[Paper Review] Caching as an Image Characterization Problem using Deep Convolutional Neural Networks

Yantong Wang, Vasilis Friderikos|arXiv (Cornell University)|Jul 16, 2019

Caching and Content Delivery23 references4 citations

TL;DR

This paper proposes a novel deep learning framework that transforms proactive caching optimization into a grayscale image classification problem, using convolutional neural networks (CNNs) to predict optimal edge caching locations. By training on MILP-generated optimal solutions, the CNN achieves real-time decision-making with performance over 400% better than randomized greedy heuristics, significantly outperforming traditional methods in speed and accuracy while maintaining high feasibility and precision.

ABSTRACT

Caching of popular content closer to the mobile user can significantly increase overall user experience as well as network efficiency by decongesting backbone network segments in the case of congestion episodes. In order to find the optimal caching locations, many conventional approaches rely on solving a complex optimization problem that suffers from the curse of dimensionality, which may fail to support online decision making. In this paper we propose a framework to amalgamate model based optimization with data driven techniques by transforming an optimization problem to a grayscale image and train a convolutional neural network (CNN) to predict optimal caching location policies. The rationale for the proposed modelling comes from CNN's superiority to capture features in grayscale images reaching human level performance in image recognition problems. The CNN is trained with optimal solutions and numerical investigations reveal that the performance can increase by more than 400% compared to powerful randomized greedy algorithms. To this end, the proposed technique seems as a promising way forward to the holy grail aspect in resource orchestration which is providing high quality decision making in real time.

Motivation & Objective

• To address the real-time decision-making challenge in proactive caching for 5G and beyond networks.
• To overcome the curse of dimensionality in NP-hard MILP formulations for content placement optimization.
• To leverage deep learning's feature extraction capability to model spatial request patterns in caching networks.
• To enable fast, high-precision, and feasible caching decisions through end-to-end training on optimal solutions.
• To demonstrate that CNN-based inference can match or exceed MILP performance with drastically reduced computation time.

Proposed method

• The original mixed-integer linear programming (MILP) caching optimization problem is mapped into a grayscale image representation of request patterns across edge clouds.
• A deep convolutional neural network (CNN) is trained to predict optimal caching policies directly from these grayscale images.
• The CNN is trained using optimal solutions generated by solving the MILP offline, enabling real-time inference.
• Spatial features in the image representation capture topological and demand relationships between users and edge clouds.
• The network architecture leverages weight-sharing and hierarchical feature learning, similar to image recognition, to generalize across network configurations.
• Performance is evaluated using metrics like mean total cost, precision, feasible ratio, and maximum cost deviation from optimal solutions.

Experimental results

Research questions

• RQ1 Can a deep CNN be effectively trained to predict optimal caching locations by treating the optimization problem as an image recognition task?
• RQ2 How does the CNN's performance compare to traditional greedy heuristics in terms of cost, precision, and feasibility under varying network scales?
• RQ3 To what extent can the CNN generalize across different network topologies and request patterns while maintaining high accuracy?
• RQ4 What is the trade-off between computational efficiency and solution quality when replacing MILP with a learned CNN policy?
• RQ5 How sensitive is the CNN's output to small variations in input request patterns, and does it exhibit adversarial robustness?

Key findings

• The CNN achieved a mean total cost of 25.07 for 10 requests, only 1.76% higher than the MILP solution (10.83), while reducing computation time from 80.96s to 0.51s.
• For 15 requests, the CNN's mean cost was 55.59, which was 40.5% lower than the RGC heuristic (98.72), representing a performance improvement of over 400%.
• The CNN maintained a 94.8% feasible ratio and 65.4% precision for 15 requests, outperforming RGC (89.8% feasible, 26.4% precision).
• In the 20-request scenario, the CNN's mean cost (124.66) was less than half of RGC's (156.36), with a 51.9% precision rate compared to RGC's 21.1%.
• The maximum cost deviation between CNN and MILP was 106.6 for 15 requests, significantly lower than RGC's 463.9, indicating better consistency.
• The CNN achieved near-optimal performance in under 2.2 seconds for 20 requests, while MILP required over 2.4 hours, demonstrating strong scalability for real-time use.