Mądrzejsze podejście do debilizacji sztucznej inteligencji w opiece zdrowotnej

Dermatologists in our modern hospitals and clinics are finding themselves turning to artificial intelligence (AI) to sort through and classify various skin conditions. A key task for these AI systems is to evaluate whether or not a skin lesion could potentially develop into something more sinister like cancer, or if it’s just a benign growth. Trouble arises, however, if these systems learn to favor certain skin tones over others, consequently failing to accurately identify patients at higher risk.

The issue of bias in AI is as old as the technology itself, persisting as one of the major challenges that researchers face. Bias can originate from the training data used to teach the AI, but it can also be birthed from the architectural design of the model itself. When bias infiltrates real-world applications of AI, the impact can be significant — particularly in medicine where the stakes are sky high.

New Measures Against Bias in AI

A novel effort against AI bias has been introduced by a team of researchers from MIT, Worcester Polytechnic Institute and Google via their new approach known as “Weighted Rotational DebiasING” (WRING). Recognized at the 2026 International Conference for Learning Representations, this method specifically targets vision language models (VLMs), such as those used by the AI company OpenAI in their OpenCLIP application. VLMs are flexible tech tools that can process and understand various types of data such as video, images, and text all at once.

The researchers’ main motivation to develop WRING was a lingering issue when using existing debiasing methods like “projection debiasing.” Coined as the “Whac-A-Mole dilemma,” this challenge, which was initially pointed out in 2023, surfaces when attempting to fight bias in AI. The projection debiasing technique consists of stripping out biased information from the model’s representation space. This sounds like a solid plan initially, but it has a hidden downside: it can unknowingly modify other relationships within the AI model.

A Smarter Way to Debias AI Vision Models

WRING ingeniously deals with these problems by modifying particular coordinates in the AI model’s high-dimensional framework. These adjustments essentially deny the model’s ability to differentiate between varying groups within a concept, while leaving all other relationships unscathed. As a post-processing approach similar to projection debiasing, WRING can be applied to already-trained VLMs, eliminating the need for new training.

One of the research paper’s authors, Walter Gerych, emphasized the efficacy and efficiency of WRING. But, he noted its current limitation: it can only be applied to Contrastive Language-Image Pre-training (CLIP) models. Looking ahead, Gerych envisions bringing the WRING approach to generative language models, such as ChatGPT. As for funding, the study was financially backed by awards from the National Science Foundation and an MIT-Google Computing Innovation Award. For a deep dive into the topic, check the full paper tutaj.

If you’re interested in incorporating AI technologies into your business, especially to automate routine tasks, consider looking into top-notch AI solutions at implementi.ai.