Architecture generation for multi-objective neural architecture search

The attention given to architecture generation within neural architecture search (NAS) has grown significantly due to its efficient nature. By learning architectural representations through a generative model and establishing a latent space, the prediction process of predictors is simplified, thereby enhancing efficiency in architecture search. However, many NAS approaches prioritise identifying architectures solely based on accuracy, often neglecting architectural complexity. This paper presents a multi-objective NAS approach that integrates a multi-objective evolutionary algorithm (MOEA) with a generative model. This approach tackles the challenge by generating promising architectures while maintaining a balance between accuracy and complexity. Moreover, incorporating ranking errors assists in gradually regulating the generative model, thus aiding in the identification of promising representations. Besides, a MOEA, constructed with reference vectors, is utilised to preserve the quality of architectures. Experimental findings illustrate the effectiveness of the proposed approach in selecting architectures that achieve a balance between accuracy and complexity.