元学习综述：迈向自动化AI的高效模型自适应

需积分: 8 108 浏览量更新于2024-07-09 收藏 1.06MB PDF 举报

元学习（Meta-learning），也被称为学习如何学习，是近年来人工智能领域的重要研究方向，尤其在高度自动化AI、少样本学习、自然语言处理和机器人技术等领域展现出强大的潜力。本文是一篇28页的综述论文，深入探讨了元学习的最新进展。论文的核心内容主要集中在以下几个方面： 1. **定义与目标**：元学习关注的是机器学习模型对未见过任务的快速和准确适应，与深度学习的主要区别在于深度学习专注于预测训练样本，而元学习则着眼于模型的泛化能力，旨在处理高维、少量样例的新任务。 2. **应用场景**：元学习对于实现高度自主的人工智能至关重要，它允许模型通过自我改进来应对未知挑战，无需从头开始训练。例如，在自动驾驶、医疗诊断等场景中，模型能够迅速适应新的环境或疾病模式。 3. **优势与补充**：相对于传统的深度学习模型，元学习可以作为一种额外的泛化手段，增强模型对未知数据的处理能力，特别是在资源有限或者新任务频繁出现的情况下，其适应性优势更加明显。 4. **方法与策略**：元学习通常采用策略如MAML (Model-Agnostic Meta-Learning) 和Prototypical Networks，它们通过学习通用的特征表示或者策略，使得模型能够在面临新的任务时通过少量调整就能达到良好的性能。 5. **协同进化视角**：论文还讨论了元学习中的协同进化思想，即在环境与代理之间进行交互学习，这对于解决那些单纯训练无法解决的复杂问题提供了可能，这种动态适应机制使得模型在面对动态变化的世界时更具竞争力。 6. **未来趋势与挑战**：尽管元学习取得了显著进展，但仍面临诸如样本效率、计算成本、可解释性等方面的挑战。随着技术的发展，如何进一步提升元学习的效率和普适性，以及在实际应用中如何更好地结合其他机器学习技术，将是未来研究的重点。这篇综述论文全面梳理了元学习的最新研究成果，为理解元学习的基本原理、方法和潜在应用提供了一个坚实的基础，也为该领域的研究人员和实践者提供了深入研究和探索的方向。

A Comprehensive Overview and Survey of Recent Advances in Meta-Learning A PREPRINT

•

Relation Network [

] contains meta-parameters in feature extraction neural network and meta-parameters

in distance metrics. Feature extraction model and distance measure are jointly tuned to represent similarity

between tasks. These metric-based meta-learning methods are very similar in structure but include different

degrees of ﬂexibility in training.

•

Prototypical Network [

] measures distance between extracted feature embeddings and these centroids from

different classes. There are many methods built upon Prototypical Network which show superior performance

in few-shot image classiﬁcation, as in [

] and [

]. [

] integrates domain adaptation into Prototypical

Network and shifts and scales class centroids in adaptation to unseen tasks. [

] utilizes adaptive margin loss

in deep metric learning of Prototypical Network to improve classiﬁcation power.

•

Dynamic few-shot [

] integrates a ConvNet-based classiﬁer, a few-shot classiﬁcation weight generator and

a classiﬁer based upon cosine distance metric. Classiﬁcation weight remains ﬁxed across all seen classes

and can be generalized to unseen tasks. For unseen classes, simply add extra neurons to the output layer and

use the few-shot classiﬁcation weight generator to estimate additional parameters. Feature extraction neural

network and classiﬁcation weight generator for testing data jointly represent similarity shared by different

tasks. Complexity of this framework may be tuned to match the complexity of real similarity between tasks.

•

MAP [

] (Mean Average Precision) designs a similarity-ranking based measure as the meta-objective function

to estimate meta-parameters. It measures the difference between predicted similarity and real similarity between

tasks. For each individual sample in a batch of data, the other samples in the same batch are ranked by predicted

similarity with this individual sample. When processing each individual sample, relevant model experiences

from relevant samples are referred to. In mAP, memory module scheme is used much more often than other

methods, since memory module is only consulted on the unit of tasks not on the unit of sample points.

Layered meta-learning relies upon task decomposition to construct generalization tree and consists of base learner and

meta-learners for different generalization scopes:

•

MAML [

] is the most popular methods within GBML (Gradient-based meta-learning) methods. Meta-

parameter is the initial value of task-speciﬁc parameter on each task. Meta-parameter is updated in meta-

learner by minimizing loss function on validation data. Task-speciﬁc parameter is updated in base learner by

minimizing loss function on training data. MAML and GBML are compatible with all learners optimized using

stochastic gradient descent. Most deep neural networks can be optimized through gradient back-propagation

and can use MAML or GBML to improve model generalization capability. Since MAML includes SGD

update in both base learner and meta-learner, MAML introduces second-order derivative of loss function which

costs computational time. Later Reptile [

] and FOMAML [

] are proposed as ﬁrst-order approximation of

MAML. Reptile and FOMAML are faster but less accurate than MAML.

•

Meta-LSTM [

] uses LSTM as meta-learner and is also compatible with any learner optimized by SGD.

LSTM is one of the most important forms of RNN. LSTM contains memory cells, self-connected recursive

neurons, multiplicative gates and forget gates. LSTM ensures that gradients do not explode or vanish in

back-propagation and can account for long-range dependence in sequential data. Memory cells contain critical

past model training experiences and forget gates ensure that memory do not explode. LSTM can model

task similarity of any complexity and self-connected recursive neurons allow self-improvement along the

meta-training process. There is correspondence between LSTM cell calculations and stochastic gradient

descent. Specifying gates in LSTM properly, calculations in LSTM cell are exactly stochastic gradient descent.

•

R2-D2 and LR-D2 [

] considers using widely applicable base learners such as ridge regression and logistic

regression. These base learners are of less complexity than deep models, but they are also widely applied in

real-life applications. Meta-learner is speciﬁed to be neural network model that has high representation power

for task similarity of any complexity. Base learner of lower complexity and Meta-learner of high complexity is

an efﬁcient way to specify layered meta-learning model. Base learner is fast to train on each few-shot task and

meta-learner slowly uncovers shared features of all tasks to update meta-parameter.

•

TPN [

] applies end-to-end transductive learning between meta-training data and meta-testing data. Feature

extraction model and label propagation model based upon graph are jointly trained to reﬂect the similarity

between unseen tasks and trained tasks. Transductive label propagation model based upon graph is regarded

as the meta-learner. Data variance is taken into account when modelling similarity metric between samples.

Meta-parameters are parameters in feature extraction module and parameters in data variance model. Label

propagation for each task predicts labels jointly for all input data using an explicit solution formula, which

constitutes a very efﬁcient base learner.

•

LEO [

] utilizes a dimension reduction encoder network, a relation network and a decoder function. Encoder

network performs feature extraction, relation network measures similarity between samples with data variance

A Comprehensive Overview and Survey of Recent Advances in Meta-Learning A PREPRINT

accounted, decoder network projects features to classiﬁer. For generalization to unseen tasks, parameters in

all these modules are updated with MAML. Meta-learner provides initial values for these parameters in base

learner during generalization to unseen tasks. In the inner loop, base learner considers task training data and

updates these parameters from initial values supplied from meta-learner.

Bayesian meta-learning recasts meta-learning to be in the probabilistic framework and integrates Bayesian great minds

thoughts into meta-learning:

•

Bayesian program learning [

] generates hand-written characters indistinguishable from real hand writings

using visual checks. Joint distribution of character types, image-speciﬁc parameters and images are utilized to

simulate data indistinguishable from real data. Simulated data is augmentation to the original few-shot dataset

and contributes to improvement in predictive accuracy in few-shot learning. Joint posterior distribution of

image-speciﬁc parameters and character types is updated during model generalization to unseen tasks.

•

AI-GAs [

] consist of meta-learning architecture, learner algorithms and generation of more complex en-

vironment. AI-GAs consider coevolution between agents and environment. AI-GAs provide solutions for

complex environments which are otherwise unsolvable by training from scratch. AI-GAs contains important

interaction between agent and environment. Within evolution, as environment becomes more complex, so does

the solver agent. Each small change in environment corresponds to each small improvement in solver agent.

Evolution of environments guides solver agent to grow and to tackle complex tasks which cannot be solved by

tranining from scratch. AI-GAs are optimization techniques which use meta-learner to guide base learner to

ﬁnd optimum closer to global optimum.

•

Neural Statistician [

100

] applies variational autoencoder (VAE) to approximate the posterior distribution of

task-speciﬁc context. Given posterior distribution, context corresponding to highest posterior probability

is the predicted label of input data. From posterior distribution, we can also obtain conﬁdence interval of

unknow label on input data. VAE is an efﬁcient Bayesian inference tool which not only provides estimation on

unknown quantity but also the uncertainty in predicted value.

•

LLAMA [

101

] recasts MAML as probabilistic inference in a hierarchical Bayesian model. It uses ﬁrst-order

and second-order Laplace approximation to Gaussian distribution in log likelihood function. For other data

distributions such as multi-modal distribution, highly skewed distribution or fat-tailed distribution, Laplace

approximation does not work well. Laplace approximation shows best results for tightly distributed, symmetric

distributions such as Gaussian distrbution. All distributions in Bayesian framework are taken to be Gaussian so

that log likelihood function can be well approximated by Laplace approximation, which uses point estimation

at the mode to approximate integral of Gaussian distribution.

•

BMAML [

] replaces SGD (Stochastic gradient descent) in MAML with SVGD (Stein variational gradient

descent), where SVGD is an efﬁcient Bayesian inference method combining MCMC and variational inference.

SVGD does not only work for all learners optimized with SGD. It uses chaser loss based upon SVGD to

estimate meta-parameter and Bayesian fast adaptation to update task-speciﬁc parameters.

•

VERSA [

] relies upon distributional Bayesian decision theory for amortized variational inference. Posterior

distribution of true label is approximated and updated during model generalization to unseen tasks.

•

Under Bayesian framework, not only task-speciﬁc parameters and meta-parameters are regarded as random

variables, task can also be viewed as a random variable. Task can be considered as countably inﬁnitely many

and Chinese restaurant process is used as the prior distribution for task. Posterior distribution of task is

continually updated and maximum posterior estimate of task is provided in the model generalization process.

All task-speciﬁc parameters are taken to be functions of task. Estimation and uncertainty measure for task-

speciﬁc parameter are derived from posterior distribution of task. Meta-parameters should be updated using

aggregation of all task posterior distributions.

2.1 Black-Box Meta-Learning

Hyperparameter optimization can be achieved through random grid search or manual search [

]. Model search space

is usually indexed by hyperparameters [

]. In adaptation to novel tasks, hyperparameters are re-optimized using data

from the novel task. Optimizers can be approximated with neural networks or reinforcement learners [

]. Neural

networks can approximate any function with good convergence results. By using neural networks, the optimizer

search space represents a wide range of functions that guarantee better potential optima. For neural network with

varying architectures, hyperparameters include number of layers, number of block modules, number of neurons etc.

These hyperparameters are discrete and optimization techniques are different from continuous hyperparameters such

as learning rate, momentum etc. Bayesian optimization, reinforcement learning optimization, genetic programming

optimization may be employed to optimize both discrete and continuous hyperparameters in deep neural network model.

剩余46页未读，继续阅读

guduhanyan

粉丝: 1
资源: 36

元学习综述：迈向自动化AI的高效模型自适应

metalearning-slides.pdf

2019_Online Meta-Learning.pdf

matlab meta-learning

spring.datasource.dynamic.datasource.meta-db.druid.initial-size=8 spring.datasource.dynamic.datasource.meta-db.druid.max-active=16 spring.datasource.dynamic.datasource.meta-db.druid.min-idle=1

Model-Agnostic Meta-Learning

META-INF.services

meta-baseline: exploring simple meta-learning for few-shot learning

ServiceLoader加载不到META-INF.services目录下的类文件

安卓编译报错 2 files found with path 'META-INF/NOTICE.md' 怎么解决

最新资源