Deep Active Ensemble Sampling For Image Classification Salman Mohamadi Gianfranco Doretto Donald A. Adjeroh

2025-04-26 0 0 1.13MB 17 页 10玖币

侵权投诉

Deep Active Ensemble Sampling For Image

Classiﬁcation

Salman Mohamadi, Gianfranco Doretto, Donald A. Adjeroh

West Virginia University, Morgantown, WV, USA

Abstract. Conventional active learning (AL) frameworks aim to reduce

the cost of data annotation by actively requesting the labeling for the

most informative data points. However, introducing AL to data hun-

gry deep learning algorithms has been a challenge. Some proposed ap-

proaches include uncertainty-based techniques, geometric methods, im-

plicit combination of uncertainty-based and geometric approaches, and

more recently, frameworks based on semi/self supervised techniques. In

this paper, we address two speciﬁc problems in this area. The ﬁrst is the

need for eﬃcient exploitation/exploration trade-oﬀ in sample selection in

AL. For this, we present an innovative integration of recent progress in

both uncertainty-based and geometric frameworks to enable an eﬃcient

exploration/exploitation trade-oﬀ in sample selection strategy. To this

end, we build on a computationally eﬃcient approximate of Thompson

sampling with key changes as a posterior estimator for uncertainty rep-

resentation. Our framework provides two advantages: (1) accurate poste-

rior estimation, and (2) tune-able trade-oﬀ between computational over-

head and higher accuracy. The second problem is the need for improved

training protocols in deep AL. For this, we use ideas from semi/self super-

vised learning to propose a general approach that is independent of the

speciﬁc AL technique being used. Taken these together, our framework

shows a signiﬁcant improvement over the state-of-the-art, with results

that are comparable to the performance of supervised-learning under

the same setting. We show empirical results of our framework, and com-

parative performance with the state-of-the-art on four datasets, namely,

MNIST, CIFAR10, CIFAR100 and ImageNet to establish a new baseline

in two diﬀerent settings.

1 Introduction

Active learning (AL) has consistently played a central role in domains where

labeling cost is of great concern. The core idea of AL frameworks revolves around

learning from small amounts of annotated data and sequentially choosing the

most informative data sample or batch of data samples to label. To this end, after

initial training using available labeled data, an acquisition function is utilized

to leverage the model’s uncertainty in order to explore the pool of unlabeled

data for most informative data points. In parallel with advancements in AL, in

the recent years, deep learning has gained tremendous attention with diverse

applications such as realistic image generation, object detection and tracking,

arXiv:2210.05770v1 [cs.CV] 11 Oct 2022

2 S. Mohamadi et al.

semantic segmentation, and iris recognition [1,2,3] due to its emergence as a high-

performing approach, primarily conditioned on the availability of large amounts

of training data. An interesting challenge is how to eﬃciently incorporate data-

hungry deep learning tools into supposedly data-eﬃcient AL frameworks.

Adjusting AL algorithms for deep neural networks has been very challeng-

ing, where extending the model complexity/capacity to that of CNNs ultimately

ended up with either a poor performance, or some minor improvements at the

cost of querying almost all samples. On the other hand, sequential training of

such expressive models as well as extending the framework to high dimensional

data injects even more complexity [4,5,6]. This challenge was relatively under-

explored, until a breakthrough work by Gal et al [7], which essentially considered

the problem of incorporating deep learning into AL for high dimensional data

as highly connected with that of uncertainty representation [8]. They thus ap-

proached the problem from the perspective of uncertainty representation in deep

learning for AL, and developed a Bayesian AL framework for image data. Later

work (such as [9]), however, argued that the approach exhibits poor scalability

to big datasets due to its limited model capacity .

Another approach that also relied on uncertainty representation, is ensemble-

based AL [9]. Here, an ensemble of classiﬁers is used, where the classiﬁers in-

dependently learn from the data in parallel. The major drawback is the poor

diversity (lack of exploration) even with larger ensembles. Our approach, while

enjoying the power of ensembles, solves this problem by oﬀering an inherent ex-

ploration/exploitation trade-oﬀ as classiﬁers maintain some dependency in the

form of a shared prior. Apart from uncertainty representation, another set of

emerging methods that primarily rely on geometrical data representation [10]

showed improved performance in deep AL. However, similar to [11], we empiri-

cally observed that these geometric approaches typically suﬀer from performance

degradation as the class diversity (number of classes) increases. Another recent

approach is the work reported in [11] where they take advantage of adversarial

training to provide improved performance over previous methods. We empirically

ﬁnd that their work provided a balanced performance on datasets at diﬀerent

scales and diversity. As we will show later, our proposed model outperforms this

approach in multiple settings with signiﬁcant margins, with results approaching

that of supervised learning models in some cases.

In the ﬁrst part of the paper, primarily motivated to eﬃciently integrate

the advantages of uncertainty and geometrical representations, we propose an

approach built upon approximate Thompson sampling. On one hand, this pro-

vides an improved representation of uncertainty over unlabeled data, and on the

other hand, supports an inherent tune-able exploration/exploitation trade-oﬀ

for diverse sampling [12,13]. Unlike conventional ensemble-based methods whose

performance tend to saturate quickly, under our tuneable model, adding a few

more classiﬁers tends to improve the uncertainty and geometric representation.

To mitigate the general sample diversity problem of ensemble models (see [14,9]

), we use an inclusive sample selection strategy. Our framework showed a no-

ticeable improvement over the state-of-the-art, with performance approaching

DAES 3

those of supervised learning methods. Further, we explore the scope and scale

of model eﬃciency improvements brought about by our proposed techniques.

Brieﬂy, due to the exploration/exploitation trade-oﬀ, Thompson sampling is

expected to improve both predictive uncertainty and sample diversity by com-

puting, sampling, and updating a posterior distribution. A serious consideration,

however, is that, for more expressive models such as deep convolutional neural

networks (CNNs) designed for high dimensional data, Thompson sampling makes

the process computationally diﬃcult. This is primarily because computation of

the posterior distribution over CNNs is complex by nature. Inferences based

on Laplace approximations or Markov chain Monte Carlo approaches would

be two possible alternatives. However, both approaches are still very expen-

sive in terms of computational cost [15,16,17]. Lu et al [17] argue that due to

the compatibility of Thompson sampling with sequential decision and updating,

an approximate version of Thompson sampling could be a promising solution.

Accordingly, we build an ensemble model relying on an eﬃcient approximate

of Thompson sampling, which improves the state-of-the-art. Interestingly, this

model possesses both the advantage of uncertainty based deep AL approaches

(exploiting most uncertain samples), and of geometric solutions (exploring for

more diverse though not necessarily highly uncertain samples).

In the second part of the paper, we investigate a new line of eﬀorts/arguments

revolving around the idea of boosting AL frameworks using self/semi supervised

learning techniques. We substantiate and unify these arguments and also design

and perform extensive experiments on multiple baselines to assess this approach

as a new general training protocol for AL frameworks. This enables our approach

to be compared against recent boosted AL frameworks.

Brieﬂy, our key contributions in this paper are as follows:

–A new framework for deep AL which enables an exploration/exploitation

trade-oﬀ for sample selection and hence oﬀers the advantages of both uncertainty-

based and geometry-based methods.

–A new general training protocol for visual AL approaches, developed by

substantiating and unifying recent arguments on boosting AL using self/semi

supervised learning, and experimentally evaluating this approach on multiple

recent baselines. We compare our framework against two sets of baselines to

show its performance.

2 Background and preliminaries

Background: Early eﬀorts on AL with image data considered mainly kernel-

based approaches [18,19,20]. Later, AL methods with image data using CNN in-

cluded uncertainty-based approaches [7,21,9,22,23], geometry-based approaches

[10], or their combination [11], e.g, based on adversarial training. Generally

speaking, uncertainty-based approaches focus on ﬁnding most uncertain sam-

ples to label, with the potential downside of less diversity in sample selection,

while geometric approaches tend to weigh on diversity of samples, resulting in

performance degradation in cases of very diverse datasets (with large number of

4 S. Mohamadi et al.

classes). Most recently, in a relatively diﬀerent setting, Gao et al. [24] leveraged

semi-supervised learning while Bengar et al. [25] applied self-supervised learn-

ing (SSL) techniques to deliver a signiﬁcant performance improvement. We will

compare our proposed approach against these related work, on the same problem

settings. Some other recent work in this general area of modern AL with high

dimensional data can be found in [26,27,28,29,30,31]. Though these are relevant,

they are not as closely related to our approach.

SSL: As the second contribution of this work relates to SSL we brieﬂy review

the literature. Brieﬂy, SSL is one of the closest modern problem domains to AL

with zero labeling eﬀort policy. Here, the goal is to leverage all unlabeled data to

train a network for a pretext task so as to prepare the network for a downstream

task, usually with small amounts of data [32]. Until recently, a major set of SSL

baselines were contrastive baselines relying on contrasting augmented views of a

sample with each other (positive contrastive pairs) and with views of other sam-

ples (negative contrastive pairs) [33,34]. Newer baselines such as [35,36], a.k.a

non-contrastive approaches, rely on contrasting positive pairs, needless of con-

trasting negative pairs. Recently, Ermolov et al. [37] reported a non-contrastive

method based on whitening the embedding space, which was eﬀective, yet con-

ceptually simple. We adopt this approach in this work.

Preliminary: We describe these two major paradigms below.

1. Uncertainty-based techniques: Two categories of well-known deep learn-

ing techniques for uncertainty representation and estimation include ensemble-

based techniques (non-Bayesian) [22,23] and Monte-Carlo (MC) dropout (Bayesian)

[21,7]. In ensemble-based methods, an ensemble of Nidentically structured neu-

ral networks are trained using identical training data Dtr, where the diﬀerent

random values are applied for weight initialization wi. For a given class cout of

multiple classes and input X, we then have:

p(y=c|x, Dtr) = 1

i=N

i=1

p(y=c|x, wi)(1)

However, MC-dropout trains a network with dropout, and during test, imple-

ments Tforward passes, each individually with a new dropout mask, resulting

in Tsets of weights wt. Given input x, the average of all Tsoftmax vectors

represents the output for a desired class c.

p(y=c|x, Dtr) = 1

t=T

t=1

p(y=c|x, wt)(2)

Here we brieﬂy describe some popular eﬀective uncertainty-based acquisition

functions [7,9] or their approximation for ensemble-based approaches, MC dropout,

and our proposed framework, all based on uncertainty sampling.

A. Selecting samples with highest predictive entropy [38].

H[y|x, Dtr] := −X

p(y=c|x, wn)).log( 1

p(y=c|x, wn)) (3)

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

DeepActiveEnsembleSamplingForImageClassificationSalmanMohamadi,GianfrancoDoretto,DonaldA.AdjerohWestVirginiaUniversity,Morgantown,WV,USAAbstract.Conventionalactivelearning(AL)frameworksaimtoreducethecostofdataannotationbyactivelyrequestingthelabelingforthemostinformativedatapoints.However,introducin...

展开>> 收起<<

Deep Active Ensemble Sampling For Image Classification Salman Mohamadi Gianfranco Doretto Donald A. Adjeroh.pdf

共17页,预览4页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Deep Active Ensemble Sampling For Image Classification Salman Mohamadi Gianfranco Doretto Donald A. Adjeroh

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: