Meta-Learning with Implicit Gradients

被引:0
|
作者
Rajeswaran, Aravind [1 ]
Finn, Chelsea [2 ]
Kakade, Sham M. [1 ]
Levine, Sergey [2 ]
机构
[1] Univ Washington, Seattle, WA 98195 USA
[2] Univ Calif Berkeley, Berkeley, CA 94720 USA
来源
ADVANCES IN NEURAL INFORMATION PROCESSING SYSTEMS 32 (NIPS 2019) | 2019年 / 32卷
关键词
D O I
暂无
中图分类号
TP18 [人工智能理论];
学科分类号
081104 ; 0812 ; 0835 ; 1405 ;
摘要
A core capability of intelligent systems is the ability to quickly learn new tasks by drawing on prior experience. Gradient (or optimization) based meta-learning has recently emerged as an effective approach for few-shot learning. In this formulation, meta-parameters are learned in the outer loop, while task-specific models are learned in the inner-loop, by using only a small amount of data from the current task. A key challenge in scaling these approaches is the need to differentiate through the inner loop learning process, which can impose considerable computational and memory burdens. By drawing upon implicit differentiation, we develop the implicit MAML algorithm, which depends only on the solution to the inner level optimization and not the path taken by the inner loop optimizer. This effectively decouples the meta-gradient computation from the choice of inner loop optimizer. As a result, our approach is agnostic to the choice of inner loop optimizer and can gracefully handle many gradient steps without vanishing gradients or memory constraints. Theoretically, we prove that implicit MAML can compute accurate meta-gradients with a memory footprint no more than that which is required to compute a single inner loop gradient and at no overall increase in the total computational cost. Experimentally, we show that these benefits of implicit MAML translate into empirical gains on few-shot image recognition benchmarks.
引用
收藏
页数:12
相关论文
共 50 条
  • [21] Meta-learning for fast incremental learning
    Oohira, T
    Yamauchi, K
    Omori, T
    ARTIFICAIL NEURAL NETWORKS AND NEURAL INFORMATION PROCESSING - ICAN/ICONIP 2003, 2003, 2714 : 157 - 164
  • [22] Learning to Propagate for Graph Meta-Learning
    Liu, Lu
    Zhou, Tianyi
    Long, Guodong
    Jiang, Jing
    Zhang, Chengqi
    ADVANCES IN NEURAL INFORMATION PROCESSING SYSTEMS 32 (NIPS 2019), 2019, 32
  • [23] Subspace Learning for Effective Meta-Learning
    Jiang, Weisen
    Kwok, James T.
    Zhang, Yu
    INTERNATIONAL CONFERENCE ON MACHINE LEARNING, VOL 162, 2022, : 10177 - 10194
  • [24] Meta-Learning Representations for Continual Learning
    Javed, Khurram
    White, Martha
    ADVANCES IN NEURAL INFORMATION PROCESSING SYSTEMS 32 (NIPS 2019), 2019, 32
  • [25] TOOLS AND TASKS FOR LEARNING AND META-LEARNING
    Jaworski, Barbara
    JOURNAL OF MATHEMATICS TEACHER EDUCATION, 2005, 8 (05) : 359 - 361
  • [26] Tools and Tasks for Learning and Meta-learning
    Barbara Jaworski
    Journal of Mathematics Teacher Education, 2005, 8 (5) : 359 - 361
  • [27] On meta-learning rule learning heuristics
    Janssen, Frederik
    Fuernkranz, Johannes
    ICDM 2007: PROCEEDINGS OF THE SEVENTH IEEE INTERNATIONAL CONFERENCE ON DATA MINING, 2007, : 529 - 534
  • [28] Probabilistic Active Meta-Learning
    Kaddour, Jean
    Saemundsson, Steindor
    Deisenroth, Marc Peter
    ADVANCES IN NEURAL INFORMATION PROCESSING SYSTEMS 33, NEURIPS 2020, 2020, 33
  • [29] Research directions in meta-learning
    Vilalta, R
    Drissi, Y
    IC-AI'2001: PROCEEDINGS OF THE INTERNATIONAL CONFERENCE ON ARTIFICIAL INTELLIGENCE, VOLS I-III, 2001, : 354 - 360
  • [30] Meta-learning in active inference
    Penacchio, O.
    Clemente, A.
    BEHAVIORAL AND BRAIN SCIENCES, 2024, 47