Scientists Aim to Provide Human Inspired Visual Generalization CapabilitiesAccording to Yunhao Ge, a computer scientist from the University of Southern California or USC notes that they were inspired by human visual generalization capabilities in order to simulate machines to be capable of human imagination. It was noted that human beings are able to separate learned knowledge by attributes and combine them in order to imagine a new object. The paper reportedly attempts to simulate the whole process through neural networks. The main key is extrapolation. This means being able to use a huge bank of training data and pass what is seen into the unseen. This is supposedly difficult for AI due to it being trained to spot certain patterns instead of that of broader attributes.
New Approach with Controllable Disentangled RepresentationThe team has been able to come up with a controllable disentangled representation learning. This is done by using an approach that is close to creating deep fakes by generally disentangling different parts of a sample. This is like separating face movement and face identity as seen on deepfake videos. According to ScienceAlert, this means if an AI is capable of seeing a red car and a blue car, it can also “imagine” what a red bike would look like despite not yet seeing one. The researchers put together a framework that is called Group Supervised Learning.
AI to Recognize Similarities and DifferencesOne of the main innovations within the technique is processing samples within groups as opposed to individually. This would then build up semantic links between them along the way. Stephen Hawkings has previously warned that AI could replace human beings.
Artificial intelligence can then recognize similarities and differences in the samples that it sees by using the knowledge in order to produce something that is completely new. According to Laurent Itti, a USC computer scientist, the new disentanglement approach will truly unleash a new sense of imagination within AI systems. This would bring them closer to a human’s way of understanding the world.
Move to Reduce AI Bias
The ideas are notably not completely new as seen on OpenReview. However, researchers have taken the concepts a step further by making the approach more flexible and also compatible with additional data types. They’ve also made the whole framework open source for other scientists to better make use of it as AI is already widely used in day-to-day activities.
In the future, the system developed could help guard directly against AI bias by taking away more sensitive attributes from total equations which helps make neural networks that aren’t sexist or racist, for example. The same approach could potentially be applied to the medicine fields and self-driving cars.
The research was presented at the official 2021 International Conference on Learning Representations. The research can be seen here.
Written by Urian B.