USC Researchers Propose a Novel Learning Framework, ‘Group-Supervised Learning (GSL)’, To Aid Neural Networks Envision Objects With Different Attributes

521
Source: https://openreview.net/pdf?id=8wqCDnBmnrT

Artificial intelligence has proven itself to be adept at many tasks, from inventing human faces that don’t exist to winning games of poker. However, these networks still struggle when it comes to something humans do naturally: imagination.

Recently, a research team from USC led by computer science Professor Laurent Itti has developed an AI model that uses human-like capabilities to imagine a never-before-seen object with different attributes. The team took their inspiration from human visual generalization capabilities. They tried to simulate human imagination in machines.

Humans can differentiate their learned knowledge by attributes, for example, shape, pose, position, color, etc., and then recombine them to visualize a new object. The research team tries to simulate this process using neural networks. This novel approach unleashes a new sense of imagination in AI systems.

One of the long-sought goals of AI has been creating models that can extrapolate, i.e., given a few examples, and the model should extract the underlying rules and apply them to a wide range of new examples it hasn’t seen before. Machines are usually trained on sample features, say pixels, without considering the object’s attributes.

https://openreview.net/pdf?id=8wqCDnBmnrT

In their new study, the research team tries to overcome this limitation using disentanglement. Disentanglement is a process that can be utilized to generate deepfakes, for example, by disentangling human face movements and identity.

The new approach uses a group of sample images instead of one sample at a time and uses the similarity between them to achieve “controllable disentangled representation learning.”

Next, it recombines this knowledge to achieve controllable novel image synthesis, which can be perceived as imagination. The process is similar to how humans extrapolate. For example, when a human sees a color from one object, they can effectively apply it to other objects by substituting the original color with the new one. Using this method, the group generated a new dataset containing 1.56 million images that could aid future research in the field.

The researchers believe that their framework can be compatible with nearly any type of data. This significantly widens the opportunity for applications. For example, it can be used in disentangling race and gender-related knowledge to make fairer AI by removing sensitive attributes from the equation.

Similarly, in the field of medicine, it could aid doctors, and biologists in discovering more useful drugs by disentangling the medicine function from other properties and then recombining them to generate new medicine. Permeating machines with imagination could also help create safer AI.

Source: https://viterbischool.usc.edu/news/2021/07/usc-researchers-enable-ai-to-use-its-imagination/

Paper: https://openreview.net/pdf?id=8wqCDnBmnrT


A message from Asif Razzaq, Co-founder of Marktechpost:

Show your support for our mission ‘making AI understandable for all’ by joining/connecting through our 34k+ FB GroupLinkedIn Page and Quora AI Group.

Advertisement/Sponsored Post:

If you are a company looking to promote your product/webinar/conference/service, feel free to reach out via email to [email protected] We offer sponsored posts and advertisements.