Abstract

Supervised learning techniques construct predictive models by learning from a large number of training examples, where each training example has a label indicating its ground-truth output. Though current techniques have achieved great success, it is noteworthy that in many tasks it is difficult to get strong supervision information like fully ground-truth labels due to the high cost of data labeling process. Thus, it is desired for machine learning techniques to work with weak supervision. This article reviews some research progress of weakly supervised learning, focusing on three typical types of weak supervision: incomplete supervision where only a subset of training data are given with labels; inexact supervision where the training data are given with only coarse-grained labels; inaccurate supervision where the given labels are not always ground-truth.

摘要

监督学习是使用大量的训练样本训练的预测模型。其中每个训练样本被真实的打标。尽管现在的技术进步很大，仍然难以忽略在很多的任务中，难以训练出一个强的监督学习器。比如因为获取样本的标签代价高昂。所以机器学习需要在弱监督(weak supervision)上取得技术进步。本文回顾了弱监督的研究进展，并聚焦在三种典型的弱监督学习：不完全监督学习(incomplete supervision)是只有一部分子集给出标签；不精确监督学习(inexact supervision)是训练集样本只给出大概的标签；误监督学习(inaccurate supervision)是训练集样本不一定可信。

---

1. Introduction

Machine learning has achieved great success in various tasks, particularly in supervised learning tasks such as classification and regression. Typically, predictive models are learned from a training data set which contains a large amount of training examples, each corresponding to an event/object. A training example consists of two parts: a feature vector (or called instance) describing the event/object, and a label indicating the ground-truth output. In classification, the label indicates the class to which the training example belongs; in regression, the label is a real-value response corresponding to the example. Most successful techniques, such as deep learning [37], require ground-truth labels be given for a big training data set; in many tasks, however, it can be difficult to attain strong supervision information due to the high cost of data labeling process. Thus, it is desired for machine learning techniques to be able to work with weak supervision.

1.简介

机器学习在许多任务中取得成功，尤其在如分类和回归问题的监督学习中。通常，模型是从大量的训练样本中训练得到的，每个训练样本都是一个事件/事物(event/object)，一个标签表示一个真实的输出。在分类问题中，标签表示一个训练样本的归属类别；在回归问题中，一个标签表示一个样本的实际输出值。在多数的任务中，如深度学习[37]，需要大量的有真实标签的样本；但是在很多问题中，由于获取标签代价高昂，很难得到一个强的监督信息。因此，需要机器学习能够进行弱监督学习。

---

Typically, there are three types of weak supervision. The first is incomplete supervision, i.e., only a (usually small) subset of training data are given with labels whereas the other data remain unlabeled. Such situation occurs in various tasks. For example, in image categorization the ground-truth labels are given by human annotators; it is easy to get a huge number of images from the internet, whereas only a small subset of images can be annotated due to the human cost. The second type is inexact supervision, i.e., only coarse-grained labels are given. Consider the image categorization task again. It is desired to have every object in the images be annotated; however, usually we only have image-level labels rather than object-level labels. The third type isinaccurate supervision, i.e., the given labels are not always ground-truth. Such situation occurs, e.g., when the image annotator is careless or weary, or some images are difficult to be categorized.

通常，有三种弱监督学习。第一种是不完全监督学习(incomplete supervision)，如只有一小部分训练样本带有标签，而其他的样本没有标签。有很多任务会出现该情况，例如在图片分类中，真实的样本标签由人类标注。容易从互联网中获取有标注的图片，但是因为人为标注的代价，依然只有一小部分图片被标注。

---

Weakly supervised learning is an umbrella covering a variety of studies which attempt to construct predictive models by learning with weak supervision. In this article, we will introduce some progress about this line of research, focusing on learning with incomplete, inexact and inaccurate supervision. We will treat these types of weak supervision separately, but it is worth mentioning that in real practice they often occur simultaneously. For the simplicity, in this article we consider binary classification concerning two exchangeable classes Y and N. Formally, with strong supervision, the supervised learning task is to learn f : X → Y from a training data set D = {(x1,y1),...,(xm,ym)}, where X is the feature space, Y = {Y,N}, xi ∈ X, and yi ∈ Y. We assume that (xi, yi)’s are generated according to an unknown identical and independent distribution D; in other words, (xi, yi)’s are i.i.d. samples. Figure 1 provides an illustration of the three types of weak supervision we will discuss in this article.

弱监督学习通过多种方法手段，进行弱监督学习来训练模型。本文对三种弱监督方法的研究进展进行分别描述，需要注意的是这三种方法往往可以同时运用。简单化，本文考虑二分类问题，类别为Y，N。在强监督学习中，训练目的是从数据D = {(x1,y1),...,(xm,ym)}中训练一个模型 f: X → Y ，X是特征空间，Y = {Y,N}, xi ∈ X, yi ∈ Y。我们假设样本之间满足独立同分布。图1描述了三种弱监督学习。

图1 长条是特征空间，红绿mark是标签，“？”是样本可能不准确，那些小云朵表示某些场景可以用多种弱监督学习解决问题

---

2.Incomplete Supervision

Incomplete supervision concerns about the situation where we are given a small amount of labeled data, which is insufficient to train a good learner, while abundant unlabeled data are available. Formally, the task is to learn f : X → Y from a training dataset D ={(x1,y1),...,(xl, yl),xl+1,...,xm}, where there are l number of labeled training examples (i.e., those given with yi ) and u = m − l number of unlabeled instances; the other conditions are the same as that of supervised learning with strong supervision, as defined in the end of Section 1. For the convenience of discussion, we also call the l labeled examples as “labeled data” whereas the u unlabeled instances as “unlabeled data”.

不完全监督学习关注只有一小部分带有标签样本的场景。这些样本不足以训练一个好的模型。不完全监督学习的目的是从所有数据集D ={(x1,y1),...,(xl, yl),xl+1,...,xm}中学习f : X → Y ，其中 l 是带有标签的样本个数，而u=m-l是无标签样本。为了方便，我们称 l 个带有标签的样本为带标签样本('labeled data')，称 u 个无标签样本为无标签样本('unlabeled data')。

---

There are two major techniques for this purpose, i.e.,active learning [65] and semi-supervised learning [16, 97,102].

Active learning assumes that there is an “oracle”, such as a human expert, can be queried to get ground-truth labels for selected unlabeled instances. In contrast, semi-supervised learning attempts to automatically exploit unlabeled data in addition to labeled data to improve learning performance, where no human intervention is assumed.There is a special kind of semi-supervised learning called transductive learning whose main difference with (pure) semi-supervised learning lies in their different assumptions about test data, i.e., data to be predicted by the trained model. Transductive learning holds a “close-world” assumption, i.e., the test data are given in advance and the goal is to optimize performance on the test data; in other words, the unlabeled data are exactly test data. Pure semi-supervised learning holds an “open-world” assumption, i.e., the test data are unknown and the unlabeled data are not necessarily to be test data. Figure 2 intuitively shows the difference between active learning, (pure) semi-supervised learning and transductive learning.

有两种主流的方法，一种是active learning[65]，一种是semi-supervised learning[16, 97, 102]。

active learning假设有一种“oracle”，例如专家经验，用来查询未标注样本的真实标签。相反，semi-supervised learning试图从带标签数据中自动学习未标注样本的标签来改善模型表现，而不需要人为介入。有一种特别的semi-supervised learning称为transductive learning，它与纯(pure)半监督学习的不同在于二者对于测试集的假设。transductive learning假设了一个'close-world'，测试集已经事先给定，未标注样本就是测试集，目的是改善测试集上的表现。而pure semi-supervised learning假设‘open-world’，测试集未知并且未标注数据也未必就是测试集。图2展示了active learning，  (pure) semi-supervised learning，以及transductive learning的区别。