Conference PaperPDF Available

Performance Analysis of Incremental Learning Strategy in Image Classification

Authors:

Abstract and Figures

Deep learning-based image classification model learns from the fixed and specific training dataset. For the generalization and adaptation of human learning behaviour, some of these models adapted incremental learning to enhance the learning and knowledge from updated and incremented dataset. An incremented dataset can be in form of increment of examples or new class dataset images. This incremented dataset is learned by deep learning models by two incremental learning strategies i.e. sample-wise and class-wise. This paper proposes a performance analysis methodology and experimentally analyze the performance of these incremental learning strategies in CNN based image classification model on prepared incremented dataset. The evaluation of performance of these deep image classification model on classification performance metrics such as accuracy, precision, recall and f1 score shows that these model’s learning and classification capabilities are increased with incremental learning on the incremented dataset. Two different incremental learning shows the relation between performances of model to the increment of dataset.
Content may be subject to copyright.
XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE
Performance Analysis of Incremental Learning
Strategy in Image Classification
Gaurav Jaiswal
ICT Research Lab, Department of Computer Science
University of Lucknow
Lucknow, India
Email: gauravjais88@gmail.com
Abstract—Deep learning-based image classification model
learns from the fixed and specific training dataset. For the
generalization and adaptation of human learning behaviour,
some of these models adapted incremental learning to enhance
the learning and knowledge from updated and incremented
dataset. An incremented dataset can be in form of increment of
examples or new class dataset images. This incremented
dataset is learned by deep learning models by two incremental
learning strategies i.e. sample-wise and class-wise. This paper
proposes a performance analysis methodology and
experimentally analyze the performance of these incremental
learning strategies in CNN based image classification model on
prepared incremented dataset. The evaluation of performance
of these deep image classification model on classification
performance metrics such as accuracy, precision, recall and f1
score shows that these model’s learning and classification
capabilities are increased with incremental learning on the
incremented dataset. Two different incremental learning shows
the relation between performances of model to the increment
of dataset.
Keywords—image classification, incremental learning, CNN,
incremental image dataset, deep learning model, performance
analysis.
I. INTRODUCTION
Image classification is the core problem in computer
vision. With the advent of deep learning and advancement in
high computation, image classification is widely applied in
various domain of real-world applications [1] [2]. Deep
supervised Image classification models learn from training
input images which are fixed and specific to the task. Some
of these deep models outperform and surpass the level of
learning and classification of the human being in various
specific vision tasks [3].
In the present scenario, daily new data, images are
generated and updated in every domain. To perform in this
scenario, the image classification model is also needed to
retrain and updated with these new datasets. For the
generalization of these model in the real-world learning
environment, it should be able to learn incrementally to be
updated as the real-world data are updated with time. As a
human being, it should increase their learning base for new
samples or new class data and able to handle new updated
knowledge to the classification of images. These challenging
tasks increase the interests of researchers. Various scientist
and researchers proposed and developed various learning and
transfer strategies for new knowledge and data to update and
increment the learning of image classification model such as
Transfer learning [4], online learning [5], incremental
learning [6], never-ending learning [7] batch learning [8] and
new domain adaptation strategy [9]. These type of learning
also provide a flexible way to input image data into the
incremental manner
Incremental learning is one of these learning strategies in
which input data continuously update the knowledge base of
the existing learned model, the process of this learning is
shown in Fig. 1. For stream data, He et al. [10] proposed an
adaptive incremental learning framework which enable
knowledge collection and transformation to benefit learning
from continuous data stream. Losing et al. [11] presented the
review and comparison of state of arts incremental learning
methods and found that SVM based incremental methods
perform better. With the advent of deep learning methods,
learning and classification task achieve better performance.
Incremental leaning in deep learning model also enable these
model to learn for new domain knowledge. Rosenfeld et al.
[12] proposed deep adaptation modules method which
preserve the previous knowledge and constrains newly
learned feature to be linear combination of existing ones.
Fine tuning deep models for class-increment learning leads
to catastrophic forgetting problem. Zhang et al. [13]
proposed deep model consolidation approach for class-
incremental learning for deep learning model and handled
this forgetting issues. Various researchers have analyzed the
performance of image classification model on different
parameters i.e. image quality parameters such as blur, noise,
contrast, image compression [14], resolution [15], model
learning parameters [16]. Performance analysis on different
incremental strategies sample-wise and class-wise increment
have not been explored.
Fig. 1 Process of Incremental Learning
The main aim of this paper is to analyze the performance
of two incremental learning strategies (increment of image
samples and increment of images of new class) in deep
learning image classification model (CNN: convolution
neural network). This analysis can also be done on other type
of data like numerical or temporal, but here, we mainly focus
on image data. For the analysis of performance on these two
learning strategies, a performance analysis methodology is
proposed. In this methodology, the performance score of
image classification model is measured by four standard
classification performance metrics such as accuracy,
precision, recall and f1 score. Based on these performance
scores, analysis is performed. The performance analysis of
incremental learning of image classification model shows
that the knowledgebase of image classification model is
increased with the increment of image dataset. Two different
incremental learning strategies shows the relation between
performances of model to the increment of dataset. The
performance is decreased for class-wise increment of dataset
while performance is increased for sample-wise increment of
dataset. The remaining paper is organized as follow: Section
II presents the steps of performance analysis methodology
for performance on incremental learning strategies. In
Section III, experimental setup and experimental results are
described with performance scores, further analysis is done
on these results with the comparison graph of performance
analysis. Finally, Section IV concludes the paper.
II. PERFORMANCE ANALYSIS METHODOLOGY
The proposed performance analysis methodology
constitutes of three submodules (i.e. image dataset, image
classification model, and performance evaluation score) and
two processes (i.e. Incremental learning and performance
evaluation). The block diagram of this methodology is
shown in Fig. 2. Incremental learning process connects the
image dataset submodule and image classification
submodule while performance evaluation process connects
the image classification submodule and performance
submodule.
A. Image Dataset
Image dataset contains the set of images of various
classes in which we want to train our image classification
model. For incremental learning environment, these image
datasets are rearranged in such a way that it provides the
continuous increasing dataset to model with the help of
incremental learning strategies. For this, these datasets are
divided and rearranged in two distribution manner such as
sample-wise distribution [17] and class-wise distribution
[18]. These distributed set of image dataset can be used in
incremental order to input the image classification model
using incremental learning strategies. Dataset preparation of
sample-wise increment dataset is shown in Fig. 3.
Fig. 3 Sample/ example wise increment in dataset
Fig. 2 Performance analysis methodology for incremental strategy in image classification
Dataset preparation of class-wise increment dataset is
shown in Fig. 4.
B. Incremental Learning Process
In performance analysis methodology, Incremental
learning is the process between image dataset and image
classification model submodule. This process provides an
algorithmic way to input data to the image classification
model and keeps updated the model learning knowledge
base for new data. The input of data can be in sample-wise
or class-wise increment.
C. Image Classification Model
Image classification model is trained incrementally with
each set of dataset. This model is updated continuously with
new increment dataset. After training of model, these models
are evaluated on test dataset. Test dataset contains increment
test data and previous dataset images. For the sake of input
and output constraint, we have fixed the input and output
node of model whether it is trained on sample-wise
increment dataset or class-wise increment dataset.
D. Performance Evaluation Process
Performance evaluation is the process between image
classification model and performance evaluation score.
Trained image classification model is evaluated on actual test
data class value to predicted test data class value. These
values are compared on performance metrics such as
accuracy, precision, recall and f1 score [1].
E. Performance Evaluation Score
After the performance evaluation process, performance
evaluation score is obtained. It contains all performance
scores (accuracy, precision, recall and f1 score) for all
increment image dataset on both incremental learning
strategies.
III. EXPERIMENTS AND ANALYSIS
A. Experimental Setup
For performing the experiments based on discussed
performance analysis methodology, first, we prepared the
incremental image dataset for incremental learning of image
classification model. For this, we selected the three standard
image database i.e. MNIST [19], Fashion-MNIST [20] and
CIFAR10 [21]. These image datasets contain the 10
different classes of image. These datasets are divided in two
way to create example wise increment or class wise
increment. The experimental configuration of these divided
datasets is given in Table I.
TABLE I DATA SET CONFIGURATION
Original Image
Dataset
Incremental
Dataset
Increment of Dataset
Class wise
increment
Example wise
increment
MNIST [19]
M-D1 Cl
ass
0-4 50% images
M-D2 Cl
ass
5-6 20% images
M-D3 Cl
ass
7-8 20% images
M-D4 Cl
ass
9 10 % images
Fashion_MNIST
[20]
FM-D1 Cl
ass
0-4 50% images
FM-D2 Cl
ass
5-6 20% images
FM-D3 Cl
ass
7-8 20% images
FM-D4 Cl
ass
9 10 % images
CIFAR10 [21]
C-D1 Cl
ass
0-4 50% images
C-D2 Cl
ass
5-6 20% images
C-D3 Cl
ass
7-8 20% images
C-D4 Cl
ass
9 10 % images
Fig. 4 Class wise increment of dataset
This performance methodology is implemented in python
based library keras and sk-learn. The architecture of CNN
based image classification model is based on CNN
architecture of Alexnet which is proposed by Krizehvasky et
al. [22]. The layer-wise architectural details of CNN models
for each dataset are shown in Table II.
TABLE II LAYER-WISE ARCHITECTURE DETAILS OF CNN MODEL
CNN Architecture for MNIST, Fashion-MNIST
Layers Layers Parameter Activation Function
Conv2D
Conv2D
Maxpooling2D
32, size=(3,3)
32,size=(3,3)
Size=(2,2)
Relu
Relu
Conv2D
Conv2D
Maxpooling2D
64,size=(3,3)
64,size=(3,3)
Size=(2,2)
Relu
Relu
Dense
Dropout
512
0.2
Relu
Dense 10 Softmax
CNN Architecture for CIFAR10
Layers Layers Parameter Activation Function
Conv2D
Conv2D
Conv2D
Maxpooling2D
Dropout
32,size=(3,3)
32,size=(3,3)
32,size=(3,3)
Size=(2,2)
0.25
Relu
Relu
Relu
Conv2D
Conv2D
Conv2D
Maxpooling2D
Dropout
64,size=(3,3)
64,size=(3,3)
64,size=(3,3)
Size=(2,2)
0.25
Relu
Relu
Relu
Dense
Dropout
512
0.5
Relu
Dense 10 Softmax
The implemented CNN model is trained on training
incremental dataset and tested on all previously combined
test dataset to check its new knowledge learning as well as
old knowledge learning.
For performance evaluation of these CNN models on
modified incremental image dataset, four standard
classification performance metrics such as accuracy,
precision, recall and f1 score are selected. These
performance metrics tell the model’s correctness,
completeness, compactness.
B. Experimental Results
After performing training and testing on each incremental
dataset, these trained image classification model are
evaluated on predicted value and actual value of test dataset
class on selected performance metrics. The obtained
experimental results and performance scores are tabulated in
Table III and Table IV.
Table III shows the performance scores of image
classification model on sample-wise incremental learning of
MNIST, Fashion-MNIST and CIFAR-10 dataset.
TABLE III PERFORMANCE SCORE ON SAMPLE WISE INCREMENT DATASET
Original
Image
Dataset
Incremental
Dataset
Performance score
(sample
-
wise increment)
accuracy precision recall
f1
score
MNIST
[19]
M
-
D1
0.9866
0.98
66
0.98
67
0.98
66
M
-
D2
0.9918
0.9917
0.9918
0.9917
M
-
D3
0.9926
0.9926
0.9926
0.9926
M
-
D4
0.9930
0.9931
0.9929
0.9930
Fashion_
MNIST
[20]
FM
-
D1
0.9076
0.9090
0.9078
0.9082
FM
-
D2
0.9122
0.9127
0.9124
0.9124
FM
-
D3
0.9146
0.9146
0.9145
0.9145
FM
-
D4
0.9151
0.9153
0.9148
0.9150
CIFAR10
[21]
C
-
D1
0.6570
0.6567
0.6538
0.6
545
C
-
D2
0.7021
0.7099
0.7028
0.70
42
C
-
D3
0.7247
0.7321
0.7252
0.7243
C
-
D4
0.7365
0.7371
0.7361
0.7365
Table IV shows the performance scores of image
classification model on class-wise incremental learning of
MNIST, Fashion-MNIST and CIFAR10 dataset.
TABLE IV PERFORMANCE SCORE ON CLASS-WISE INCREMENT DATASET
Original
Image
Dataset
Incremental
Dataset
Performance score
(class-wise increment)
accuracy precision recall
f1
score
MNIST
[19]
M
-
D1
0.9986
0.9986
0.9986
0.9986
M
-
D2
0.9954
0.9955
0.9953
0.9954
M
-
D3
0.9939
0.9938
0.9939
0.9938
M
-
D4
0.9930
0.9931
0.9929
0.9930
Fashion_
MNIST
[20]
FM
-
D1
0.9426
0.9428
0.9426
0.9427
FM
-
D2
0.9236
0.9237
0.9236
0.9237
FM
-
D3
0
.
9167
0.9161
0.9167
0.9165
FM
-
D4
0.9151
0.9153
0.9148
0.9150
CIFAR10
[21]
C
-
D1
0.7648
0.7659
0.7643
0.7651
C
-
D2
0.7465
0.7462
0.7467
0.7464
C
-
D3
0.7394
0.7385
0.7398
0.7391
C
-
D4
0.7365
0.7371
0.7361
0.7365
The experimental results of these experiments show that
the increment of dataset and their learning increase the
knowledge base of image classification model.
0.982
0.984
0.986
0.988
0.99
0.992
0.994
accuracy precision recall f1 score
performance score
MNIST dataset
MNIST M-D1 MNIST M-D2
MNIST M-D3 MNIST M-D4
0.902
0.904
0.906
0.908
0.91
0.912
0.914
0.916
accuracy precision recall f1 score
performance score
Fashion-MNIST dataset
Fashion_MNIST FM-D1 Fashion_MNIST FM-D2
Fashion_MNIST FM-D3 Fashion_MNIST FM-D4
0.6
0.62
0.64
0.66
0.68
0.7
0.72
0.74
0.76
accuracy precision recall f1 score
performance score
CIFAR10 dataset
CIFAR10 C-D1 CIFAR10 C-D2
CIFAR10 C-D3 CIFAR10 C-D4
Fig. 5 Comparison graph of performances for sample-wise incremental learning
0.99
0.991
0.992
0.993
0.994
0.995
0.996
0.997
0.998
0.999
1
accuracy precision recall f1 score
performance score
MNIST dataset
MNIST M-D1 MNIST M-D2
MNIST M-D3 MNIST M-D4
0.9
0.905
0.91
0.915
0.92
0.925
0.93
0.935
0.94
0.945
accuracy precision recall f1 score
performance score
Fashion-MNIST dataset
Fashion_MNIST FM-D1 Fashion_MNIST FM-D2
Fashion_MNIST FM-D3 Fashion_MNIST FM-D4
0.72
0.725
0.73
0.735
0.74
0.745
0.75
0.755
0.76
0.765
0.77
accuracy precision recall f1 score
performance score
CIFAR10 dataset
CIFAR10 C-D1 CIFAR10 C-D2
CIFAR10 C-D3 CIFAR10 C-D4
Fig. 6 Comparison graph of performances for class-wise incremental learning
C. Performance Analysis
Now, comparison graphs of these performance scores are
generated for detail performance analysis of the incremental
learning on CNN based image classification model for
MNIST, Fashion-MNIST, and CIFAR10 datasets. The
performance comparison graph of sample-wise incremental
learning and class-wise incremental learning for three
datasets is shown in Fig. 5 and Fig. 6 respectively.
During the analysis of the performance comparison graph
of the classification model for both dataset, it is noticeable
that the knowledge base of learning model is increased. So it
can be able to perform classification task on new
incremented data also. It successfully learned and classified
the new class or sample data.
Further, we also analyzed the relation between
incremental learning strategies to the performance of image
classification model. We found that increment data in
sample-wise, it increased the performance as it increases the
better understanding and learning of known fact or image
data. While, increment in class-wise, decrease the
performance of image classification model. The increment in
class-wise explore the new knowledge and for understanding
and learning on these new data, model has to change and
adapt this by updating the model.
IV. CONCLUSION
For the generalization and adaptation of human learning
behaviour, image classification model adapted incremental
learning to enhance the learning and knowledge from
updated and incremented dataset. An incremented dataset
can be in form of increment of examples or new class dataset
images. This incremented dataset is learned by deep learning
models by two incremental learning strategies. This paper
proposed a performance analysis methodology for analyzing
the performance of incremental learning in image
classification. The performance analysis of incremental
learning of image classification model shows that the
knowledgebase of image classification model is increased
with the increment of image dataset. Two different
incremental learning strategies shows the relation between
performances of model to the increment of dataset. The
performance is increased for sample-wise increment of
dataset while performance is decreased for class-wise
increment of dataset. The performance analysis of two
learning strategies can be very useful to the researchers for
further exploration of this learning areas.
REFERENCES
[1] I. Goodfellow, Y. Bengio, A. Courville and Y. Bengio, Deep learning,
vol. 1, MIT press Cambridge, 2016.
[2] N. Ketkar and E. Santana, Deep Learning with Python, vol. 1,
Springer, 2017.
[3] K. He, X. Zhang, S. Ren and J. Sun, "Delving deep into rectifiers:
Surpassing human-level performance on imagenet classification," in
Proceedings of the IEEE international conference on computer
vision, 2015.
[4] K. Weiss, T. M. Khoshgoftaar and D. Wang, "A survey of transfer
learning," Journal of Big data, vol. 3, p. 9, 2016.
[5] J. Kivinen, A. J. Smola and R. C. Williamson, "Online learning with
kernels," IEEE transactions on signal processing, vol. 52, p. 2165–
2176, 2004.
[6] Y. Wu, Y. Chen, L. Wang, Y. Ye, Z. Liu, Y. Guo and Y. Fu, "Large
scale incremental learning," in Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition, 2019.
[7] T. Mitchell, W. Cohen, E. Hruschka, P. Talukdar, B. Yang, J.
Betteridge, A. Carlson, B. Dalvi, M. Gardner, B. Kisiel and others,
"Never-ending learning," Communications of the ACM, vol. 61, p.
103–115, 2018.
[8] G. I. Parisi, R. Kemker, J. L. Part, C. Kanan and S. Wermter,
"Continual lifelong learning with neural networks: A review," Neural
Networks, vol. 113, p. 54–71, 2019.
[9] M. Wang and W. Deng, "Deep visual domain adaptation: A survey,"
Neurocomputing, vol. 312, p. 135–153, 2018
[10] H. He, S. Chen, K. Li and X. Xu, "Incremental learning from stream
data," IEEE Transactions on Neural Networks, vol. 22, p. 1901–1914,
2011.
[11] V. Losing, B. Hammer and H. Wersing, "Incremental on-line
learning: A review and comparison of state of the art algorithms,"
Neurocomputing, vol. 275, p. 1261–1274, 2018.
[12] A. Rosenfeld and J. K. Tsotsos, "Incremental learning through deep
adaptation," IEEE transactions on pattern analysis and machine
intelligence, 2018.
[13] J. Zhang, J. Zhang, S. Ghosh, D. Li, S. Tasci, L. Heck, H. Zhang and
C.-C. J. Kuo, "Class-incremental learning via deep model
consolidation," in The IEEE Winter Conference on Applications of
Computer Vision, 2020.
[14] S. Dodge and L. Karam, "Understanding how image quality affects
deep neural networks," in 2016 eighth international conference on
quality of multimedia experience (QoMEX), 2016.
[15] S. P. Kannojia and G. Jaiswal, "Effects of varying resolution on
performance of CNN based image classification: An experimental
study," Int. J. Comput. Sci. Eng, vol. 6, p. 451–456, 2018.
[16] N. Sharma, V. Jain and A. Mishra, "An analysis of convolutional
neural networks for image classification," Procedia computer science,
vol. 132, p. 377–384, 2018.
[17] S. P. Kannojia and G. Jaiswal, "Ensemble of hybrid CNN-ELM
model for image classification," in 2018 5th International Conference
on Signal Processing and Integrated Networks (SPIN), 2018.
[18] S. P. Kannojia and G. Jaiswal, "Meta Fusion of Trained
Heterogeneous Image Classifier using Classifier Selection," in
National Conference on Intelligent Computing, Communication and
Data Science (NCICCD-2018), 2018.
[19] Y. LeCun, L. Bottou, Y. Bengio and P. Haffner, "Gradient-based
learning applied to document recognition," Proceedings of the IEEE,
vol. 86, p. 2278–2324, 1998.
[20] H. Xiao, K. Rasul and R. Vollgraf, "Fashion-mnist: a novel image
dataset for benchmarking machine learning algorithms," arXiv
preprint arXiv:1708.07747, 2017.
[21] A. Krizhevsky, G. Hinton and others, "Learning multiple layers of
features from tiny images," 2009.
[22] A. Krizhevsky, I. Sutskever and G. E. Hinton, "Imagenet
classification with deep convolutional neural networks,"
Communications of the ACM, vol. 60, p. 84–90, 2017.
... In turn, considering only incremental data makes the model prone to "catastrophic forgetting". In response to these problems, many scholars have proposed different incremental learning frameworks [26][27][28]. For example, Cui et al. proposed an adaptive feedback handover (SAFH) algorithm to solve the dynamic handover problem in IoV, using an incremental feedback mechanism to dynamically capture the splitting properties of decision trees [29]. ...
Article
Full-text available
With the rapid increase of data, centralized machine learning can no longer meet the application requirements of the Internet of Vehicles (IoV). On the one hand, both car owners and regulators pay more attention to data privacy and are unwilling to share data, which forms the isolated data island challenge. On the other hand, the incremental data generated in IoV are massive and diverse. All these issues have brought challenges of data increment and data diversity. The current common federated learning or incremental learning frameworks cannot effectively integrate incremental data with existing machine learning (ML) models. Therefore, this paper proposes a Federated Learning Framework Based on Incremental Weighting and Diversity Selection for IoV (Fed-IW&DS). In Fed-IW&DS, a vehicle diversity selection algorithm was proposed, which uses a variety of performance indicators to calculate diversity scores, effectively reducing homogeneous computing. Also, it proposes a vehicle federated incremental algorithm that uses an improved arctangent curve as the decay function, to realize the rapid fusion of incremental data with existing ML models. Moreover, we have carried out several sets of experiments to test the validity of the proposed Fed-IW&DS framework’s performance. The experimental results show that, under the same global communication round and similar computing time, the Fed-IW&DS framework has significantly improved performance in all aspects compared to the frameworks FED-AVG, FED-SGD, FED-prox & the decay functions linear, square curve and arc tangent. Specifically, the Fed-IW&DS framework improves the Acc (accuracy), loss (loss), and Matthews correlation coefficient (MCC) by approximately 32%, 83%, and 66%, respectively. This result shows that Fed-IW&DS is a more reliable solution than the common frameworks of federated learning, and it can effectively deal with the dynamic incremental data in the IoV scenario. Our findings should make a significant contribution to the field of federated learning.
Conference Paper
Full-text available
Several fusion methods are used to integrate different trained classification models to generalize classification over the distributed dataset. This paper proposes a Meta fusion approach for integrating trained heterogeneous image classifier using classifier selection. In this approach, classifier selector selects a trained classifier from the set of trained classifier on distributed heterogeneous data. The selected model classifies an input image into its predicted class. Integrated model, based on Meta fusion approach, of these trained classifier works as a generalized classification model for combined distributed heterogeneous image data. Meta fusion approach is verified on MNIST image dataset in an environment of distributed heterogeneous data and achieves competitive results.
Article
Full-text available
Image quality is affected by different types of quality factors (such as resolution, noise, contrast, blur, compression). Resolution of the image affects the visual information of the image. Image with higher resolution contains more visual information details while an image with lower resolution contains less visual details. Convolutional neural network (CNN) based image classifiers always take input as an image, automatically learn image features and classify into its output class. If input image resolution varies, then it hinders classification performance of CNN based image classifier. This paper proposes a methodology (training testing methods TOTV, TVTV) and presents the experimental study on the effects of varying resolution on CNN based image classification for standard image dataset MNIST and CIFAR10. The experimental result shows that degradation in image resolution from higher to lower decreases performance score (accuracy, precision and F1 score) of CNN based Image classification.
Conference Paper
Full-text available
To leverage feature representation of CNN and fast classification learning of ELM, Ensemble of Hybrid CNN-ELM model is proposed for image classification. In this model, image representation features are learned by Convolutional Neural Network (CNN) and fed to Extreme Learning Machine (ELM) for classification. Three hybrid CNN-ELMs are ensemble in parallel and final output is computed by majority voting ensemble of these classifier’s outputs. The experiments show this ensemble model improves the classifier’s classification confidence and accuracy. This model has been benchmarked on MNIST benchmark dataset and effectively improved the accuracy in comparison of single hybrid CNN-ELM classifier with accuracy upto 99.33%. Proposed ensemble model has been also compared with Core CNN, Core ELM, Hybrid CNN-ELM and achieves competitive accuracy.
Article
Full-text available
This paper presents an empirical analysis of theperformance of popular convolutional neural networks (CNNs) for identifying objects in real time video feeds. The most popular convolution neural networks for object detection and object category classification from images are Alex Nets, GoogLeNet, and ResNet50. A variety of image data sets are available to test the performance of different types of CNN’s. The commonly found benchmark datasets for evaluating the performance of a convolutional neural network are anImageNet dataset, and CIFAR10, CIFAR100, and MNIST image data sets. This study focuses on analyzing the performance of three popular networks: Alex Net, GoogLeNet, and ResNet50. We have taken three most popular data sets ImageNet, CIFAR10, and CIFAR100 for our study, since, testing the performance of a network on a single data set does not reveal its true capability and limitations. It must be noted that videos are not used as a training dataset, they are used as testing datasets. Our analysis shows that GoogLeNet and ResNet50 are able to recognize objects with better precision compared to Alex Net. Moreover, theperformance of trained CNN’s vary substantially across different categories of objects and we, therefore, will discuss the possible reasons for this.
Article
Full-text available
Whereas people learn many different types of knowledge from diverse experiences over many years, and become better learners over time, most current machine learning systems are much more narrow, learning just a single function or data model based on statistical analysis of a single data set. We suggest that people learn better than computers precisely because of this difference, and we suggest a key direction for machine learning research is to develop software architectures that enable intelligent agents to also learn many types of knowledge, continuously over many years, and to become better learners over time. In this paper we define more precisely this never-ending learning paradigm for machine learning, and we present one case study: the Never-Ending Language Learner (NELL), which achieves a number of the desired properties of a never-ending learner. NELL has been learning to read the Web 24hrs/day since January 2010, and so far has acquired a knowledge base with 120mn diverse, confidence-weighted beliefs (e.g., servedWith(tea,biscuits)), while learning thousands of interrelated functions that continually improve its reading competence over time. NELL has also learned to reason over its knowledge base to infer new beliefs it has not yet read from those it has, and NELL is inventing new relational predicates to extend the ontology it uses to represent beliefs. We describe the design of NELL, experimental results illustrating its behavior, and discuss both its successes and shortcomings as a case study in never-ending learning. NELL can be tracked online at http://rtw.ml.cmu.edu, and followed on Twitter at @CMUNELL.
Article
Full-text available
Humans and animals have the ability to continually acquire and fine-tune knowledge throughout their lifespan. This ability, referred to as lifelong learning, is mediated by a rich set of neurocognitive mechanisms that together contribute to the development and specialization of our sensorimotor skills as well as to long-term memory consolidation and retrieval. Consequently, lifelong learning capabilities are crucial for computational learning systems and autonomous agents interacting in the real world and processing continuous streams of information. However, lifelong learning remains a long-standing challenge for machine learning and neural network models since the continual acquisition of incrementally available information from non-stationary data distributions generally leads to catastrophic forgetting or interference. This limitation represents a major drawback also for state-of-the-art deep and shallow neural network models that typically learn representations from stationary batches of training data, thus without accounting for situations in which the number of tasks is not known a priori and the information becomes incrementally available over time. In this review, we critically summarize the main challenges linked to lifelong learning for artificial learning systems and compare existing neural network approaches that alleviate, to different extents, catastrophic forgetting. Although significant advances have been made in domain-specific learning with neural networks, extensive research efforts are required for the development of robust lifelong learning on autonomous agents and robots. We discuss well-established and emerging research motivated by lifelong learning factors in biological systems such as neurosynaptic plasticity, critical developmental stages, multi-task transfer learning, intrinsically motivated exploration, and crossmodal learning.
Article
Full-text available
Deep domain adaption has emerged as a new learning technique to address the lack of massive amounts of labeled data. Compared to conventional methods, which learn shared feature subspaces or reuse important source instances with shallow representations, deep domain adaption methods leverage deep networks to learn more transferable representations by embedding domain adaptation in the pipeline of deep learning. There have been comprehensive surveys for shallow domain adaption, but few timely reviews the emerging deep learning based methods. In this paper, we provide a comprehensive survey of deep domain adaptation methods for computer vision applications with four major contributions. First, we present a taxonomy of different deep domain adaption scenarios according to the properties of data that define how two domains are diverged. Second, we summarize deep domain adaption approaches into several categories based on training loss, and analyze and compare briefly the state-of-the-art methods under these categories. Third, we overview the computer vision applications that go beyond image classification, such as face recognition, semantic segmentation and object detection. Fourth, some potential deficiencies of current methods and several future directions are highlighted.
Article
Full-text available
Recently, incremental and on-line learning gained more attention especially in the context of big data and learning from data streams, conflicting with the traditional assumption of complete data availability. Even though a variety of different methods are available, it often remains unclear which of them is suitable for a specific task and how they perform in comparison to each other. We analyze the key properties of eight popular incremental methods representing different algorithm classes. Thereby, we evaluate them with regards to their on-line classification error as well as to their behavior in the limit. Further, we discuss the often neglected issue of hyperparameter optimization specifically for each method and test how robustly it can be done based on a small set of examples. Our extensive evaluation on data sets with different characteristics gives an overview of the performance with respect to accuracy, convergence speed as well as model complexity, facilitating the choice of the best method for a given application.