Tuesday, April 16, 2024

Fractals can help AI learn to see more clearly – or at least fairly

Must read


The problem is that manually assembling a dataset like ImageNet takes a lot of time and effort. Images are usually tagged with poorly paid crowd workers. Data sets can also contain sexist or racist labels that may skew a model in a hidden way as well as images of people who have been included without their consent. There is evidence that these biases can seep into even in pre-training.

However, fractal patterns can be found in everything from trees and flowers to clouds and waves. This prompted the team from Japan’s National Institute of Advanced Industrial Science and Technology (AIST), Tokyo Institute of Technology, and Tokyo Denki University to question whether these models could be used to teach an image recognition system the basics of image recognition, instead of using photos of real objects.

Researchers created FractalDB, an endless number of computer-generated fractals. Some look like leaves, others like snowflakes or snail shells. Each group of similar models automatically received a label. They then used FractalDB to pre-train a convolutional neural network, a type of deep learning model commonly used in image recognition systems, before completing its training with a set of real images. They found that it performed almost as well as models trained on cutting edge data sets, including ImageNet and Places, which contains 2.5 million images of outdoor scenes.

Does it work? Anh Nguyen of Auburn University in Alabama, who was not involved in the study, is not convinced that FractalDB is a match for ImageNet. He studied how abstract models can confuse image recognition systems. “There is a link between this work and examples that deceive machines,” he says. He would like to explore in more detail how this new approach works. But Japanese researchers believe that by fine-tuning their approach, computer-generated datasets like FractalDB could replace existing ones.

Why fractals: The researchers also tried to train their AI using other abstract images, including those produced using Perlin noise, which creates speckled patterns, and Bézier curves, a type of curve used in computer graphics. But fractals have given the best results. “Fractal geometry exists in basic knowledge of the world,” says senior author Hirokatsu Kataoka from AIST.


- Advertisement -spot_img

More articles


Please enter your comment!
Please enter your name here

- Advertisement -spot_img

Latest article