Popularity-Aware Neural Network-based Caching

초록

In-network caching has been widely adopted as an essential technology for enhancing service quality and promoting efficient network utilization. Recently, this concept has been extended to machine learning-based services, enabling improved service performance. In such applications, a 'cache-hit' does not necessarily indicate the presence of identical content in the cache memory. Rather, it signifies that a request can be served with sufficient confidence by leveraging the similarity of outputs from intermediate layers of a neural network (NN). Due to this distinction, unlike traditional in-network caching, there has been limited focus on optimizing caching efficiency in NN-based caches with consideration of data popularity. To address this, we propose the Popularity-Aware NN-caching (PANN) system, which biases the training of the NN-cache toward frequently requested data, thereby delivering more accurate responses to popular queries. In our approach, the popularity distribution of user requests is periodically collected by the service provider (e.g., Cloud) and used to train the PANN-cache via transfer learning. The trained cache is then deployed to end-user devices, embedded at the end of the NN's hidden layers within mobile platforms. Our comprehensive evaluation, conducted using three well-known datasets - CIFAR-100, Places365, and Food-101 - and two backbone models - ResNet-18 and ResNet-50 - demonstrates that the proposed PANN-cache outperforms existing NN-caching schemes by handling up to 38.86% more requests and improving prediction accuracy by 8.04%. Furthermore, compared to knowledge distillation approaches under identical memory constraints, the PANN-cache achieves up to 60.25% more predictions with sufficient confidence scores, along with their accuracy improvement of 27.04%. © 2026 Copyright held by the owner/author(s).

키워드