리뷰: A Survey on Few-ShotLearning

Generalizing from a Few Examples: A Survey on Few-ShotLearning

키워드:

• Few-Shot Learning
• One-Shot Learning
• Low-Shot Learning
• Small SampleLearning
• Meta-Learning
• Prior Knowledge

1 소개

• 튜링 머신
• 하드웨어 발전 (GPU, 분산 컴퓨팅)
• 모델의 발전
• 인간 챔피온을 이긴 모델들
• 현 딥러닝 모델의 단점
• Few-Shot Learning (FSL) 소개
  • one-shot imitation [147]
  • multi-armed bandits [33]
  • visual navigation [37]
  • continuous control [156]
• 학계에서 주목
  • meta-learning [37, 106, 114]
  • embedding learning [14, 126, 138]
  • generative modeling [34, 35, 113]

1.2 표기

• Task: $T$

• 데이터셋: $D = { D_\text{train}, D_\text{test}}$

• 훈련셋: $D_\text{train} = { (x_i, y_i )}_{i=1}^I$

• 시험셋: $D_\text{test} = { x^\text{test}}_{i=1}^I$

• Ground-truth 결합확률분포: $p(x, y)$

• $x$에서 $y$를 구하는 가설: $\hat h$

• $h(\cdot;\theta)$에 대한 가설 공간: $\mathcal H$
  • $h$의 모든 매개변수: $\theta$

• FSL 알고리즘은 $\theta$를 발견하여 $\mathcal H$를 찾기 위한 최적화 전략.

• 최선의 매개변수화: $h^\ast \in \mathcal H$

• 성능 측정을 위한 손실함수: $\mathcal l(\hat y, y)$

  • $\hat y = h(x;\theta)$

2.1 문제정의

Definition 2.1 (Machine Learning [92, 94] ). A computer program is said to learn from experience $E$ with respect to some classes of task $T$ and performance measure $P$ if its performance can improvewith $E$ on $T$ measured by $P$.

$E$: 경험 (experience) - 데이터

$T$: 작업 (task)

$P$: 성능 측정 (performance measure)

성능 $P$는 $T$에 기반한 $E$에 의해 향상되고 $P$에 의해 측정

Definition 2.2.

Few-Shot Learning(FSL) is a type of machine learning problems (specified by $E$, $T$ and $P$), where $E$ contains only a limited number of examples with supervised information for the target $T$.

FSL 예

• 영상 분류 [138]
• 단문 의미 분류 [157]
• 객체인식 [35]

Few-shot classification

• $h$: 분류기 (classfier)

각 입력 $x_i$에 대해 라벨 $y_i$를 예측

N-way-K-shot classification [37, 138]

• $D_\text{train}$ 은 $I = K N$을 포함.

• $N$: 클래스

• $K$: examples

• $I$: 샘플 수

Few-shot regression [37, 156]

• $h$: 회귀 추정 함수

• $y_i$ observed value of the dependent variable $y$

• $x_i$: observed value of independent variable $x$

Few-short reinforcement learning [3, 33]

finding a policy given only a few trajectories consisting of state-action pairs

3 typical scenarios of FSL

• Acting as a test bed for learning like human
• Learning for rare cases
• Reducing data gathering effort and computational cost

experience $E$: prior knowledge

One typical type of FSL methods: Bayesianlearning [35,76]

• It combines the provided training set $D_\text{train}$with some prior probability distribution which is available before $D_\text{train}$ is given.

One-shot learning: one example with supervised information in $E$, FSL called One-shot learning [14, 35, 138]

Zero-shot learningproblem (ZSL): $E$ does not contain any example with supervised information for the target $T$, FSL becomes azero-shot learningproblem.

2.2 Relevant Learning Problems

Weakly supervised learning [163] : learns from experience $E$ containing only weak supervision (such as incomplete, inexact, inaccurate or noisy supervised information)

• Semi-supervised learning [165]: learns from a small number of labeled samples and (usually a large number of) unlabeled samples in $E$. Example applications are text and webpage classification
  • Positive-unlabeled learning [81]: a special case of semi-supervisedlearning, in which only positive and unlabeled samples are given.
    • example, to recommend friends in social networks, we only know the users’ current friends according to the friend list, while their relationships to other people are unknown
• Active learning [117]: Elects informative unlabeled data to query an oracle for output $y$. This is usually used for applications where annotation labels are costly, such aspedestrian detection.

Weakly supervised learning with incomplete supervision: only a smallamount of samples have supervised information

• includes only classification and regression
  • while FSL also includes reinforcement learning problems
• mainly uses unlabeled data as additional information in $E$

Imbalanced learning [54]: learns from experienceEwith a skewed distribution for $y$

• some values of $y$ are rarely taken
• as in fraud detection and catastropheanticipation applications
• trains and tests to choose among all possible $y$
• FSL trains and tests forywith a few examples, while possibly taking the other $y$’s as prior knowledge for learning

Transfer learning [101]:

• transfers knowledge from the source domain/task, where training data is abundant, to the target domain/task
• used inapplications such as cross-domain recommendation
• Domain adaptation [11]: the source/target tasks are the same but the source/target domains are different
  • For example, in sentiment analysis, the source domain data contains customer comments on movies, while the target domain data contains customer comments on daily goods
• Transfer learning methodsare popularly used in FSL [7,82,85]
  • prior knowledge is transferred from thesource task to the few-shot task

Meta-learning [59]: improves $P$ of the new task $T$ by the provided data set and the meta-knowledge extracted across tasks by a meta-learner

• meta-learner gradually learns generic information (meta-knowledge) across tasks, and the learner generalizes the meta-learner for a new task $T$ using task-specific information
• learning optimizers [5, 80]
• dealing with the cold-start problemin collaborative filtering [137]
• guiding policies by natural language [25]
• meta-learner is taken as prior knowledge to guide each specific FSL task
• formal definition, using –> Appendix A 참조 TODO

Core Issue

the core issue of FSL based on error decomposition in supervised machine learning [17,18]

FSL supervised learning includingclassification and regression

provide insights for understanding FSL reinforcement learning

Empirical Risk Minimization [94, 136]

where:

• $\mathcal l$: 성능측정을 위한 손실 함수

• $h$: 가설 (hypothesis)

• $R$: we want to minimize its expected risk

• $p(x, y)$: 찾고자 하는 함수

empirical risk

\(R_I (h) = \frac{1}{I} \sum _{i=1}^I \mathcal l(h(x_i), y_i)\) sed as a proxy forR(h), leading toempirical risk minimization[94,136] (with possiblysome regularizers)

• $I$: 샘플 수

sed as a proxy forR(h), leading to empirical risk minimization [94,136] (with possibly some regularizers)

• $\hat h = \arg \min _h R(h)$: minimizes the expected risk

• $h^\ast = \arg \min _{h \in \mathcal H} R(h)$: minimizes the expected risk

• $h_I = \arg \min _{h \in \mathcal H} R_I(h)$: minimizes the empirical risk

Total error [17, 18]: $\mathbb E$ $$ \mathbb E[R(h_I) - R(\hat h)] = \mathbb E[R(h^\ast) - R(\hat h)]

• \mathbb E[R(h_I) - R(h^\ast)] $$

• $\mathbb E[R(h^\ast) - R(\hat h)]$: $\mathcal E _\text{app}(\mathcal H)$

• $\mathbb E[R(h_I) - R(h^\ast)]$: $\mathcal E _\text{est}(\mathcal H, I)$

나중에…

• one-shot imitation [147]
• multi-armed bandits [33]
• visual navigation [37]
• continuous control [156]
• meta-learning [37, 106, 114]
• Meta-learning [59]
• embedding learning [14, 126, 138]
• 영상 분류 [138]
• 단문 의미 분류 [157]
• 객체인식 [35]
• N-way-K-shot classification [37, 138]
• Few-shot regression [37, 156]
• Few-short reinforcement learning [3, 33]
• One typical type of FSL methods: Bayesianlearning [35, 76]
• One-shot learning [14, 35, 138]
• generative modeling [34, 35, 113]
• Weakly supervised learning [163]
• Positive-unlabeled learning [81]
• Active learning [117]
• Imbalanced learning [54]
• Transfer learning [101]
• Transfer learning methods are popularly used in FSL [7,82,85]
• Semi-supervised learning [165]
• learning optimizers [5, 80]
• Empirical Risk Minimization [94, 136]