Improving Neural Architecture Search with Class Visualization and Bilevel Optimization for Imbalanced Data

Neural Architecture Search (NAS) has proven to be a highly effective technique for automating neural network design, overcoming challenges posed by complex datasets, class imbalance, and the high time consumption associated with manual architecture selection. Although NAS shows promise, it frequently suffers from high computational costs and the risk of data loss due to insufficient training. This paper presents a novel NAS method that incorporates class visualization to mitigate computational costs and a bilevel optimization framework to address class imbalance. The hierarchical structure comprises an upper level formulated as a multi-objective optimization problem to maximize performance, measured by the Matthews correlation coefficient, while minimizing model complexity. The lower level focuses on neural network training using a reduced dataset obtained by class visualization. The proposed approach combines NSGA-II (for the upper-level problem) with stochastic gradient descent (for the lower-level problem). This method significantly reduces NAS runtime on MNIST and CIFAR-10 subsets while maintaining accuracy and robustness.