#South Korea #Seoul #AI Training #SEOULTECH #VFF-Net

VFF-Net: The Future of AI Training

Seoul National University of Science and Technology has recently unveiled a groundbreaking algorithm, VFF-Net, which introduces a novel approach to training artificial intelligence (AI) models. This innovative technique presents a significant departure from the conventional backpropagation (BP) methods that have long dominated this realm.

Understanding Backpropagation's Limitations

Backpropagation has been the cornerstone of deep neural networks (DNNs) – complex systems that enable AI to learn from data inputs such as images, audio, and text. While BP has facilitated remarkable predictions across various sectors, it is not without flaws. Some of the primary issues include slow convergence rates, tendencies towards overfitting, high computational demands, and limited transparency, often referred to as the 'black box' issue.

Although forward-forward networks (FFNs) were developed as an alternative to BP, their application within convolutional neural networks (CNNs), particularly popular in image processing, proved challenging. This was primarily due to the inherent risk of information loss when directly applying FFNs.

Innovations and Solutions in VFF-Net

A research team led by Mr. Gilha Lee and Associate Professor Hyun Kim tackled this issue by creating VFF-Net, a sophisticated algorithm designed to work efficiently with CNNs. This approach is centered around three pioneering methodologies:

1. Label-Wise Noise Labeling (LWNL): This method trains the network using three distinct types of data - clean images, positively labeled images, and negatively labeled images. This strategy is crucial in preserving pixel information, thereby boosting accuracy during training.

2. Cosine Similarity-Based Contrastive Loss (CSCL): In a departure from conventional training algorithms, CSCL assesses the similarity of feature representations based on cosine angles. This unique perspective preserves vital spatial information, enhancing the overall effectiveness of image classification.

3. Layer Grouping (LG): VFF-Net overcomes the challenges associated with layer-specific training by strategically grouping layers with similar output characteristics, along with auxiliary layers that bolster performance gains.

Thanks to these methodological advancements, VFF-Net has demonstrated significant enhancements in image classification accuracy. For instance, when tested on CNN models with four convolutional layers, error rates dropped by 8.31% on the CIFAR-10 dataset and by 3.80% on CIFAR-100. Moreover, VFF-Net achieved a remarkable error rate of just 1.70% when evaluating the MNIST dataset, showcasing its potential to redefine performance standards.

The Future of AI Training

Dr. Hyun Kim emphasizes the exciting potential of VFF-Net: "By moving away from backpropagation, we are ushering in a new era of AI training that requires less computational power, making it accessible for integration into personal devices and smart electronics. This will not only streamline operations but also promote sustainability by reducing reliance on energy-intensive data centers."

In conclusion, VFF-Net holds immense promise as it paves the way for advanced, more efficient AI systems that mimic human-like learning processes. As AI continues to advance, VFF-Net's contributions may well redefine the landscape of intelligent technology, rendering it faster, more reliable, and environmentally friendly.

For those interested in exploring the details, the full study on VFF-Net was published in the journal Neural Networks on October 1, 2025, and is available online.

Reference

• - Title of original paper: VFF-Net Evolving forward–forward algorithms into convolutional neural networks for enhanced computational insights
• - Journal: Neural Networks
• - DOI: 10.1016/j.neunet.2025.107697

About SEOULTECH

• - Website: SEOULTECH