Multi-Task Learning at the Mobile Edge: An Effective Way to Combine Traffic Classification and Prediction

A General Approach for Traffic Classification in Wireless Networks Using Deep Learning

“…To overcome this problem, MTL should be explored to reduce the computing/storage requirements, achieve higher performance, and simplify the training procedure. As recently investigated and demonstrated in [21], and combined with distributed learning approaches, such as Gossip Learning (GL) used in [23], MTL might speed up the learning process and increase the system scalability. A related problem is the possible bias and damage on the model performance caused by using fixed-length input as required by CNNs in the presence of significant variations in the input data's length.…”

“…As pointed out by the authors in [21], most of the literature in the field of NMS are focused on single-task learning, e.g., each model is designed and trained to solve one specific learning task such as TC, traffic prediction, or anomaly detection. As a solution, Multi-Task Learning (MTL) approaches have been proposed in [21] and [22], where TC is used as one of the learning tasks, to leverage useful information contained in multiple related tasks aiming to improve the generalization capabilities of all them while learning.…”

“…As pointed out by the authors in [21], most of the literature in the field of NMS are focused on single-task learning, e.g., each model is designed and trained to solve one specific learning task such as TC, traffic prediction, or anomaly detection. As a solution, Multi-Task Learning (MTL) approaches have been proposed in [21] and [22], where TC is used as one of the learning tasks, to leverage useful information contained in multiple related tasks aiming to improve the generalization capabilities of all them while learning. The authors in [21] used a MTL approach to jointly solve the TC and traffic Their proposal is a two step procedure where they first use an AE to extract common feature representations among tasks and then use the encoder part of it to train a traffic classifier (softmax layer with a softmax activation function) and a traffic predictor (a dense layer with the Rectified Linear Unit (ReLU) activation function) simultaneously.…”

“…As a solution, Multi-Task Learning (MTL) approaches have been proposed in [21] and [22], where TC is used as one of the learning tasks, to leverage useful information contained in multiple related tasks aiming to improve the generalization capabilities of all them while learning. The authors in [21] used a MTL approach to jointly solve the TC and traffic Their proposal is a two step procedure where they first use an AE to extract common feature representations among tasks and then use the encoder part of it to train a traffic classifier (softmax layer with a softmax activation function) and a traffic predictor (a dense layer with the Rectified Linear Unit (ReLU) activation function) simultaneously. Compared to conventional single-task learning approaches, which do not use AE and tackle classification and prediction tasks separately, the MTL approach always provided the highest performance.…”

Distributed and Multi-Task Learning at the Edge for Energy Efficient Radio Access Networks

“…Another ML paradigm that can help in reducing the implementation complexity without sacrificing ANN's universal function approximation property is represented by MTL [36]. As KTL, MTL is animated by the human learning principle of transferring the acquired knowledge: often people apply some ability, learned from previous tasks, to help learn a new task.…”