AutoMAML-HPO: Meta-Learning to Configure the Inner Loop Using Dataset Meta-Features

Meta-learning has been successfully used to recommend algorithms and hyperparameters in classical ML (Treder-Tschechlov et al., 2023; Marinho et al., 2024; Wang et al., 2024), but rarely to configure the inner workings of MAML itself. AutoMAML-HPO trains a meta-recommender that, given task meta-features (e.g., class imbalance, support-set diversity, gradient Lipschitz proxies, modality cues), outputs:

• Number of inner steps, per-layer learning rates, and regularization strength.
• Whether to include gradient-similarity loss (Tak & Hong, 2024) or trust-region constraints (Occorso et al., 2022).
• Data augmentation knobs (e.g., channel exchanging; Zhang et al., 2022) suited to the task’s heterogeneity.
  This builds a bridge between CASH-style meta-learning and gradient-based meta-learning: the outer meta-learner doesn’t just learn an initialization but also a policy for configuring adaptation. Unlike fixed hyperparameters in Finn et al. (2017), AutoMAML-HPO provides task-specific adaptation schedules, which is particularly important when domain shift is large (Satrya & Yun, 2023) or resources are tight (wireless/O-RAN scenarios; Lotfi & Afghah, 2024). The anticipated impact is better robustness and reduced manual tuning for practitioners, turning MAML into a mostly “plug-and-play” few-shot solution.

References:

1. Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks. Chelsea Finn, P. Abbeel, S. Levine (2017). International Conference on Machine Learning.
2. Enhancing Model Agnostic Meta-Learning via Gradient Similarity Loss. Jae-Ho Tak, Byung-Woo Hong (2024). Electronics.
3. Hyper-parameter Optimization for Wireless Network Traffic Prediction Models with A Novel Meta-Learning Framework. Liangzhi Wang, Jie Zhang, Yuan Gao, Jiliang Zhang, Guiyi Wei, Haibo Zhou, Bin Zhuge, Zitian Zhang (2024). arXiv.org.
4. Meta Reinforcement Learning Approach for Adaptive Resource Optimization in O-RAN. Fatemeh Lotfi, F. Afghah (2024). IEEE Wireless Communications and Networking Conference.
5. Optimization on selecting XGBoost hyperparameters using meta‐learning. Tiago Lima Marinho, Diego Carvalho do Nascimento, Bruno A. Pimentel (2024). Expert Syst. J. Knowl. Eng..
6. ML2DAC: Meta-Learning to Democratize AutoML for Clustering Analysis. Dennis Treder-Tschechlov, M. Fritz, H. Schwarz, B. Mitschang (2023). Proc. ACM Manag. Data.
7. Improving Generalization of Model-Agnostic Meta-Learning by Channel Exchanging. Ce Zhang, Ruixuan Chen, Yifeng Zeng, Shaolong Ren, Qingshan Cui (2022). 2022 International Conference on Electronics and Devices, Computational Science (ICEDCS).
8. Trust Region Meta Learning for Policy Optimization. Manuel Occorso, Luca Sabbioni, A. Metelli, Marcello Restelli (2022). Meta-Knowledge Transfer @ ECML/PKDD.
9. Combining Model-Agnostic Meta-Learning and Transfer Learning for Regression. Wahyu Fadli Satrya, Ji-Hoon Yun (2023). Italian National Conference on Sensors.