Towards in-network neural networks

Abstract

In recent years, machine learning (ML) techniques have been shown to be efficient when applied to network operation and management problems, and there are several recent studies that seek to apply these techniques in specific network areas and technologies. In this work, the first step is taken towards an intelligent data plan implementation using In-Network Neural Networks. The problem consists in mapping a set of Artificial Neural Networks Artificial Neural Networks (ANN) in a programmable network infrastructure, considering a series of restrictions (e.g., mapping only a single neuron per network device) that seek to optimize the operation of the ANNs in the network. To solve this problem, an optimization problem is initially formalized, using the Mixed-Integer Linear Programming (MILP) model. To try to find feasible solutions to the problem, we developed a math-heuristic based on solutions generated by a constructive heuristic, besides two other algorithms with polynomial complexity, based on random and greedy decisions. Both seek to find a valid solution to the problem with a minimum amount of resources used (for example, memory and processing). The objective is to evaluate the solutions found by the proposed strategies, in comparison to a reduction of the problem for an instance of the VNE, a problem similar to the one being solved, but with simpler restrictions. In addition, it seeks to assess the impact of some parameters and metrics, such as the number of available / used flows, on the quality of the solutions generated. The results show that the math-heuristic and VNE techniques can not generate scalable solutions (in terms of processing and memory) to the problem, being able to solve only small instances of the problem. On the other hand, the greedy and random algorithms can map the maximum number of ANNs possible, considering ANNs with 5 neurons and a fat-tree topology of 20 devices. Similarly, in the topology of 80 network devices, we map 87.5 % of the possible ANNs. In addition, the algorithms can map these ANNs with a low number of available network streams, finding, for example, the maximum number of mappings possible with only 20 flows available for the topology of 20 devices.