Automated Model Design using Gated Neuron Selection in Telecom

TL;DR

TabGNS introduces a novel gradient-based neural architecture search for tabular telecom data, leveraging neuron-level gating to automatically design compact, high-performance MLPs within the model lifecycle management framework. By replacing layer-level search with individual neuron gates and using Gumbel-Softmax with Straight-Through Estimation, TabGNS performs joint architecture search and training, yielding architectures that match or exceed large MLP baselines while reducing parameters by 51–82% and achieving up to 36× faster search times. Across six tabular datasets, including telecom RSS and VoD metrics plus two generic benchmarks, TabGNS demonstrates robust performance gains and consistent convergence where other tabular NAS methods struggle. The approach enables automated, scalable, and efficient model design for resource-constrained telecom environments, supporting rapid retraining and deployment in lifecycle management contexts, with future work extending to other layer types and datasets.

Abstract

The telecommunications industry is experiencing rapid growth in adopting deep learning for critical tasks such as traffic prediction, signal strength prediction, and quality of service optimisation. However, designing neural network architectures for these applications remains challenging and time-consuming, particularly when targeting compact models suitable for resource-constrained network environments. Therefore, there is a need for automating the model design process to create high-performing models efficiently. This paper introduces TabGNS (Tabular Gated Neuron Selection), a novel gradient-based Neural Architecture Search (NAS) method specifically tailored for tabular data in telecommunications networks. We evaluate TabGNS across multiple telecommunications and generic tabular datasets, demonstrating improvements in prediction performance while reducing the architecture size by 51-82% and reducing the search time by up to 36x compared to state-of-the-art tabular NAS methods. Integrating TabGNS into the model lifecycle management enables automated design of neural networks throughout the lifecycle, accelerating deployment of ML solutions in telecommunications networks.

Figures (5)

• Figure 1: Workflow of the operational steps in the ML model lifecycle management etsi_study_2024 augmented with TabGNS for automated neural architecture search and training.
• Figure 2: TabGNS outperforms previous tabular NAS methods on all of the critical dimensions; prediction error, architecture size, and search time. Scatter-point size represents architecture size. Results for VoD dataset.
• Figure 3: Representation of the learning process of TabGNS. (a) Gate for each neuron is initialised to a be open at a low probability. (b) Architecture and weights are iteratively updated until convergence. (c) Neurons with gates with probability $p_{\text{open}}>0.5$ are extracted together with their weights. (d) Final trained fully connected neural network, can be further fine tuned if needed.
• Figure 4: Size of the architecture during a search process given initialisation of gates with low/high probability of being open, for the VoD dataset. Initialisation of the gates determines the starting size of the architectures and consequentially the size of the final architecture.
• Figure 5: Architecture size and prediction error given initial gate state on VoD dataset. With low initial gate probability, TabGNS is able to find small architectures with low error.