Researchers Develop AI Model to Improve Image Recognition Reliability

A research team led by Gergő Orbán at the HUN-REN Wigner Physics Research Center, working with Argentine researchers, has developed a new method to make image-processing artificial intelligence more reliable by improving how it estimates uncertainty in its decisions.

For deep learning systems, it is crucial not only to make the right decision but also to assess confidence in that decision. The method developed by Hungarian and Argentine experts improves this uncertainty management, which could lead to significant advances in medical imaging and also bring us closer to a more precise understanding of how human vision works.

The research is related to the field of NeuroAI, which examines parallels between the functioning of the nervous system and that of artificial intelligence (AI). The goal is twofold: on the one hand, a better understanding of brain function can help in developing more effective machine intelligence, while the study of artificial systems can provide new insights for research into the nervous system.

The researchers sought to determine how artificial intelligence could more accurately indicate the degree of certainty it has in its inferences about an image.

The study focused on variational autoencoders (VAE). These are deep generative models that capture the essential features of an image in a compact, internal representation and are then able to reconstruct the image from it. They can not only draw conclusions but also estimate the uncertainty associated with them.

However, the researchers found that the uncertainty estimation in traditional VAEs can fail in several common image processing scenarios—such as when image contrast changes, when dealing with noisy or corrupted data, or when images differ significantly from the training examples.

To address this issue, the research team supplemented the model with a solution known from classical computer vision and created a new variant called the Explaining-Away Variational Autoencoder (EA-VAE). With the help of this new built-in factor, the model is able to distinguish whether a change visible in the image stems from the depicted object itself or is merely due to a change in the overall brightness or contrast of the image.

Tests conducted on various types of image data confirmed that the new model provides more consistent uncertainty estimates.

The press release highlights that what makes this research unique is that the so-called divisive normalization—a phenomenon widely observed in the biological visual system—has been observed in the functioning of the EA-VAE. This computational mechanism regulates the activity of neurons relative to the activity of surrounding cells.

Thus, the researchers not only improved the AI’s performance but also developed a model based on a mechanism found in the biological nervous system.

Accurately assessing uncertainty is of paramount importance in applications where a falsely confident decision could have serious consequences. When analyzing a medical image, for example, it matters not only what lesion the algorithm identifies but also how confident it is in that identification. A well-calibrated system can indicate when a given image is unusual or difficult to interpret and therefore requires review by a human expert.

The method can also be beneficial in any image-processing system that must cope with varying lighting conditions, noise, corrupted data, or previously unseen images. The research thus points the way toward both more reliable artificial intelligence and a more accurate modeling of biological vision, according to a summary by the HUN-REN Wigner Research Center for Physics.

Via MTI; Featured photo: Pixabay

The post Researchers Develop AI Model to Improve Image Recognition Reliability appeared first on Hungary Today.