Personalized Neural Architecture Search

Existing approaches for Neural Architecture Search (NAS) aim at efficiently maximizing individual or sets of objectives (e.g. high accuracy or a low number of parameters) by exploiting Reinforcement Learning (RL), evolutionary algorithms, or Bayesian optimization. Most multi-objective NAS algorithms assume that all objectives are fully known and require them to be broadly explored to successfully approximate the Pareto front, which results in computational expensive search algorithms. To address this problem, we propose an interactive machine learning approach based on preference elicitation which enables end-users to explore and find a custom loss function and can be directly used for State-of-the-Art single-objective black-box optimization. We integrate our approach into State-of-the-Art single objective NAS algorithms and evaluate it against multi-objective approaches on the NATS-Bench benchmark dataset. Furthermore, we show that diverse end-user preferences can be successfully approximated in terms of loss functions, leading to suitable neural architectures.