A team from DGIST has developed an image translation model that can reduce data biases in AI models. The model uses spatial self-similarity loss and texture co-occurrence to generate high-quality images with consistent content and similar textures. It outperforms existing techniques in debiasing and image translation, making it useful for applications like autonomous vehicles and healthcare.
Data Bias Removal: A Breakthrough in AI
Data biases can significantly impact the performance of AI models when applied to new data. To address this challenge, a team from Daegu Gyeongbuk Institute of Science and Technology (DGIST) has developed an innovative image translation model that effectively eliminates data biases without prior knowledge of their existence.
Practical Solutions and Value
By using spatial self-similarity loss, texture co-occurrence, and GAN losses, the team has created a model that generates high-quality images with consistent content and similar local and global textures. These images can then be used to train debiased classifiers or modified segmentation models.
The key contributions of this research are:
- Texture co-occurrence and spatial self-similarity losses can be used to translate images and obtain optimal pictures for debiasing and domain adaptation.
- A strategy for learning downstream tasks that mitigates unexpected biases during training by enriching the training dataset without bias labels.
- Superior performance compared to existing debiasing and image translation techniques, demonstrated through comparisons with biased datasets and domain adaptation datasets.
This deep learning model outperforms preexisting algorithms by creating a dataset through texture debiasing and using it for training. It has shown superior performance on datasets with texture biases, such as classification datasets for distinguishing numbers, dogs and cats with different hair colors, and distinguishing COVID-19 from bacterial pneumonia.
If you’re interested in leveraging AI to remove biases in your datasets and stay competitive, consider exploring this breakthrough research. Connect with us at hello@itinai.com for AI KPI management advice and visit itinai.com/aisalesbot to discover our AI Sales Bot solution for automating customer engagement and managing interactions across all customer journey stages.
