GPU: A New Enabling Platform for Real-Time Optimization in Wireless Networks

Quality Monitoring and Assessment of Deployed Deep Learning Models for Network AIOps

“…Such quality assessment can apply (i) to individual inferences, enriching the DL output with third-party confidence about its accuracy, which can be useful to decide whether to trust punctual decisions of a DL model, e.g., a traffic classification result [2]. Moreover, quality assessment can span (ii) over multiple inferences, which allows tracking the overall quality of a DL model over time, e.g., which can be useful to trigger model retraining, or model switching, if environmental conditions change [1].…”

“…Finally, (iii) flexibility to assess quality of individual model inferences, as well as ability to track the overall model quality over time, are important practical features. 5) Efficiency and timeliness: From a (i) computational complexity viewpoint, it appears necessary for DL model quality assessment to be a relatively simple operation [1]. Ideally, quality assessment of a model output should have a cost comparable to the inference cost of the same modelwhich would allow to track the quality of deployed models at a fine grain.…”

“…In particular, spectacular successes of Deep Learning (DL) techniques have gained significant share of newspaper attention in the image classification field, with numerous models already deployed in production. In a nutshell, these DL advances have equally benefited from ground-breaking theoretic research 1 , from software plaforms efficiently exploiting hardware acceleration 2 , and from the availability of large corpuses of labeled data such as ImageNet (comprising tens of millions of images for tens of thousands of classes).…”

“…The network research community, as many other fields, is exploring the use of DL models to relieve and assist human operation for diverse tasks, such as dynamic resource management [1], traffic classification and management [2] and troubleshooting. Yet, the network fast paced evolution, combined to the lack of a universal network data format and privacy/business-sensitive matters, have so far made it impossible to construct and share large corpuses of networking data, which were crucial for DL success in the image field.…”

“…This yields to several challenges for DL deployment in the network field. First, the knowledge gained by DL models during training is hard to transfer, thus special attention should be payed to assess whether the model is fit to the environmental conditions where it has been deployed: e.g., a DL model for resource management [1] trained on outdoor settings is hardly applicable to indoor ones. Second, as per the just observed fast-paced evolution of the network traffic and technological landscape, it is reasonable to expect that any deployed DL model should be regularly updated, to remain fit to the job it has been originally designed for: for instance, a DL model for traffic classification [2] will encounter new applications it was not trained to recognize.…”