A new study has demonstrated that aligning neural network connections just once before training can effectively guide how artificial intelligence (AI) learns.

A research team led by Professor Se-Bum Paik of the Department of Brain and Cognitive Sciences at KAIST has shown that the initial organization of neural connections plays an important role in subsequent learning. The team developed a new learning method that achieves high learning performance and stability without requiring repeated weight transport during training.

Most modern AI systems learn by calculating the difference between their predictions and the correct answers, then propagating these errors backward through the neural network. This process, known as "backpropagation," requires precise information about the network's forward connections to be used in backward calculations during every learning step.

In contrast, the biological brain has separate forward and feedback connections that operate independently. Inspired by this structure, researchers have developed learning methods that use separate, fixed connections to transmit error signals. However, these approaches have often suffered from slower learning and reduced accuracy.

The research team found a solution not in the learning process itself, but in how neural connections are organized before learning begins.

Their newly developed method, called "Initial Feedback Alignment (IFA)," aligns the forward connections that transmit information and the feedback connections that carry error signals only once before training begins. The feedback connections are then kept fixed throughout learning. Much like setting the right direction before starting a journey, this simple initial alignment allows AI to effectively reduce errors without repeatedly matching the two sets of connections during training.

< Initial Feedback Alignment improves learning performance and stabilizes loss optimization>

The researchers demonstrated that the new learning method works effectively not only in image classification tasks but also in fine-tuning pretrained large language models (LLMs). Across various experiments, IFA generally achieved higher learning performance and greater stability than conventional random feedback methods, while approaching the performance of standard backpropagation under multiple conditions. These findings demonstrate that learning through initial alignment can be extended to a wide range of AI models.

Further analysis of how AI minimizes errors revealed that initial alignment promotes more stable learning trajectories and convergence toward flatter minima, where small changes in network parameters do not cause sharp increases in error. Interestingly, the researchers also found that perfect symmetry between forward and feedback connections was not necessary for effective learning. Under certain conditions, IFA even showed greater robustness than conventional backpropagation against adversarial attacks designed to mislead AI systems.

The significance of this study lies in demonstrating that the initial organization of neural connections can effectively guide subsequent learning without continuously matching connection weights throughout training.

These findings provide new insights into how biological neural circuits may support learning and suggest potential applications in efficient AI learning systems, including brain-inspired neuromorphic computing and energy-efficient artificial intelligence.

The research was conducted by Jeonghwan Cheon (M.S. in the Department of Brain and Cognitive Sciences, currently in the ROK Army) as the first author and Jaehyuk Bae, a Ph.D. student, as the second author, with Professor Se-Bum Paik serving as the corresponding author, all from the Department of Brain and Cognitive Sciences at KAIST. The research was accepted at NeurIPS 2026, one of the world's leading international conferences in artificial intelligence.

Paper Information

※ Title: One-Time Soft Alignment Enables Resilient Learning without Weight Transport

※ Authors: Jeonghwan Cheon, Jaehyuk Bae, Se-Bum Paik

※ Conference: NeurIPS (Neural Information Processing Systems) 2026