MI-ELM: Highly efficient multi-instance learning based on hierarchical
extreme learning machine
Qiang Liu
a,
n
, Sihang Zhou
a
, Chengzhang Zhu
a
, Xinwang Liu
a
, Jianping Yin
b
a
School of Computer, National University of Defense Technology, Changsha 410073, China
b
State Key Laboratory of High Performance Computing, National University of Defense Technology, Changsha 410073, China
article info
Article history:
Received 7 May 2015
Received in revised form
28 July 2015
Accepted 21 August 2015
Communicated by G.-B. Huang
Available online 6 September 2015
Keywords:
Multi-instance learning
Hierarchical extreme learning machine
Optimization
Double-hidden layers feedforward network
abstract
Multi-instance learning (MIL) is one of promising paradigms in the supervised learning aiming to handle
real world classi fication problems where a classification target contains several featured sections, e.g., an
image typically contains several salient regions. In this paper, we propose a highly efficient learning
method for MI classification based on hierarchical extreme learning machine (ELM), called MI-ELM.
Specifically, a double-hidden layers feedforward network (DLFN) is designed to serve as the MI classifier.
Then, the MI classification is formulated as an optimization problem. Moreover, the output weights of
DLFN can be analytically determined by solving the aforementioned optimization problem. The merits of
MI-ELM are as follows: (i) MI-ELM extends the single-layer ELM to be a hierarchical one that well fits for
training DLFNs in MI classification. (ii) The input and hidden-layer parameters of DLFNs are assigned
randomly rather than tuned iteratively, and the output weights of DLFNs can be determined analytically
in one step. Therefore, MI-ELM significantly enhances the efficiency of the DLFN without notable loss of
the classification accuracy. Experimental results over several real-world data sets demonstrate that the
proposed MI-ELM method significantly outperforms existing kernel methods for MI classi fication in
terms of the classification accuracy and the classification time.
& 2015 Elsevier B.V. All rights reserved.
1. Introduction
Nowadays, real world applications become much more com-
plex, especially in the era of big data. For example, biometric
recognition in bioinformatics, automatic target recognition in
public safety, Landmark Recognition [1,2], anomaly detection in
mobile Internet. Typically, a common characteristic of these
applications is that the target to be classified comprises of multiple
featured sections, e.g., a target in the automatic target recognition
contains several salient regions. However, a key challenge of these
classification problems is that the specific labels of featured sec-
tions are partially known for use, while the knowledge of category
labels is available. Hence, it is vital to address the challenge made
by the label missing with respect to some sections of a target.
Suppose that each training category with known label contains
one or more instances, the learner has incomplete information
about the exact labels of all training instances, but instead, the
learner only knows the category label of these instances, resulting
in the multi-instance (MI) learning problem. MI learning, which is
first introduced in [3], is the most promising supervised learning
technology to solve the above problem. In MI classification pro-
blems, a category and its label are termed as the bag and the bag
label, respectively. It is worthwhile to mention that only bag labels
are assigned in priori for the goal of classification, while instance
labels are partially unknown. The typical framework of supervised
MI classi fication is illustrated in Fig. 1, where each bag contains a
collection of instances.
In order to solve the MI learning problem, some kernel meth-
ods have been proposed to design MI classi fiers with considerable
performance [4 –6]. Considering that these learning methods were
concentrated on improving the predictive performance in terms of
the classification accuracy, the authors in [7] further proposed a
direct method to learn sparse kernel classifiers for MI classification
by investigating both the accuracy and the efficiency. However,
these kernel methods are iterative MI learning methods that tune
parameters of MI classifiers iteratively, inducing much longer
training time. Thus, it is intuitive to find another non-iterative
method to significantly reduce the training time while maintain-
ing or even improving the predictive accuracy. In 2006, Huang
et al. [8] first proposed a rigorous learning algorithm called
extreme learning machine (ELM) for single-hidden layer feedfor-
ward neural networks (SLFNs). The notable merits of ELM include
two aspects: (i) The input weights and hidden layer biases of
Contents lists available at ScienceDirect
journal homepage: www.elsevier.com/locate/neucom
Neurocomputing
http://dx.doi.org/10.1016/j.neucom.2015.08.061
0925-2312/& 2015 Elsevier B.V. All rights reserved.
n
Corresponding author.
E-mail addresses: libra6032009@gmail.com, qiangl.ne@hotmail.com (Q. Liu),
306114653@qq.com (S. Zhou), kevin.zhu.china@gmail.com (C. Zhu),
xinwangliu@nudt.edu.cn (X. Liu), jpyin@nudt.edu.cn (J. Yin).
Neurocomputing 173 (2016) 1044–1053