usually controlled and can’t occur without being programmed.

In this article, I’ll explain how AI dreams occur and why they’re useful. I’ll also compare human and AI dreams.

Just Like Human Dreams, AI Dreams Involve Distorted Memories

When discussing dreams and whether AI can have them, our point of reference is the human dream. We dream when parts of our brains become active during sleep, resulting in a series of unconscious sensations and thoughts.

In addition to dreams, one of the most impressive capabilities of our brains is that they can retain and replay memories. You can relive an experience you went through a dozen years ago because your brain remembers it.

Memories make up part of our dreams. But to come up with dreams, the brain usually distorts memories and sometimes uses them as a base to develop unique experiences.

In a way, human dreams can be thought of as confused memories. But so can AI dreams.

How AI Dreams Occur

Scientists stumbled into AI dreams while working to improve the performance of the neural networks that enable processes like image classification.

To train an AI for facial recognition, scientists feed it numerous images. The AI is programmed to analyze the images and come up with the concept of a face. It then compares inputs to this concept to determine whether they’re faces.

A major breakthrough came when scientists learned of inaccuracies in some of the concepts the classification systems were developing. As an example, an AI had developed a concept of a dumbbell that included an arm. So, if an image of a dumbbell on the floor was shown to the AI, it wouldn’t count as a dumbbell because it didn’t have an arm.

This “concept” of images that AI systems develop during learning is similar to a human memory. What is more, AI can be induced to distort these images into variations so unique that they’re likened to human hallucinations, and which are the AI version of dreams.

Dreaming in AI involves essentially letting the AI hallucination run free in a process called inceptionism. Google’s Inception program has used this technique to produce a whole gallery of AI dreams in the form of images and videos.

Inception is a neural network built for image recognition. To identify images, it uses a complex neural network with multiple layers. Each layer characterizes a certain aspect of an image. For example, one layer can be responsible for identifying doors. This information will later be collated with that from other layers to identify an image as a building.

Inceptionism is the amplification of the work of one layer in the neural network, achieved by maximizing the activation of the nodes on that layer. The result is surreal images — dreams.

How Computer Scientists Get AI to Dream

AI dreams are intentional mistakes. The scientist who induces AI to dream isn’t sure of the product, but they allow the AI to make mistakes.

Since scientists control these dreams, they can regulate them, essentially limiting the output so that it’s useful in some way.

There are two major approaches to getting computers to dream.

Generative Adversarial Networks (GANs)

A generative adversarial network is an AI setup that consists of two neural networks:

A generator
A discriminator

The generator is a neural network whose goal is to generate images that the discriminator classifies as either false or true.

The goal of the training is to get the generator to produce fake images that are so good that the discriminator can’t tell them apart from real images. During training, the discriminator will give feedback to help the generator learn in a process called backpropagation.

By the time the training is complete, the system can produce fakes (dreams) that are virtually unrecognizable from real objects.

Reinforcement Learning

This is a type of AI training technique that gets AI to figure out solutions to problems with the help of a reward system. The setup is such that the AI associates solutions with rewards and is programmed to maximize the rewards that it gets.

Usually, while the AI interacts with the environment, it learns and alters its behavior to increase the probability of rewards.

Getting the AI to dream involves programming it to simulate future states of the environment and try to maximize the rewards in those “imaginary states.”

The AI could learn something through an interaction in a dreamed-up environment and use it when it encounters a similar situation in the real environment. It would have improved its performance without undergoing the cost of learning — usually failure.

Why Would Scientists Train AI to Dream?

Most people view dreaming as a highly-advanced activity that’s a preserve of humans. From that point of view, we shouldn’t teach AI how to dream.

Moreover, AI dreaming essentially involves allowing AI to make mistakes. It’s easy to think of a scenario where these mistakes come in the form of unpleasant surprises — like an AI getting the idea to wipe out half the world’s population to solve the problem of sustainability.

Why would we risk such outcomes just to train AI to dream?

Dreams are helpful. For example, they help humans process a day’s experiences and integrate them into long-term memory. They’re essential to our advanced intelligence. So it’s not a stretch to assume that they would help AI perform better in solving problems like cancer.

And there’s evidence to show that dreams do enhance the effectiveness of AI.

Dreaming Improves the Efficiency of AI Models

DeepMind is a subsidiary of Google that was formed with the goal of creating AI that can reason like humans. In 2016, they published results showing that they had significantly enhanced the data efficiency, learning speed, and robustness of an AI by empowering it to dream.

The AI was using reinforcement learning to learn games like Labyrinth. With dreams, it learned ten times faster than previous versions of AI that didn’t dream.

While they were using the concept of dreams used to improve performance in games, it could easily be applied to other fields where scientists are trying to make an impact with AI, including energy management and healthcare.

Dreaming Helps Create Deepfakes

Some people don’t think deepfakes are a good thing, and with good reason, but they can be useful. For example, using deepfakes in content generation can drastically reduce cost. You get to use humans that don’t exist anywhere in the planet as your actors. They look so real that even humans can’t tell they’re fake — yet they’re computer generated.

The magic of deepfakes is made possibly by the dreams produced by GANs.

To produce fake faces, you train a GAN to generate human faces that it can’t recognize are fake —- and then you’re in business. Such a GAN will be trained to do facial recognition and will have impeccable knowledge of what makes up a human face. It’ll then use this information to generate infallible results.

These faces are dreams because they’re total fabrications.

If AI can be made to dream, can they also develop emotions like humans? Find out all about it in our article “Can AI Develop Emotions?” and explore how AI can mimic human type emotions.

The Takeaway

Can AI dream? Yes, but not the way humans do. Dreams in humans are uncontrolled, whereas in AI, they’re controlled. AI dreams may not be as glorious as humans, but AI can still fabricate images and concepts, which is essentially what dreams are.

Not only can computer scientists induce AI to dream, but they can also make use of those dreams, for example in deepfakes and to enhance the learning rate of AI. Read our other article Can AI Create Another AI? to learn more about the learning rate of AI and how it can create other AI’s in order to achieve certain goals.