Researchers allow AI to use its’ imagination ‘- closer to humans’ understanding of the world
A team of USC researchers are helping AI imagine what is invisible. This approach could also lead to fairer AI, new drugs, and improved safety for self-driving cars.
Imagine an orange cat. Now imagine the same cat. However, he has black fur. Now imagine a cat pretending to be along the Great Wall of China. When you do this, the activation of a series of neurons in the brain is reminiscent of variations in the images presented, based on previous world knowledge.
In other words, as a human it is easy to imagine objects with different attributes. However, despite advancements in deep neural networks that rival or exceed human performance on certain tasks, computers still struggle with the very human skill of “the imagination.”
Currently, a USC research team composed of Professor Laurent Itti in computer science and doctoral students Yunhao Ge, Sami Abu-El-Haija and Gan Xin have used human characteristics to date. We have developed an AI that does not imagine different objects. attribute.Title processed Zero-shot synthesis by learning with supervised group learningWas published at the 2021 International Conference on Representing Learning, held on May 7.
“We were inspired by human visual generalization, which attempts to simulate the human imagination with machines,” said Ge, the study’s lead author.
“Man can separate the knowledge acquired by attributes (shape, pose, position, color, etc.) and combine them to imagine new objects. Our article uses neural networks. I am trying to simulate this process. “
AI generalization problem
For example, suppose you want to create an AI system that produces an image of a car. Ideally, the algorithm would provide images of the car, allowing it to generate different types of cars of any color from different angles, from Porsche to Pontiac to pickup trucks.
It’s one of AI’s long-standing goals to create scale-up models. This means that, given a few examples, the model should be able to extract the underlying rules and apply them to a wide range of new examples that we’ve never seen before. However, machines are most often trained on example features, such as pixels, without considering the attributes of the object.
In this new study, researchers attempt to overcome this limitation by using a concept called unraveling. Unraveling can be used to generate deepfake, for example, by disentangling the movements and identities of the human face. By doing this, “people can synthesize new images and videos that replace the identity of the original person with another person, but retain the original movement,” Ge said.
Likewise, the new approach takes a group of sample images instead of one sample at a time as in traditional algorithms, explores the similarities between them and “learning tangled expressions in a controllable way”. To achieve what is called.
We then recombine this knowledge to achieve what we call a “new controllable image composition”, or imagination. “For example, in the movie Transformers, you can take the shape of a Megatron car, the color and pose of a yellow Bumblebee car, and the background of Times Square in New York City. As a result, this sample during a training session. Even if it was not seen, a Bumblebee colored Megatron car would be driven in Times Square. “
This is similar to how humans extrapolate. When humans see the color of an object, they can easily apply it to other objects by replacing the original color with a new color. Using their technology, the group generated a new dataset containing 1.56 million images that could be useful for future research in this area.
Understanding the world
Untangling isn’t a new idea, but researchers say their framework is compatible with almost all types of data and knowledge. This opens up new application opportunities. For example, create fairer AI by disentangling knowledge related to race and gender and removing sensitive attributes from equations altogether.
In medicine, disentangling the functions of medicine from other properties and recombining them to synthesize new drugs can help physicians and biologists discover more useful drugs. You can also create safer AI by infusing your machine with imagination, for example, allowing self-driving cars to imagine and avoid dangerous scenarios never seen before during training.
“Deep learning has already demonstrated exceptional performance and expectations in many areas, but it is often done by superficial imitation and does not have a deep understanding of the individual attributes that make each object unique. Said Itti. “For the first time, this new approach to disentangling unleashes the new imagination of AI systems and brings it closer to understanding the human world.”
Reference: May 7, 2021, “Zero Shot Synthesis by Group Supervised Learning” by Yunhao Ge, Sami Abu-El-Haija, Gan Xin, Laurent Itti 2021 International Conference on the Representation of Learning..
Researchers allow AI to use its’ imagination ‘- closer to humans’ understanding of the world Source link Researchers allow AI to use its’ imagination ‘- closer to humans’ understanding of the world