A Fractional order method for Meta-Learning about aerospace target classification
Zhongliang Yu, Jianfeng Lv · IEEE Transactions on Aerospace and Electronic Systems · 2022
Meta-learning can sum up the experience from the learned tasks, and solve problems where datasets are scarce or expensive and model generalization, which traditional machine learning is inadequate. Meta-learning can obtain a well initial model, which can quickly generalize after a few adjustments to solve new tasks with good performance. However, MAML faces the problem of convergence difficulties, and easily prone to concussion late in training. We employ fractional order to MAML, which can retain past task gradient for a better stabilized and generalize model. We proposed a method based on fractional-order and model-agnostic meta-learning (MAML), and named as FracMAML, which can apply to a method based on meta-learning such as MAML, Reptile. Our proposed method can obtain a better initial model via fractional order. We take advantage of meta-learning on few-shot problems to solve the aerospace targets classification problem, and proposed an Aerospace dataset. We employed FracMAML on MiniImagenet and Omniglot, which obtain state of art accuracy than some classical meta-learning method such as MAML and Reptile and so on. Finally, we verify FracMAML method on aerospace targets classification task based on our Aerospace dataset, which performed well, further verify the versatility of our algorithm.