Source-linked AI summary
Task-Agnostic Meta-Learning for Few-shot Learning
Muhammad Abdullah Jamal, Guo-Jun Qi, Mubarak Shah
TL;DR
Few-shot meta-learning can produce an initial model biased toward existing tasks, making adaptation to divergent unseen tasks difficult with very few examples. TAML addresses this with entropy maximization for classification and inequality minimization for broader loss-based settings, and its methods outperform compared meta-learning algorithms in few-shot classification and reinforcement learning.
Problem
Existing meta-learning may bias its initial model toward meta-training tasks, limiting generalization to discrepant unseen tasks when few adaptation examples are available.
Method
TAML learns a task-agnostic initial model by maximizing prediction entropy for classification or minimizing inequality among initial losses across tasks.
Results
TAML consistently outperforms existing meta-learning algorithms on few-shot classification and reinforcement learning tasks.
Takeaways & Limitations
An unbiased task-agnostic prior is presented as a more generalizable starting point for adapting to unseen tasks.
Takeaways & Limitations
Entropy-based TAML is limited to discrete labels in classification tasks, while broader settings require an alternative loss-based inequality metric.
Abstract
from arXiv · showhide
Meta-learning approaches have been proposed to tackle the few-shot learning problem.Typically, a meta-learner is trained on a variety of tasks in the hopes of being generalizable to new tasks. However, the generalizability on new tasks of a meta-learner could be fragile when it is over-trained on existing tasks during meta-training phase. In other words, the initial model of a meta-learner could be too biased towards existing tasks to adapt to new tasks, especially when only very few examples are available to update the model. To avoid a biased meta-learner and improve its generalizability, we propose a novel paradigm of Task-Agnostic Meta-Learning (TAML) algorithms. Specifically, we present an entropy-based approach that meta-learns an unbiased initial model with the largest uncertainty over the output labels by preventing it from over-performing in classification tasks. Alternatively, a more general inequality-minimization TAML is presented for more ubiquitous scenarios by directly minimizing the inequality of initial losses beyond the classification tasks wherever a suitable loss can be defined.Experiments on benchmarked datasets demonstrate that the proposed approaches outperform compared meta-learning algorithms in both few-shot classification and reinforcement learning tasks.
1 Introduction
Few-shot learning seeks rapid adaptation from limited labeled examples, while meta-learning aims to generalize across tasks. The paper introduces TAML to reduce bias toward meta-training tasks and presents entropy- and inequality-based approaches, evaluated in classification and reinforcement learning.
- Motivation: Few-shot learning aims to enable rapid adaptation from a small number of labeled examples, reducing the need for large training sets.The motivation is to learn from prior experience while avoiding cumbersome large-scale label collection.
- Meta-learning background: Meta-learning trains an initial model and updating strategy across varied tasks to generalize to new tasks.The approach seeks parameters that transfer across a variety of tasks.
- Problem: Existing meta-learning can bias the initial model toward sampled meta-training tasks, weakening generalization when unseen tasks differ and examples are scarce.The paper addresses this by preventing overperformance on particular tasks or minimizing performance inequality across tasks.
- Proposed approach: TAML proposes entropy-based and inequality-minimization approaches to learn a less task-biased initial model.The entropy approach targets uncertainty in classification outputs, while inequality minimization targets losses across tasks.
- Scope and evaluation: The entropy-based approach is limited to discrete outputs, whereas inequality minimization can use losses in scenarios beyond classification.The paper reports experimental studies on few-shot classification and reinforcement learning.
2 Approach
TAML addresses task bias in few-shot meta-learning by shaping an initial model that avoids over-performing on particular training tasks. It uses entropy-based regularization for classification and inequality minimization over task losses for broader settings.
- Meta-learning setup: The meta-learning setup trains an initial model and update rule across sampled tasks so the learner can rapidly adapt to a new task from few examples.In the MAML example, task-specific parameters are updated by gradient descent using the task loss.
- Motivation: TAML targets initial-model bias toward meta-training tasks, which can impair adaptation to discrepant unseen tasks when few examples are available.The approach also aims to avoid an over-performing initial model that could hinder learning a consistent update rule.
- Entropy-Maximization/Reduction TAML: Entropy-based TAML maximizes the initial model’s label entropy to prevent over-performance on particular classification tasks, then encourages entropy reduction after adaptation.The entropy term is combined with the usual meta-training objective as a regularizer, with λ balancing the terms.
- Inequality-Minimization TAML: Entropy-based TAML is limited to discrete-label classification, whereas inequality-minimization TAML operates on task losses and can use any suitable loss beyond classification.Inequality minimization is computed over losses from sampled tasks, alongside expected post-update loss.
- Inequality-Minimization TAML: Inequality-minimization TAML treats each task’s initial loss as an outcome and minimizes loss inequality across tasks to obtain a task-agnostic initial model.The framework considers inequality statistics including the Theil Index and uses ℓ_i = L_Ti(f_θ), with the mean loss computed across M sampled tasks.
3 Related Work
Meta-learning methods learn model initialization and update strategies across tasks for rapid adaptation, but the literature also includes memory-based and architecture-specific alternatives.
- Meta-learning approaches: Meta-learning trains a learner through a meta-learner that learns both an initialization and an updating strategy generalizable across tasks.MAML instead learns an initialization adapted through fixed gradient descent, while Meta-SGD also learns update step sizes.
- Meta-learning approaches: LSTM-based meta-learners use gradients to propose learner updates, while MAML learns initialization through fixed gradient descent.These approaches target rapid adaptation using few or one gradient steps.
- Comparison methods: The paper positions TAML against established meta-learning methods including Meta-LSTM, MAML, Meta-SGD, and SNAIL.SNAIL is described as a temporal convolution and attention-based meta-learner for few-shot classification and reinforcement learning.
- Alternative approaches: Other approaches use memory-augmented networks or architecture-specific methods such as Siamese and differentiable nearest-neighbor models.These alternatives address few-shot classification without necessarily using the same meta-learning framework.
4 Experiments
Experiments evaluate TAML on benchmark few-shot classification datasets and a 2D navigation reinforcement-learning task, comparing it with established baselines under few-shot adaptation settings.
- Few-shot classification: Experiments cover Omniglot and Mini-Imagenet few-shot classification, using 5-way and 20-way settings with 1-shot and 5-shot conditions.Omniglot contains 1623 characters, while Mini-Imagenet uses 64 training, 12 validation, and 20 test classes.
- Few-shot classification: TAML outperforms MAML and Meta-SGD on Omniglot 20-way convolutional classification under the reported fair-comparison splits.Compared results were re-run on the same training/test split used for TAML models.
- Few-shot classification: TAML outperforms compared methods on Mini-Imagenet 5-way classification and outperforms both MAML and Meta-SGD in the 20-way setting.Entropy-based TAML performs best against inequality-minimization TAML for the 5-shot 5-way problem.
- Reinforcement learning: Reinforcement-learning experiments sample trajectories instead of images, using vanilla policy gradient for inner updates and TRPO as meta-optimizer.The experiments were performed with the rllab suite.
- Reinforcement learning: The reinforcement-learning evaluation uses a 2D navigation task where an agent reaches a randomly sampled goal in a unit square.The policy network matches the architecture used in prior work, with three ReLU layers and step size α = 0.1.
- Reinforcement learning: GE(0), Theil, and GE(2) TAML perform on-par with MAML after two gradient steps but outperform it afterward.Baselines include an oracle policy, conventional pre-training, and random initialization.
5 Conclusion
The paper proposes TAML to train task-agnostic initial models, reporting consistent improvements over existing meta-learning algorithms in few-shot classification and reinforcement learning.
- Conclusion: TAML trains a meta-learner to remain unbiased across tasks before adapting its initial model to unseen tasks.The paper presents entropy-based and inequality-minimization variants.
- Conclusion: The reported experiments demonstrate consistent TAML improvements over existing meta-learning algorithms on few-shot classification and reinforcement learning tasks.The conclusion frames task-agnostic priors as potentially more generalizable to new tasks.