AMC-IoT: Automatic Modulation Classification

Using Efficient Convolutional Neural Networks for

Low Powered IoT Devices
Muhammad Usman and Jeong-A Lee
Department of Computer Engineering, Chosun University, Republic of Korea
Email: [email protected], [email protected]

Abstract—Automatic modulation classification (AMC) is used recognition can extract the digital baseband information based
to identify the modulation for the received signal. IoT devices use on very low or no prior information about the device type and
modern communication methods which are based on multiple transmission schemes. AMC is regarded as an intermediate
input multiple output (MIMO) in which the signals are received
from various sources. The identification of modulation is vital. step between the signal detection and demodulation at the
Feature based AMC methods combined with deep learning receiving end. The AMC problem can be divided into to
techniques has the potential to meet the latency requirement in categories: liklihood based (LB) and feature based (FB). In
the IoT applications. An efficient convolutional neural network the LB methods the ratio of the liklihood of the signals
based on depthwise separable convolution has been proposed to are measured and the decision is made on the basis of
classify the modulation of the received signals. The proposed
architecture has 58% less parameters than the conventional the comparison of ratios to a threshold. Particularly, various
convolutional architecture and the performance is comparable. liklihood solutions with unknown modulation parameters on
Index Terms—automatic modulation classification, convolu- the basis of power spectral density (PSD) of the received signal
tional neural networks, depthwise separable, IoT are tested. It has been observed that the LB methods yields
good results but at the cost of computational complexity. Most
I. I NTRODUCTION of the studies have focused on the accurate prediction of the

T HE advent of internet of things (IoT) and evolution in

the mobile networks is driven by the need to satisfy the
consumer’s demand of enhanced performance, high speeds,
modulation schemes using LB method did not consider the
computation complexity [5], [6]. The FB methods, developed
as a suboptimal classifier for the practical implementation,
seamless links elasticity, and portability in the telecommuni- relies on the extraction of useful features from the received
cation network. The 5th Generation technology abbreviated signal and then classify them using a classifier. The features
as 5G is the next generation telecommunication standard that can be of instantaneous time, transform based, statistical or
assures to meet the ever growing needs of the communication based on constellation shape. Extensive research have been
applications. carried out in both of these methods and it is found that the
IoT is a potential candidate to utilize the communication LB method provides an optimal solution with good accuracy,
network resources and it is estimated that more than 50 billion but that demands high computation and prior knowledge about
devices are now connected to the internet [1]. Considering the the signal. On the contrary, the FB method presents a sub-
current 5G communication, the receiver is expected to receive optimal solution but has relatively low latency since it does not
the signals from multiple directions. Therefore, the well- require any prior knowledge. Conventional FB methods rely
known problem of multi-path fading is inevitable, and would on the expert’s knowledge and may perform better in certain
lead to the difficulty in identification of signals. Multiple-input situations if not all. Furthermore, to meet the latency standards
multiple output (MIMO) shall be actively utilized in the 5G of 5G as defined in [7], artificial intelligence, specially deep
communication which means the signals would be received learning (DL) has been effectively utilized for the purpose
from several sources. Furthermore, the software defined radio of modulation classification [8], [9] . A trivial issue in the
(SDR) and the cognitive radio (CR) based technologies are utilization of ML methods is the feature engineering, for
getting immense popularity in which the devices are expected which expert’s experience is required. To obviate feature
to adopt the neighboring conditions and adjust their trans- engineering in the FB methods, several machine learning
mission parameters, modulation schemes etc. It is therefore (ML) based classifiers, particularly deep learning methods
of utmost importance to be able to classify the modulation have been thoroughly studied as they present the tendency
type at the receiver’s end without needing the prior knowl- to learn complex functions with more significance compared
edge about the transmitter’s parameters. Several researchers to their shallow counterparts. With the rapid advancement
have proposed the method of automatically classifying the in the DL technologies, many methods have been proposed
modulation schemes which is termed as automatic modulation that can self-learn the features, furthermore, DL is favoured
classification (AMC) [2]–[4], and presented that modulation because it requires large datasets which can be easily obtained

978-1-7281-6758-9/20/$31.00 ©2020 IEEE 288 ICTC 2020

in the communication systems. One of the major concern proposed algorithm achieved 99% recognition performance.
using DL could be its complexity, as DL involves the phases The DL algorithm have been utilized by [22], [23] and
of training and testing of the model. However, an efficient [24], in the complex-valued temporal signal domain. Deep
convolutional network architecture which performs depthwise autoencoders (DAE) that are formed by two symmetrical deep
convolution has been utilized in many applications. Due the belief networks are the unsupervised learning algorithm that
depthwise convolutions, the model size is significantly reduced generates an output similar to its input but reconstructs it using
without affecting the accuracy. The model has less number of fewer dimensions. The dimension reduction is performed in
parameters compared to the conventional convolution, thus is the latent space which is composed of few neurons. This
suitable to match the design requirements of the small size latent space is regularized to be sparse, which helps it to
devices in the IoT environment. In this study, the efficacy learn the most important features to reproduce the input.
of the depthwise convolution method is compared with the The discriminative features learnt by the latent space can be
conventional convolution neural network. utilized to perform the classification task. The autoencoder
The rest of the paper is organized as follows: Section II is composed of three sections: encoder, decoder and latent
presents the review of automatic modulation classification space. Several configurations of autoencoder can be found in
based on deep learning based methods, the details of the literature that have been tuned to perform the required tasks.
proposed method is discussed in Section III. Results are Most common of them are the regularized autoencoders in
presented in Section IV and the paper is concluded in Section which the encoding function is not linear and allows the AE
V. to show sparsity [25], [26]. One of the most successful DL
architectures, convolutional neural network (CNN), has been
II. BACKGROUND widely adopted to address the modulation recognition problem.
The deep learning is the subset of artificial neural networks For instance, in [27], CNN has been trained to classify the
that tries to develop the relationship between the input and modulation by converting the modulated signals in to images.
the target by mimicking the human brain’s mechanism. It is Images including the constellation diagrams have been used to
a layered architecture consisting of an input layer, a hidden train a famous CNN architecture named as AlexNet [28]. Some
layer/layers in the middle and an output layer. The number advancements towards the representation of modulated data
of hidden layers governs the depth of the architecture and into gridlike topology for training of CNN have been presented
theoretically any architecture with more than one hidden in [29]. In particular, methods to convert complex signals to
layer is termed as a deep network. Deep learning algorithms images have been presented along with analysis of impacts
have been used in variety of classification problems ranging of selection of conventional parameters, and finally a CNN
from computer vision [10], bio-informatics [11]–[13], natural model is trained to show the effects of the proposed study. A
language processing [14], speech recognition [15], and signal robust method for the modulation classification is proposed in
processing [16], [17] etc. It is a branch of machine learning which 15 modulation types are used to train a CNN without
that deals with large data and treats it as an input which is needing to extract the features from the signals [30]. The use
carried into multilayered architecture containing hidden nodes. of the robust features provided good accuracy results. In [31],
The final layer of the architecture is called the output layer three architectures including Convolutional Long Short-term
which provides the desired target classes. By virtue of these Deep Neural Network (CLDNN), a Long Short-Term Mem-
hidden layered architecture, the DL-based methods outperform ory neural network (LSTM), and a deep Residual Network
classical machine learning techniques. Typically the deep (ResNet) have been studied with the aim of maintaining a
neural network learn the features from the data autonomously, good performance in while adopting strategies to reduce the
therefore manual feature representation to the machine is amount of time required in training the networks. The ability
not required. This makes it superior to the conventional ML to classify the modulation in short amount of time using a
approaches that require detection and extraction of effective comparatively less complex architecture is of great interest,
features which is an arduous and time consuming task. Deep since the devices in the IoT, unmanned air vehicles (UAVs)
neural networks assemble the simple features to develop a and other technologies in the 5G have limited computation
complex feature by themselves through multiple non-linear and power resources. A significant amount of research has
transformations [18]. Several research has been conducted to been carried out by Shea et.al who have presented CNN based
address the modulation classification problem which has been solutions to classify the modulation and presented open access
handled through conventional signal processing in the past dataset for training of the neural networks which have been
[19], [20]. adopted in several of the above mentioned literature. We also
With the success of ML and DL many researchers have utilize the datasets from [22], and train a convolutional neural
opted these methods propose automatic modulation classifica- network that performs depthwise convolution and compare its
tion methods. In [21], an artificial neural network (ANN) based results with that of a conventional neural network architecture.
method coupled with the genetic algorithm (GA) has been Hence we explore a less complex automatic modulation clas-
proposed, and utilized resilient backpropagation (RPROP) sification method to assist the low powered devices employed
for the first time to recognize digital modulation. GA was in the wireless communication. This development would be
employed to extract the six discriminating features and the beneficial for the deployment of deep learning based solution

289
in the wireless communication. • Depthwise Convolution uses single filter for each input
channel i.e, if there are 3 input channels, there will be 3
A. Convolutional Neural Network DkxDk spatial convolutions.
• Pointwise Convolution is used to change the dimension
CNNs are regularized form of multilayer perceptrons, in-
of the output of the depthwise convolution and makes
spired by the biological process of connection of neurons.
is similar to the output of the convention convolutional
They are effectively utilized in various classification tasks
neural network output layer. It does so by applying a
because require little preprocessing in comparison with other
1x1 convolution having a depth equal to the number of
classification algorithms. A convolutional neural network con-
channels of the depthwise convolution layer.
denses the input to the important features that helps in distin-
guishing the input. Both 2D and 3D convolution architectures In standard convolution the filtration and combination is
are commonly employed according to the requirement of the performed in a single step where as in depthwise separable
user. It is essentially composed of four layers. convolutions, the filtration and combination task are separated
in two layers as shown in Fig. 3.
• Convolution Layer: The convolution layer is responsible
The general form of 2D convolution is shown in (1)
to perform the convolution operation between the filter
and the input map. After the filter is passed over the z[m, n] = x[m, n] ∗ k[m, n]
underlying input is changed in a way that certain features ∞ ∞
(1)
of the input are emphasized. = x[i, j] · k[m − i, n − j]
• Pooling Layer: Some applications require compression j=−∞ i=−∞

of the output of the convolution layer. Pooling is a which is equivalent to (2), since the convolution operation
way of summarizing the features by downsampling the is commutative
feature map. This adds robustness for the features to the
position changes. Most commonly used pooling methods z[m, n] = x[m, n] ∗ k[m, n]
are average pooling and the max pooling. ∞ ∞
  (2)
• Fully connected layer: The classification task is per- = k[i, j] · x[m − i, n − j]
formed using the dense layers in which each neuron is j=−∞ i=−∞
usually connected to each neuron in the succeeding layer
with some weights and activations. For separable convolution, the kernel is separated into
• Output Layer: The output layer is the last layer which [Mx1] and [1xN] as shown in (3)
has a certain activation function to calculate the proba- k[m, n] = k1 [m] · k2 [n] (3)
bility response.
Substituting k[m, n] in (2)
A typical CNN architecture is shown in Fig. 1.
z[m, n] = x[m, n] ∗ k[m, n]
∞ ∞
= k[i, j] · x[m − i, n − j]
j=−∞ i=−∞
 ∞
∞ 
(4)
= k1 [i] · k2 [j] · x[m − i, n − j]
−∞ −∞
∞
 ∞

 
= k2 [j] k1 [i] · x[m − i, n − j]
−∞ −∞

The 1D convulotion is given as (5)

Fig. 1. Convolutional Neural Network
z[n] = x[n] ∗ k[n]
∞
(5)
= x[p] · h[n − p]
B. Depthwise Separable Convolution
p=−∞
The concept of depthwise separable convolution was first In separable convolution, the input is convolved twice, once
proposed in [32] and was used by the Inception models in the horizontal direction and once in the vertical direction.
for complexity reduction in the first few layers [33]. Other Due to associative property of convolution, the order does not
methods including shrinking of the pre-trained network, com- matter, therefore, any of the following convolutions shown in
pression on the basis of product quantization [34], hashing and (6) can be performed first followed by the other.
pruning have been adopted to obtain a small network.
1) Architecture: Depthwise separable convolution is a z[m, n] = (h1 [m] · h2 [n] ∗ x[m, n])
depthwise convolution followed by a pointwise convolution = h2 [n] ∗ (h1 [m] ∗ x[m, n]) (6)
as shown in Fig. 2. = h1 [m] ∗ (h2 [n] ∗ x[m, n])

290
 – A conventional convolutional neural network using
the Keras Conv2D API is trained on each dataset
and performance is measured.
– Depthwise spatial convolution using the Keras
DepthwiseConv2D, which is the first step of the
depthwise separable convolution has been utilized for
the classification of each dataset.
– Depthwise separable convolution that performs
depthwise spatial convolution followed by the point-
wise convolution has been used using Keras Separa-
bleConv2D API to classify the two dataset.
– The performance of the models and the correspond-
ing number of parameters have been presented.
The implementation is done on Python using Theano/Keras
with GPU RTX2070, 2560 CUDA cores, and 8GB GDDR6
Fig. 2. Depthwise Separable Convolution
VRAM on top of Intel(R) Core(TM) i7-8700 CPU @ 3.20GHz
with 64 GB DDR3 RAM.
A difference between the standard and the depthwise sepa- A. Dataset
rable convolution is depicted in Fig. 3.
The dataset used in this work is obtained from GNU Radio
[35]. The datasets are named as RML2016a and RML2016c
respectively. The data is deterministically generated to match
the real system’s data, furthermore, the parameters of pulse
shaping, modulation and carried data are made identical. The
dataset named RML2016a is composed of two data sources.
Continuous signals including voice and the publicly available
Serial Episode #1 are used as an input for the analog modu-
lations. For digital communication entire Gutenberg works of
Shakespeare in ASCII is used. To make the digital modulated
data equiprobable, block normalizer has been utilized for
data whitening. The synthetic signals are then passed through
several channels to get the effect of unknown scaling and
translation. The GNU radio channel model block generates
the final data set, of which the time series signals are sliced
into the training and test dataset by 128 samples rectangular
window. The dataset RML2016c has been constructed by
collecting the widely used modulations. Both discrete and
continuous alphabets for digital and analog modulation data is
modulated over modem and is exposed to channel effects using
GNU Radio. The sampling is performed and 128 samples
with a shift of 64 samples are extracted. The dataset is
finally labeled into 11 classes corresponding to the modulation
scheme.
Fig. 3. Standard convolution in (a) replaced by two layers of depthwise
convolution in (b) and pointwise convolution in (c)
IV. R ESULTS AND E VALUATION
We utilize the two datasets mentioned in section III-A to
III. P ROPOSED D EEP M ODULATION C LASSIFIER train the three neural networks and compare their perfor-
mances. The weights of the convolutional layers in all the net-
In this study we use depthwise separable convolution neural works are initialized using ”Glorot” initializer, and the weights
network to minimize the number of parameters and conse- in the fully connected layers of the networks are initialized
quently the number of multiplication operations during the using ”He” initializer. The weights are learned through Adam
convolution process. optimizer by keeping a learning rate of 0.01 and categorical
• A lightweight modulation classification algorithm based cross-entropy loss function is minimized. The networks are
on convolutional neural network has been proposed for trained for 100 epochs with a batch size of 1024 samples.
two different datasets. The total number of parameters were 921611, 596491 and
• Three models have been evaluated: 385307 for Conv2D, Depthwise and Separable convolutional

291
 Accuracy
Architecture Parameters Reduction
RML2016a RML2016c
Conv2D 921611 – 71.30% 83.40%
Depthwise 596491 35.2% 69.90% 92.10%
Separable 385307 58.2% 71.25% 83.03%
TABLE I
C OMPARISON OF ACCURACY ACHIEVED ON 18 D B SNR OF BOTH
DATESETS ON DIFFERECT ARCHITECTURE CONFIGURATIONS

neural networks respectively. The accuracy achieved by these

architectures is shown in Table I and graphically depicted
in Fig. 4 and Fig. 5 for dataset RML2016a and RML2016c
respectively. Fig. 6. Confusion matrix depicting the ability of the conventional CNN to
The accuracy curves for different SNR values have been classify several modulation schemes at SNR=18 on RML2016a dataset (a)
and RML2016c dataset (b)
plotted for both datasets trained on the convolutional neural
network architectures. For RML2016a dataset, the perfor-
mances of the separable convolutional network is comparable
to the conventional neural network architecture as shown in
Fig. 4. For RML2016c dataset, the depthwise and separable
CNN configurations outperforms the conventional CNN ap-
proach as depicted in Fig. 5, suggesting efficient performance
of the network with reduced parameters.

Fig. 7. Confusion matrix depicting the ability of the deptwise CNN to

classify several modulation schemes at SNR=18 on RML2016a dataset (a)
and RML2016c dataset (b)

Fig. 6, 7 and 8 depicts the confusion matrices of the CNN,

depthwise CNN and separable CNN respectively on samples
from both datasets with SNR 18dB.

V. C ONCLUSION
Fig. 4. Classification Accuracy of the convolutional neural network architec-
tures on RML2016a dataset Depthwise and separable convolutional neural network ar-
chitectures have been presented to classify the modulation
schemes. The low powered IoT devices does not have the
capability to process large number of parameters as in the
conventional neural networks. The proposed methods achieves
similar and in some cases better performance compared to
the conventional convolution approaches with significantly less
number of parameters. Such architectures are highly favorable
for implementation in the devices that are constrained by
power and area.

ACKNOWLEDGMENT
This research was supported by Basic Science Research
Program through the National Research Foundation of Ko-
rea (NRF) funded by the Ministry of Education (NRF-
2020R1I1A3063857) and in part by Korea Institute of Energy
Fig. 5. Classification Accuracy of the convolutional neural network architec- Technology Evaluation and Planning, and the Ministry of
tures on RML2016c dataset Trade, Industry and Energy of the Republic of Korea (No.
20184010201650).

292
 [14] T. Young, D. Hazarika, S. Poria, and E. Cambria, “Recent trends in
deep learning based natural language processing,” ieee Computational
intelligenCe magazine, vol. 13, no. 3, pp. 55–75, 2018.
[15] D. Amodei, S. Ananthanarayanan, R. Anubhai, J. Bai, E. Battenberg,
C. Case, J. Casper, B. Catanzaro, Q. Cheng, G. Chen et al., “Deep
speech 2: End-to-end speech recognition in english and mandarin,” in
International conference on machine learning, 2016, pp. 173–182.
[16] M. Usman, M. S. Ibrahim, J. Ahmed, S. S. Hussain, and M. Moinuddin,
“Quantum calculus-based volterra lms for nonlinear channel estimation,”
in 2019 Second International Conference on Latest trends in Electrical
Engineering and Computing Technologies (INTELLECT), 2019, pp. 1–4.
[17] S. Khan, M. Usman, I. Naseem, R. Togneri, and M. Bennamoun, “Vp-
flms: a novel variable power fractional lms algorithm,” in 2017 Ninth
International Conference on Ubiquitous and Future Networks (ICUFN).
IEEE, 2017, pp. 290–295.
[18] I. Goodfellow, Y. Bengio, and A. Courville, Deep learning. MIT press,
Fig. 8. Confusion matrix depicting the ability of the separable CNN to 2016.
classify several modulation schemes at SNR=18 on RML2016a dataset (a) [19] A. Swami and B. M. Sadler, “Hierarchical digital modulation classifica-
and RML2016c dataset (b) tion using cumulants,” IEEE Transactions on communications, vol. 48,
no. 3, pp. 416–429, 2000.
[20] O. A. Dobre, A. Abdi, Y. Bar-Ness, and W. Su, “Survey of auto-
matic modulation classification techniques: classical approaches and new
R EFERENCES trends,” IET communications, vol. 1, no. 2, pp. 137–156, 2007.
[21] M. D. Wong and A. K. Nandi, “Automatic digital modulation recognition
[1] M. Usman, I. Ahmed, M. I. Aslam, S. Khan, and U. A. using artificial neural network and genetic algorithm,” Signal Processing,
Shah, “Sit: A lightweight encryption algorithm for secure internet vol. 84, no. 2, pp. 351–365, 2004.
of things,” International Journal of Advanced Computer Science [22] T. J. O’Shea, J. Corgan, and T. C. Clancy, “Convolutional radio modula-
and Applications, vol. 8, no. 1, 2017. [Online]. Available: tion recognition networks,” in International conference on engineering
http://dx.doi.org/10.14569/IJACSA.2017.080151 applications of neural networks. Springer, 2016, pp. 213–226.
[2] W. Chen, Z. Xie, L. Ma, J. Liu, and X. Liang, “A faster maximum- [23] T. O’Shea and J. Hoydis, “An introduction to deep learning for the
likelihood modulation classification in flat fading non-gaussian chan- physical layer,” IEEE Transactions on Cognitive Communications and
nels,” IEEE Communications Letters, vol. 23, no. 3, pp. 454–457, 2019. Networking, vol. 3, no. 4, pp. 563–575, 2017.
[3] J. Nie, Y. Zhang, Z. He, S. Chen, S. Gong, and W. Zhang, “Deep [24] T. J. O’Shea, T. Roy, and T. C. Clancy, “Over-the-air deep learning based
hierarchical network for automatic modulation classification,” IEEE radio signal classification,” IEEE Journal of Selected Topics in Signal
Access, vol. 7, pp. 94 604–94 613, 2019. Processing, vol. 12, no. 1, pp. 168–179, 2018.
[4] J. Jagannath, N. Polosky, D. O’Connor, L. N. Theagarajan, B. Sheaffer, [25] Q. Ji, Y. Sun, J. Gao, Y. Hu, and B. Yin, “Nonlinear subspace clustering
S. Foulke, and P. K. Varshney, “Artificial neural network based automatic via adaptive graph regularized autoencoder,” IEEE Access, vol. 7, pp.
modulation classification over a software defined radio testbed,” in 2018 74 122–74 133, 2019.
IEEE International Conference on Communications (ICC). IEEE, 2018, [26] W. Xu, S. Keshmiri, and G. Wang, “Adversarially approximated autoen-
pp. 1–6. coder for image generation and manipulation,” IEEE Transactions on
[5] T. Yucek and H. Arslan, “A novel sub-optimum maximum-likelihood Multimedia, vol. 21, no. 9, pp. 2387–2396, 2019.
modulation classification algorithm for adaptive ofdm systems,” in 2004 [27] S. Peng, H. Jiang, H. Wang, H. Alwageed, and Y.-D. Yao, “Modulation
IEEE Wireless Communications and Networking Conference (IEEE Cat. classification using convolutional neural network based deep learning
No. 04TH8733), vol. 2. IEEE, 2004, pp. 739–744. model,” in 2017 26th Wireless and Optical Communication Conference
[6] W. Wei and J. M. Mendel, “A new maximum-likelihood method for (WOCC). IEEE, 2017, pp. 1–5.
modulation classification,” in Conference Record of the Twenty-Ninth [28] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification
Asilomar Conference on Signals, Systems and Computers, vol. 2. IEEE, with deep convolutional neural networks,” in Advances in neural infor-
1995, pp. 1132–1136. mation processing systems, 2012, pp. 1097–1105.
[7] I. Parvez, A. Rahmati, I. Guvenc, A. I. Sarwat, and H. Dai, “A survey [29] S. Peng, H. Jiang, H. Wang, H. Alwageed, Y. Zhou, M. M. Sebdani,
on low latency towards 5g: Ran, core network and caching solutions,” and Y.-D. Yao, “Modulation classification based on signal constellation
IEEE Communications Surveys & Tutorials, vol. 20, no. 4, pp. 3098– diagrams and deep learning,” IEEE transactions on neural networks and
3130, 2018. learning systems, vol. 30, no. 3, pp. 718–727, 2018.
[8] Y. Wang, M. Liu, J. Yang, and G. Gui, “Data-driven deep learning for [30] S. Zhou, Z. Yin, Z. Wu, Y. Chen, N. Zhao, and Z. Yang, “A robust
automatic modulation recognition in cognitive radios,” IEEE Transac- modulation classification method using convolutional neural networks,”
tions on Vehicular Technology, vol. 68, no. 4, pp. 4074–4077, 2019. EURASIP Journal on Advances in Signal Processing, vol. 2019, no. 1,
[9] Y. Wang, J. Yang, M. Liu, and G. Gui, “Lightamc: Lightweight au- p. 21, 2019.
tomatic modulation classification via deep learning and compressive [31] S. Ramjee, S. Ju, D. Yang, X. Liu, A. E. Gamal, and Y. C. Eldar, “Fast
sensing,” IEEE Transactions on Vehicular Technology, vol. 69, no. 3, deep learning for automatic modulation classification,” arXiv preprint
pp. 3491–3495, 2020. arXiv:1901.05850, 2019.
[10] N. Akhtar and A. Mian, “Threat of adversarial attacks on deep learning [32] L. Sifre and S. Mallat, “Rigid-motion scattering for image classification,”
in computer vision: A survey,” IEEE Access, vol. 6, pp. 14 410–14 430, Ph. D. thesis, 2014.
2018. [33] S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep
[11] M. Usman, S. Khan, and J.-A. Lee, “Afp-lse: Antifreeze proteins network training by reducing internal covariate shift,” arXiv preprint
prediction using latent space encoding of composition of k-spaced amino arXiv:1502.03167, 2015.
acid pairs,” Scientific Reports, vol. 10, no. 1, pp. 1–13, 2020. [34] J. Wu, C. Leng, Y. Wang, Q. Hu, and J. Cheng, “Quantized convolutional
[12] M. Usman and J. A. Lee, “Afp-cksaap: Prediction of antifreeze proteins neural networks for mobile devices,” in Proceedings of the IEEE
using composition of k-spaced amino acid pairs with deep neural net- Conference on Computer Vision and Pattern Recognition, 2016, pp.
work,” in 2019 IEEE 19th International Conference on Bioinformatics 4820–4828.
and Bioengineering (BIBE). IEEE, 2019, pp. 38–43. [35] T. J. O’shea and N. West, “Radio machine learning dataset generation
[13] S. M. Ibrahim, M. S. Ibrahim, M. Usman, I. Naseem, and M. Moinuddin, with gnu radio,” in Proceedings of the GNU Radio Conference, vol. 1,
“A study on heart segmentation using deep learning algorithm for no. 1, 2016.
mri scans,” in 2019 13th International Conference on Mathematics,
Actuarial Science, Computer Science and Statistics (MACS). IEEE,
2019, pp. 1–5.

293