Best Paper Award at the Korea Computer Congress 2025 (KCC 2025) - High-Performance AI Systems Laboratory

At the Korea Computer Congress 2025 (KCC 2025), held over three days from 2 to 4 July, the paper 'Automatic Search for Tiny Neural Networks Supporting Early Exit', presented by master's student Lee Seung-woo (supervisor: Lee Young-min), received the Best Paper Award.

 

The paper is the result of research written jointly by Lee Seung-woo (master's) and Han Seong-hwan (undergraduate) under the supervision of Professor Lee Young-min.

▲ From left: Lee Seung-woo (master's), Han Seong-hwan (undergraduate)

Existing early-exit neural networks suffer limits to computational efficiency and scalability in MCU environments, where SRAM and flash resources are extremely constrained, because their auxiliary classifiers are structurally complex and require a great deal of memory. To overcome these constraints, this research designs Slimmed T-RECX, a lightweight auxiliary classifier structure optimized for MCU environments, and builds a supernet combining position-based weighted self-distillation training, proposing a method that automatically searches for the optimal tiny early-exit neural network architecture.

 

The method satisfied the hardware constraints (SRAM 880KB, flash 1.8MB) while achieving Top-1 accuracy of 77.57% on CIFAR-100 and an early-exit rate of 59.7%, holding the accuracy loss to a minimal 0.44 percentage points while reducing computation (FLOPs) by 39.1%. Using a genetic algorithm (GA) based search to find the optimal positions and thresholds for auxiliary classifiers quickly and automatically across a wide design space, the work showed that the number of parameters per architecture can be minimized while expressive power is maintained, demonstrating practical effectiveness in tiny environments.

 

The Korea Computer Congress 2025 (KCC 2025) is an information science and technology conference at which the latest research results across all fields of information technology — artificial intelligence, computer systems, information and communications, software engineering and more — are presented.

High-Performance AI Systems Laboratory: https://aisys.sogang.ac.kr https://aisys.sogang.ac.kr/