A smarter artificial intelligence (AI) technology, more reminiscent of the human brain, has emerged.
The Institute for Basic Science (IBS), led by Director Noh Do-young, announced on the 22nd that a team under the leadership of Chang-Jun Lee, head of the Cognitive and Sociality Research Group, along with Professor Kyung-Woo Song’s team from the Department of Applied Statistics at Yonsei University, developed a new technique that enhances AI image recognition capabilities by applying the method through which the brain’s visual cortex selectively processes visual information.
The human visual system can recognize objects at a glance and quickly prioritize important information even in complex environments.
In contrast, existing AI models like convolutional neural networks (CNNs) have limitations in understanding wide context or the relationships among distant pieces of information due to their structure of analyzing images by dividing them into small square filters.
Vision transformers, which address these limitations, require massive computations and large datasets, lessening their practicality.
The research team focused on how the human brain’s visual cortex processes visual information, responding only to features or important parts.
By implementing this approach, the team proposed the ‘Lp-convolution’ technique, significantly enhancing the performance of CNN models. It is designed to enable AI to prioritize key information when analyzing images.
The ‘mask’, an automatically generated filter for each image, highlights important parts like neurons in the visual cortex and excludes less important ones. The mask adjusts its form during the learning process and maintains a focus on crucial features across various environments.
Kwon Jae, the first author and a postdoctoral researcher at IBS (now at the Max Planck Institute for Security and Privacy in Germany), explained that “Lp-convolution, inspired by the brain’s information processing methods, aids AI in utilizing computational resources efficiently and enables more precise analysis.”
The research team further applied and evaluated this technology on various CNN models. As a result, the accuracy of image classification significantly improved compared to existing CNN models. Performance did not degrade even when filter size increased; instead, accuracy improved. Typically, broadening the analysis range reduces accuracy, but this limitation was overcome.
Moreover, the research team assessed how similarly Lp-convolution mimics the actual brain’s information processing. By showing natural images to mice and recording visual cortex neuron activities, they trained AI models to predict neural responses to each image. The model with Lp-convolution predicted neuron responses more accurately than traditional CNN models and reduced errors.
Chang-Jun Lee, the head of the research team, stated that “Lp-convolution can substantially contribute to mimicking and understanding how the brain processes information,” and “it sets a good example of a novel convergence model where AI and neuroscience can advance together.”
This research will be presented at ‘ICLR 2025,’ held in Singapore from April 24th to 28th.