Approaching Human Understanding of the World: Researchers Enable AI to Have Imagination

2024-03-12

According to foreign media reports, first imagine an orange cat, now imagine this cat with black fur, and then imagine a cat strutting high on the Great Wall of China. This series of thoughts will stimulate the rapid activation of neurons in your brain, and come up with various pictures based on your previous understanding of the world. In other words, as humans, it is easy to imagine an object with different attributes.

However, despite deep neural networks achieving performance comparable to or exceeding that of humans in certain tasks, computers are still struggling with their unique "imagination" skills.

Now, a research team at the University of Southern California (USC), consisting of computer science professor Laurent Itti, doctoral students Yunhao Ge, Sami Abu El Haija, and Gan Xin, has developed an artificial intelligence that can use human like abilities to imagine an unseen object with different properties.

"We were inspired by human visual generalization ability and attempted to simulate human imagination on machines," said Ge, the lead author of the study. "Humans can separate what they have learned through attributes such as shape, posture, position, and color, and then recombine them to imagine a new object. Our paper attempts to simulate this process using neural networks."

The generalization problem of AI

Suppose you want to create an AI system that generates car images. Ideally, you could provide the algorithm with some images of cars, so that it can generate cars of various colors from multiple angles.

Thisone of the long-term goals pursued by AI: to create models that can be extrapolated. This means that with just a few examples, the model should be able to extract basic rules and apply them to a large number of new examples it has not seen before. But machines are usually trained on sample features, such as pixels, without considering the properties of objects.

The Science of Imagination

In this new study, researchers attempted to overcome this limitation using a concept called disentanglement. Decontanglement can be used to generate deep forgery. Ge pointed out that by doing so, "people can synthesize new images and videos to replace the original person with another person's identity, but still maintain the original movement."

Similarly, the new method adopts a set of sample images - instead of collecting one sample at a time like traditional algorithms - and mines their similarities to achieve what is known as "controllable disentanglement representation learning.".

Then, it recombines this knowledge to achieve "controllable new image synthesis" or something that can be called imagination. " Thissimilar to our human inference: when a person sees a color from an object, we can easily apply it to other objects by replacing the original color with a new one. The research team generated a new dataset containing 1.56 million images using their technology, which may contribute to future research in this field.

Understanding this world

Although "de entanglement" is not a new concept, researchers suggest that their framework can be compatible with almost any type of data or knowledge. This expands the opportunities for application. Removing sensitive attributes of race and gender related knowledge from the equation to create a more equitable AI.

In the field of medicine, it can help doctors and biologists discover more useful drugs to separate their functions from other properties, and then recombine them to synthesize new drugs. Giving machine imagination can also help to create safer AI, such as allowing autonomous vehicle to imagine and avoid dangerous scenes that cannot be seen in training.

Itti said, "Deep learning has shown excellent performance and prospects in many fields, but this often occurs through shallow imitation, without a deeper understanding of the individual properties that make each object unique. This new separation method truly unleashes the new imagination of AI systems for the first time, bringing them closer to human understanding of the world."

Architects used 3D printing technology to build a concrete bridge in Venice