A domestic research team has succeeded in applying the way the human brain's visual cortex selectively processes visual information to artificial intelligence (AI) image recognition technology.
On the 22nd, Director General Lee Chang-jun of the Institute for Basic Science (IBS) Cognitive and Social Research Center noted that his research team, in collaboration with Professor Song Kyung-woo from Yonsei University's Department of Applied Statistics, developed a technology to enhance AI's image recognition capabilities by applying the methodology used by the brain's visual cortex to process visual information.
The human visual system has remarkable recognition capabilities. It can recognize objects at a glance and quickly filter out important information even in complex environments. In contrast, existing AI models still show limitations.
To improve existing AI systems, the research team focused on the way the human brain's visual cortex selectively processes visual information. The human visual cortex does not treat all information equally but reacts selectively, focusing on noticeable features or important parts. In this process, neurons detect a wide range smoothly and selectively respond only to the necessary information. The research team proposed the 'Lp-convolution' technology to significantly enhance the performance of convolutional neural network (CNN) models.
Lp-convolution is a technology designed to allow AI to prioritize core information when analyzing images, similar to how humans do. It emphasizes important parts like the neurons in the visual cortex while naturally excluding less important parts using a 'mask.' This mask adjusts its shape during the learning process, allowing the AI to consistently focus on significant features across various environments.
The research team evaluated the performance of this technology by applying it to various CNN models. As a result, models that utilized Lp-convolution showed a noticeable improvement in image classification accuracy compared to existing CNN models. Notably, even when the filter size was significantly increased to examine a wider area at once, the models operated stably without any performance degradation, resulting in enhanced accuracy. Generally, widening the analysis range increases computational load and tends to lower accuracy; however, Lp-convolution effectively overcame these limitations.
Additionally, the research team conducted experiments to determine how similarly Lp-convolution mimics the information processing method of the actual brain. They recorded the activity of visual cortex neurons while showing various natural images to mice and trained the AI model to predict how neurons would respond to each image based on this data. Compared to the actual neuron responses, the model trained with Lp-convolution was able to predict neuron reactions more precisely than existing CNN models, and the prediction errors were also reduced.
Lee Chang-jun, Director General of IBS, remarked, "Lp-convolution can greatly contribute not only to enhancing AI performance but also to mirroring and understanding how the brain processes information," adding that "it will serve as a good example of a new integrated model where AI and neuroscience can advance together."
This research has been accepted at the renowned AI conference 'International Conference on Learning Representations (ICLR).' It is scheduled to be presented at 'ICLR 2025,' which will take place in Singapore from April 24 to 28.