Additive regularization schedule for neural architecture search

TL;DR

The paper proposes a neural architecture search framework that uses additive regularization scheduled across layers to balance accuracy and structural complexity. It introduces GA-REG to optimize per-layer regularization metaparameters and GA-NAS to prune structure, enabling joint optimization of weights, structure, and regularization strengths. Empirical results across diverse datasets show that properly scheduled additive regularization improves accuracy (up to ~30%) while reducing network size by more than a factor of two, and that regularized models exhibit greater stability. The approach integrates an autoencoder reconstruction objective with a predictive objective via a composite loss $S = \lambda_x E_x + \lambda_y E_y + \sum_i \lambda_i \mathcal{R}_i(\mathbf{W})$, and demonstrates practical NAS automation through expert and algorithmic scheduling.

Abstract

Neural network structures have a critical impact on the accuracy and stability of forecasting. Neural architecture search procedures help design an optimal neural network according to some loss function, which represents a set of quality criteria. This paper investigates the problem of neural network structure optimization. It proposes a way to construct a loss function, which contains a set of additive elements. Each element is called the regularizer. It corresponds to some part of the neural network structure and represents a criterion to optimize. The optimization procedure changes the structure in iterations. To optimize various parts of the structure, the procedure changes the set of regularizers according to some schedule. The authors propose a way to construct the additive regularization schedule. By comparing regularized models with non-regularized ones for a collection of datasets the computational experiments show that the proposed method finds efficient neural network structure and delivers accurate networks of low complexity.

Figures (7)

• Figure 1: Functional diagram of the proposed method
• Figure 2: Expert task of the optimization schedule
• Figure 3: Changes in the accuracy, complexity, and robustness of models with iterations of the GA-REG algorithm
• Figure 4: Changing the regularization metaparameters over iterations
• Figure 5: Left: the error (11) depends on the number of layers of the neural network given the method of tuning metaparameters. Right: the model error on the validation set over iterations of training.