Study creates AI model for effectively reducing biases in dataset

WASHINGTON: Researchers had developed a new image translation model that could effectively reduce biases in data.
Data biases may occur during the development of an artificial intelligence (AI) model using photos acquired from diverse sources, contrary to the user’s goal, due to a variety of circumstances. Despite the lack of knowledge on such aspects, the created model may reduce data biases, resulting in a good image-analysis performance.
This approach is expected to help with breakthroughs in self-driving cars, content development, and medicine.
Deep learning algorithms are biased by the datasets used to train them. Because of the possibility of Covid-19 infection, image-collecting conditions may alter while generating a dataset to distinguish bacterial pneumonia from coronavirus illness 2019.
As a result, these differences cause modest disparities in the images, causing existing deep-learning models to identify diseases based on image technique inconsistencies rather than the critical traits for practical disease detection. In this case, depending on the data utilised in the training process, these models perform well.
To solve these issues, Prof. Park’s research group created an image translation model that could construct a dataset using texture debiasing and then run the learning process on that dataset. Existing picture translation models are frequently hampered by the issue of texture modifications causing unintentional content changes, as textures and contents are inextricably linked.
To overcome this issue, Prof Park’s research team created a novel model that uses error functions for both textures and contents at the same time.
This research team’s novel image translation model works by extracting and combining information about the contents of an input image and textures from a separate domain.
The created model is trained to utilise both error functions for spatial self-similarity and texture co-occurrence to store information on not only the contents of input photos but also the texture of the new domain. The model may generate an image with the texture of a different domain while retaining knowledge about the contents of the input image using these procedures.
The new deep learning model outperforms the previous methods because it generates a dataset using texture debiasing and uses the obtained dataset for training. When tested on datasets with texture biases, such as a classification dataset for distinguishing numbers, a classification dataset for distinguishing dogs and cats with different hair colours, and a classification dataset using different image protocols to distinguish Covid-19 from bacterial pneumonia, it outperformed existing debiasing and image translation techniques.
Furthermore, when applied to datasets with different biases, such as a classification dataset for identifying multi-label integers and one for distinguishing pictures, drawings, animations, and sketches, it outperformed previous algorithms.
Furthermore, Prof. Park’s research team’s picture translation technique can be used in image alteration. The research team discovered that the suggested technology merely altered an image’s textures while maintaining its original content. This analytic result confirmed the superior performance of the developed method compared to existing image manipulation methods.
Furthermore, this solution can be employed effectively in various situations. The researchers evaluated the created method’s effectiveness to that of existing picture translation systems based on other domains, such as medical and self-driving images. According to the analytical data, the developed approach performed better than the current methods.
Prof. Park stated, “The technology developed in this research offers a significant performance boost in situations where biased datasets are inevitably used to train deep learning models in industrial and medical fields.”
He also added, “It is expected that this solution will make a substantial contribution to enhancing the robustness of AI models commercially used or distributed in diverse environments for commercial purposes.”