Meta-Learning without Memorization

By Mingzhang Yin, George Tucker, Mingyuan Zhou, Sergey Levine, Chelsea Finn. 2019

202 citations recorded by the source corpus.

The ability to learn new concepts with small amounts of data is a critical aspect of intelligence that has proven challenging for deep learning methods. Meta-learning has emerged as a promising technique for leveraging data from previous tasks to enable efficient learning of new tasks. However, most meta-learning algorithms implicitly require that the meta-training tasks be mutually-exclusive, such that no single model can solve all of the tasks at once. For example, when creating tasks for few-shot image classification, prior work uses a per-task random assignment of image classes to N-way classification labels. If this is not done, the meta-learner can ignore the task training data and learn a single model that performs all of the meta-training tasks zero-shot, but does not adapt effectively to new image classes. This requirement means that the user must take great care in designing the tasks, for example by shuffling labels or removing task identifying information from the inputs. In some domains, this makes meta-learning entirely inapplicable. In this paper, we address this challenge by designing a meta-regularization objective using information theory that places precedence on data-driven adaptation. This causes the meta-learner to decide what must be learned from the task training data and what should be inferred from the task testing input. By doing so, our algorithm can successfully use data from non-mutually-exclusive tasks to efficiently adapt to novel tasks. We demonstrate its applicability to both contextual and gradient-based meta-learning algorithms, and apply it in practical settings where applying standard meta-learning has been difficult. Our approach substantially outperforms standard meta-learning algorithms in these settings.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Meta & Continual Learning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Meta & Continual Learning

Citations in the corpus

Incremental Object Detection via Meta-Learning (2020, positive): “(b) Each task in the meta-train and meta-test stages is carefully designed to avoid task-overfitting [13].”
Meta-Learning in Neural Networks: A Survey (2020, background): “The most common application of meta-learning is few-shot multi-class image recognition, where the inner and outer loss functions are typically the cross entropy over training and validation data respectively [14], [23], [75], [77], [78], [88], [90], [98], [99], [102], [105], [201], [202], [203], [20”
Variable-Shot Adaptation for Online Meta-Learning (2020, methodology): “Following Finn et al. (2019) and Yin et al. (2019), we construct a multi-task dataset with 65 tasks based on the Pascal 3D data (Xiang et al., 2014) where each task corresponds to a different object with a random camera angle.”
Bootstrapped Meta-Learning (2021, background): “…lifetime performance (Lv et al., 2017; Wichrowska et al., 2017), collapsed exploration (Stadie et al., 2018; Chen et al., 2016), biased learner updates (Stadie et al., 2018; Zheng et al., 2018), and poor generalisation performance (Wu et al., 2018; Yin et al., 2020; Triantaﬁllou et al., 2020).”
Progressive Meta-Learning With Curriculum (2022, background): “Recent studies have shown that meta-learning models are prone to overﬁtting when there do not exist sufﬁcient diversity training tasks to generalize [55], [56].”
Sharp-MAML: Sharpness-Aware Model-Agnostic Meta Learning (2022, methodology): “In particular, model-agnostic meta learning (MAML) is one of the most popular optimization-based meta learning frameworks for few-shot learning (Finn et al., 2017; Vuorio et al., 2019; Yin et al., 2020; Obamuyide & Vlachos, 2019).”
How to Train Your MAML to Excel in Few-Shot Classification (2021, methodology): “3), which make the created tasks mutually-exclusive (Yin et al., 2020) — i.”
Improving Generalization by Controlling Label-Noise Information in Neural Network Weights (2020, background): “In a recent work, Yin et al. (2020) consider a similar term in the context of meta-learning and use it as a regularization to prevent memorization of meta-testing labels.”
A survey of deep meta-learning (2020, background): “Clever task design and meta-regularization may prove useful to avoid such problems (Yin et al., 2020).”
Curriculum-Based Meta-learning (2021, background): “Recent studies demonstrate that meta-learning models are prone to overfitting when there do not exist sufficient diversity training tasks for the meta-learners to generalize[19, 46, 51] (we think this is the major reason from the results in Section 5.”
Invariant Causal Prediction for Block MDPs (2020, background): “Recent prior works (Amit and Meir, 2018; Yin et al., 2019) have developed generalization bounds for the multi-task problem, but they depend on the number of tasks seen at training time, which can be prohibitively expensive given how sample inefﬁcient RL is even in the single task regime.”
Meta-Learning Requires Meta-Augmentation (2020, methodology): “Yin et al. [37] propose using a Variational Information Bottleneck [2] to regularize the model by restricting the information ﬂow between x q and y q .”
AdaRL: What, Where, and How to Adapt in Transfer Reinforcement Learning (2021, unknown): “With the choice of π = ∫ P(θ)Q(SD,θ)dθ and ρ = ∫ Q(θ)Q(SD,θ)dθ, we have that DKL(ρ||π) ≤ DKL(Q||P) (Yin et al., 2019).”
Training Language Models to Self-Correct via Reinforcement Learning (2024, unknown): “Conceptually, this is similar to the memorization challenge in meta-learning (Yin et al., 2019), which suggests that when provided with mutually exclusive tasks, meta-learning is likely to recover the supervised learning solution (without using context from the few shots) that directly predicts the…”
Self-Supervised Meta-Learning for Few-Shot Natural Language Classification Tasks (2020, background): “While this enables generalization to new labels, it limits generalization to unseen tasks due to over-ﬁtting to the training task distribution (Yin et al., 2020).”