1st International Conference on Advances in Science, Engineering and Robotics Technology 2019 (ICASERT 2019)

GRU-based Attention Mechanism for Human

Activity Recognition
Md. Nazmul Haque M. Tanjid Hasan Tonmoy Saif Mahmud
Institute of Information Technology Dept. of CSE Dept. of CSE
University of Dhaka University of Dhaka University of Dhaka
Dhaka, Bangladesh Dhaka, Bangladesh Dhaka, Bangladesh
[email protected] [email protected] [email protected]

Amin Ahsan Ali Muhammad Asif Hossain Khan Mohammad Shoyaib

Dept. of CSE Dept. of CSE Institute of Information Technology
Independent University of Bangladesh University of Dhaka University of Dhaka
Dhaka, Bangladesh Dhaka, Bangladesh Dhaka, Bangladesh
[email protected] [email protected] [email protected]

Abstract—Sensor data based Human Activity Recognition This techniques mainly depend on heuristic based hand-crafted
(HAR) has gained interest due to its application in practical field.feature engineering that rely on low level representations. As
With increasing number of approaches incorporating feature the traditional machine learning models use low level repre-
learning of sequential time-series sensor data, in particular the
deep learning based ones has performed reasonably in uniform sentations, it lacks the characteristics of generalization [4], [5].
labeled data distribution scenario. However, most of these meth- High level abstraction along with low level representations are
ods do not capture properly the temporal context of time-steps necessary for a likely solution to this generalization problem.
in sequential time-series data. Moreover, the situation becomes Deep learning based methods deal with both low and high
worse for imbalanced class distribution which is a usual case for level representations of data. Therefore, recently two variants
HAR using body-worn sensor devices. To solve this issues, we
have integrated hierarchical attention mechanism with recurrent of deep learning methods namely convolutional [6] and re-
units of neural network in order to obtain temporal context current [7] neural network models are become dominant over
within the time-steps of data sequence. The introduced model in traditional methods in terms of performance. For example, to
this paper has achieved better performance with respect to the identify HAR, convolutional neural network(CNN) is used in
well-defined e valuation m etrics i n b oth u niform a nd imbalanced[8] and [9] and recurrent neural network(RNN) is used in [7]
class distribution than the existing state-of-the-art deep learning
based model. and [10].
Index Terms—Human Activity Recognition, Attention Mecha- Although deep learning based methods show promising
nism, Gated Recurrent Unit performance, conventional sliding window based approach for
CNN is unable to fully capture the temporal context of the
I. I NTRODUCTION sensor reading [11] which is required for better classification
Human Activity Recognition(HAR) is a domain of research of activities. For sequence data, RNN performs better than
aimed at recognizing human actions and movements from a CNN in most cases as it captures sequence information [12].
series of observations. The increasing public adoption of smart However, RNN faces long term dependency problems [13]
devices with sensors such as accelerometer and gyroscope has when the sequence is long enough. Note that, the sequence
created the opportunity to organize considerable amount of information found in HAR data is usually long. So, it is
sensor data for classification o f h uman a ctivity. T he research necessary to capture the long term dependency information
activities focused on HAR incorporates the compilation of for better classification.
sensor readings into sequential time-series data and develops Gated Recurrent Unit (GRU) is a variant of RNN which
models for recognition of activities by analyzing the acquired incorporate long term dependency information [14]. It is
sequential sensor readings. HAR poses a variety of promising expected that GRU will perform better in case of HAR data
application domain which includes physical activity annotation as it is able to capture temporal context of sensor data. It is
in the field o f m edical d ata a nalysis [ 1], p ersonal assistant noteworthy to mention here that, all temporal context are not
system [2], augmented and virtual reality [3] and many others. equally important for classification, some are more important
In the past years, the state-of-the-art solutions to HAR than others. Hence, it is necessary to give more attention to the
mainly based on the traditional machine learning techniques. important temporal context than others. Moreover, during the

c
978-1-7281-3445-1/19/$31.00 2019 IEEE
acquisition of HAR data, usually the training data found for Recurrent Neural Networks (RNN) based appoach proposed
different activities are not equal which causes class imbalance in [7] to recognize human activities and abnormal behaviour,
problem. The emphasis on important temporal context also shows some promise but leaves room for improvements.
helps to solve this problem. When the sequence is long, RNN faces log term dependency
To capture the nature of different continuous movements problems. To solve this problem, a combination of convnet and
and to extract the salient features, in this paper, we propose long short term memory [10] is used in [24] that outperforms
an attention mechanism based GRU model architecture. This other models on the KTH dataset. Gated Recurrent Units [25]
architecture plays a crucial role in capturing context of the are a variant of RNN that also addresses long term dependency
sensor reading and extracts the class imbalance tolerance issues. Adam optimizer [26] is a popular choice for training
characteristics. The main contributions of this paper are as such neural networks.
follows: Attention mechanism, introduced for sequence to sequence
• We propose to use a hierarchical temporal attention with tasks tasks such as neural machine translation [27], speech
GRU for capturing important temporal contexts. recognition [28] has also been used for classification tasks in
• The hierarchical model propose to use here is paralleliz- the domain of natural language processing [29]. The context
able. vector computed by attention helps the network to learn where
• The model is able to handle class imbalance problem. to focus on the representation generated by the encoder part for
The paper describes related work in section II where differ- generating output sequence at each time step instead of com-
ent approaches for recognition of human activity are discussed. pressing entire sequence to a fixed vector at once. Simplified
Section III has been used for describing the proposed method- form of attention mechanism [30] has been proposed for feed
ology. Section IV contains the results as well as interpretations forward network which captures some long term dependencies.
of the outcome of the proposed method. Section V concludes The approaches described so far for activity recognition fail
the paper. to capture the temporal context of sensor reading at different
time steps of activity data which is required for better accuracy
II. R ELATED W ORK and generalization. Another approach proposed by [31] uses
The most referred work proposed in [15] use fast Fourier attention mechanism on top of a complex DeepConvLSTM
transform algorithm for feature extraction, useful in recog- architecture for finding relevant temporal context for activity
nizing different activities, that produce satisfactory results recognition. In this work attention score is generated by
with numerous sensors set on distinctive parts of the body applying attention after convolutional and pooling layers in
in conjunction with various data mining algorithms. Different DeepConvLSTM. This score does not reflect the hierarchy of
approaches like K-nearest neighbors [16], decision trees [17], simple features detected from raw sensor data and complex
multi-class support vector machine [18] are used to classify features detected from hidden state outputs in case of RNN (or
human activities. All of these approaches require the use of from deeper layers in CNN). For finding relevant features for
hand-crafted features and show poor results for classifying in activity recognition, feature selection approaches used in [32],
similar type of activities like walking down and walking up. [33] can be applied. However, In this work, we propose an
In the modern age, deep learning has become prominent in attention mechanism with GRU which distills more complex
the area of learning models that represent features from low- features that would be helpful for better classification.
level to high-level abstraction used in [4], [5] which allow
to extract features automatically without hand-crafted feature III. P ROPOSED M ETHOD
engineering. A common form of neural network called fully The proposed method combines several building blocks for
connected neural network (FCNN) with Principal Component constructing the network. We use Gated Recurrent Units, two
Analysis based feature technique is used in [8] and [9] for different types of attention mechanism which are described in
HAR and sensor data. But FCNN is very expensive in terms the following sections.
of memory (weights) and computation (connections). It also
has a great chance of overfitting problem as every node is A. Gated Recurrent Unit
connected with every node in every layer. To extract additional
GRUs are a variant of Recurrent Networks that have been
features a new technique called Shift-invariant sparse coding
shown to be able to capture long term dependencies in tem-
[9] was proposed and used in combination with FCNN and
poral data while not suffering from similar vanishing gradient
handcrafted features. Convolutional neural network (CNN or
problem as regular RNN and requiring fewer parameters than
convnet) [19] with dropout [20] for reducing overfitting is a
LSTM. Hidden state h<t> calculation is based on (3) with the
recent breakthrough for feature extraction. It is used by [21] in
input vector X<t> and previous hidden state h<t−1> going
gesture recognition that give state-of-the-art result. A hierar-
through update and reset gates in (1) and (2).
chical model using convnets is proposed in [22]. To recognize
human activity for unlabeled as well as labeled data, [23] used Z<t> = σ(Wzx · X<t> + Wzh · h<t−1> + bz ) (1)
semi-supervised convnet model to learn discriminative hidden
features. Where convnet learns to recognize features of an
object and combine these features to recognize larger object. Γt = σ(WΓx · X<t> + WΓh · h<t−1> + bΓ ) (2)
 Fig. 1: Stacked 2 layer GRU model with simplified and context sensitive attention mechanism

2) Simplified Attention: Sensor data of human activities

differ from natural language data in that words are context
h<t> = (1 − Z<t> )  h<t−1> + dependent in most cases where sensor data may not depend
Z<t>  tanh(Whx · X<t> + on the context. Simplified attention does not have to learn the
Whh · (h<t−1>  Γt ) + bh ) (3) es parameter.
e<ti > = tanh (Was · h<ti > + bas ) (7)
Here, σ is used in (1) and (3) which refers to sigmoid
function and  in (3) represents element-wise multiplication. exp(e<ti > )
α<ti > = P (8)
t exp(e<ti > )
B. Attention Mechanism for HAR
Context vector is obtained using (6) using the relative
Context vector generated by attention allows a classification weight α<ti > obtained from (8)
model to give varying levels of importance to the different
temporal features generated by a recurrent network. Different C. Proposed Model Architecture
variations of attention mechanisms such as simplified attention We propose a two layer stacked GRU architecture with
and hierarchical context based attention could be used for this attention applied to the hidden state outputs of each recurrent
type of tasks. These two types of attention could be used layer. The stacked layers help to learn more complex features
separately or a combination of their learned attention scores from the sensor data. The attention scores from both layers
may be used. We take the latter approach. are concatenated to create a hierarchy context vectors before
1) Attention With Context: Because of susceptibility of sen- feeding it to the densely connected layers. Both simplified and
sor data to noise, context of a sensor value in relation to data context-based attention scores are computed independently.
on other time steps are helpful for creating a representation Batch normalization is used before applying attention and
of data for classification. By using attention on both layers after the concatenation of the the attention scores. There
of stacked GRU, the model is able to exploit hierarchy of are three fully connected layers after attention module. The
temporal features in the sensor data. first two layers are used to learn respective weights for the
different types of features obtained from the attention modules.
e<ti > = tanh (Wac · h<ti > + bac ) (4) Rectified Linear Unit (ReLU) is used for the activation of
these layers. Dropout is also performed for regularization with
exp(eT<ti > · es ) probability d1 and d2 . Dropout prevents a neural network
α<ti > = P (5)
t exp(e<ti > es ) from becoming heavily dependent on a specific weight of
X a single neuron by turning off a fraction of the neurons
c<i> = α<ti > h<ti > (6) randomly during training. For classifying human activities,
t
softmax activation is applied to the final layer. The model
Wac and bac in (4) are parameters to be learned. es in (5) architecture is illustrated in Fig. 1. We train the model with
allows the preservation of context information and is learned learning rate α and decay factor λ.
jointly when training the network. A summation of the relative Note that, a simpler model using a single GRU layer
weights of the time steps is generated as the context vector in with same attention mechanism may also be used. Such a
(6). model requires fewer parameters to learn and resulting in
less computational complexity. However, such simpler model We also performed five fold training (in each case, we
may result in lower performance. In the proposed model consider 17 individuals information of all classes except the
architecture, sequential computation overhead may be reduced imbalanced class) and test (in each case, test data is fixed
by computing the attention scores as well as the hidden states with 9 individuals as mentioned in the dataset description
for the second GRU layer in parallel. In this case we can subsection) with 50 percent dropout of the training data of
achieve better performance with much less time. a specific class that is considered as an imbalanced class.
We use attention mechanism with GRU. Attention score
may be calculated on derived features of CNN layers which
may lead to the loss of context information. In comparison, C. Result and Discussion
using attention mechanism on the hidden states computed
In this experiment, we have compared our proposed model
directly from sensor data facilitates better learning of the
architectures with a state-of-the-art baseline CNN based
temporal contexts. In our proposed model, we use two types
method [6]. The column two of TABLE I presents the compar-
of hierarchy: firstly, hierarchy with two different types of
ison according to the experimental setup described in the HAR
attention (e.g. simplified and with context) and secondly, at-
dataset. It is observed from this table that in terms of accuracy,
tentions from multiple layers of GRU. Using such a hierarchy
our proposed stacked GRU with attention performs better from
of attention allows the incorporation of less complex features
both the proposed simplified model (with GRU + Attention)
(low level information) with more complex features (high level
and the baseline CNN based method. Such an improvement is
representation of features) required for classification.
expected as we consider hierarchy in the attention model and
IV. E XPERIMENT R ESULT AND D ESCRIPTION two layer stacked GRU.
A. Dataset Description We have demonstrated the impact of class imbalance in
TABLE I (Column 3-8) and TABLE II for different methods.
In this experiment, we use the benchmark HAR dataset In both cases (1/2 and 2/3 class drop), our stacked GRU with
[34] which provides time series information of six differ- attention wins for all classes in terms of accuracy. Besides
ent activities(walking, walking upstairs, walking downstairs, these, to test the generalization ability of the proposed method,
sitting, standing and laying). Two types of sensor namely we perform five fold training and test as mentioned in the im-
accelerometer and gyroscope are used to capture these infor- plementation detail subsection. These results is demonstrated
mation and randomly partitioned into 70% train and 30% test in TABLE III. In this case, we consider AUC as a performance
sets. Data has been preprocessed with noise-filters and sampled metric as it performs better for imbalance dataset [36]. From
in fixed length sliding windows of 2.56 sec and 50% overlap this table, it is also observed that when imbalanced is injected,
which yields 128 readings per window. The training data is the proposed method wins for every classes individually which
generated with a total of 7352 examples incorporating data of demonstrated the superiority of the proposed method.
21 randomly selected individuals. On the other hand, test set
Note that, the improvement of the proposed method with the
is composed of 2947 examples incorporating the remaining 9
baseline CNN model observed in the aforementioned tables is
individual subjects selected for this specific dataset.
not that significant. The reason for such slight improvements
B. Implementation Detail can be explained by the simplicity of the dataset. In the dataset,
the activities are separable by a large margin which is shown in
During training, Adam optimizer is used with its default
Fig. 4. The figure is generated with dimension reduction using
parameters (β1 =0.9 and β2 =0.999) for backpropagation. More-
t-SNE [37]. So that when we create class imbalance,there is
over, the initial α is set to 0.001 which is presented in [35]
no significant drop in accuracies for both of the methods and
with λ = 0.2 ( decay based on validation loss). The dropout
the difference between these two methods in terms of accuracy
probability d1 and d2 are set to 0.25 and 0.1 respectively in
also remains small. We believe, the improvement will be much
the fully connected layers.
larger for more challenging dataset.
The standard accuracy metric defined in 9 is used to evaluate
the methods used in this experiment. We demonstrated the confusion matrix of the proposed
method and baseline CNN in Fig. 2 and 3 respectively. The
TP + TN confusion matrix indicates that misclassification occurs rarely
Accuracy = (9)
TP + TN + FP + FN and also within the vicinity of ground truth class even in
where, TP, TN, FP and FN are defined as True Positive, True imbalanced class scenario. However, from these two figures,
Negative, False Positive and False Negative respectively. it is observed that the proposed method has fewer misclasssi-
To evaluate the performances of the methods in the class fication compare to the baseline CNN.
imbalance scenario, we first drop half of the training data Based on the results in this experiment, it is evident that
and second drop two-third of the training data. In both the proposed method is better in terms of accuracy and AUC
cases, we keep unchanged test data (mentioned in the Dataset metric. This method is able to produce such result by giving
Description subsection). For measuring the performance of the the attention to both simple and complex features which
methods in the class imbalance scenario, we use area under are learned from the proposed combination of two attention
the curve (AUC) as an evaluating metric in this case. mechanisms.
TABLE I: 1/2 Drop of Specific Class Data in Training Set and Measurement using given Test Data in terms of accuracy (%)
Model Architecture Performance (Accuracy %) Class - 1 Class - 2 Class - 3 Class - 4 Class - 5 Class - 6
CNN 94 .022 89.37902952 88.49677638 90.02375297 90.6345436 92.33118426 92.09365456
GRU + Attention 93.79 90.90600611 92.50084832 90.77027486 92.195453 92.36511707 92.05972175
Stacked GRU (2 Layers)
94.16355 92.02578894 91.78825925 92.53478113 91.9239905 93.07770614 94.46895148
+ Attention

TABLE II: 2/3 Drop of Specific Class Data in Training Set and Measurement using given Test Data in terms of accuracy (%)
Model Class - 1 Class - 2 Class - 3 Class - 4 Class - 5 Class - 6
Baseline (CNN) 90.56667798 89.68442484 89.41296233 89.88802172 91.44893112 91.58466237
2-Stacked GRU
90.73634204 91.38106549 92.26331863 92.02578894 92.56871395 93.31523583
+ Attention

TABLE III: Subject-wise Class Drop in K-Fold and Measurement using given Test Data in terms of AUC
Model Architecture k - fold Class - 1 Class - 2 Class - 3 Class - 4 Class - 5 Class - 6
0 0.989897104 0.988005576 0.989058376 0.990233114 0.986666755 0.992600584
1 0.982893556 0.986482948 0.984618942 0.981161394 0.99192285 0.984878402
2 0.980885725 0.987111175 0.98570861 0.987078789 0.99061635 0.983950708
Baseline (CNN)
3 0.989421716 0.988371167 0.984282503 0.983464878 0.99116295 0.982648438
4 0.990690272 0.990301677 0.99302656 0.986080276 0.987868021 0.989988681
Average 0.9867576746 0.9880545086 0.9873389982 0.9856036902 0.9896473852 0.9868133626
0 0.9891089821 0.9917239879 0.9937438938 0.9885901347 0.9892608509 0.9915665231
1 0.9898042512 0.9845693577 0.9922561365 0.9879438003 0.9901914546 0.9906566782
Stacked GRU (2 Layers) 2 0.992904892 0.990711794 0.992136949 0.988169278 0.991100119 0.992741053
+ Attention 3 0.988042117 0.992132516 0.990958787 0.987103942 0.989934521 0.993510194
4 0.991783724 0.993914724 0.989005982 0.992895329 0.992081285 0.992756454
Average 0.9903287933 0.9906104759 0.9916203497 0.9889404968 0.9905136461 0.9922461805

Fig. 2: Confusion Matrix when half of the training data for Fig. 3: Confusion Matrix when half of the training data for
class ’Walking Downstairs’ is dropped for stacked GRU model class ’Walking Downstairs’ is dropped for CNN model

V. C ONCLUSION in terms of activity duration and feature complexity which we

will address in future.
The proposed method performs reasonably well for the
recognition of human activity while not being sensitive to class ACKNOWLEDGMENT
imbalance in the training data as it learns temporal features
This research is supported by the University Grants Com-
using attention mechanism. More parrallelizable models could
mission, Bangladesh under the Dhaka University Teachers
be developed in the future by constructing embedding from
Research Grant No Reg/Admin-3/54292-94.
the temporal data. In this paper, we intend to show the
effectiveness of the proposed mechanism using a benchmark R EFERENCES
HAR dataset. We believe the proposed method will also
[1] H. Song, D. Rajan, J. J. Thiagarajan, and A. Spanias, “Attend and
perform better for more complex datasets, collected from diagnose: Clinical time series analysis using attention models,” in Thirty-
heterogeneous devices that have more varied class distribution Second AAAI Conference on Artificial Intelligence, 2018.
 with deep convolutional neural networks,” in Advances in neural infor-
mation processing systems, pp. 1097–1105, 2012.
[20] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhut-
dinov, “Dropout: A simple way to prevent neural networks from overfit-
ting,” Journal of Machine Learning Research, vol. 15, pp. 1929–1958,
06 2014.
[21] S. Duffner, S. Berlemont, G. Lefebvre, and C. Garcia, “3d gesture
classification with convolutional neural networks,” in 2014 IEEE In-
ternational Conference on Acoustics, Speech and Signal Processing
(ICASSP), pp. 5432–5436, IEEE, 2014.
[22] J. Yang, M. N. Nguyen, P. P. San, X. L. Li, and S. Krishnaswamy, “Deep
convolutional neural networks on multichannel time series for human
activity recognition,” in Twenty-Fourth International Joint Conference
on Artificial Intelligence, 2015.
[23] M. Zeng, T. Yu, X. Wang, L. T. Nguyen, O. J. Mengshoel, and I. Lane,
“Semi-supervised convolutional neural networks for human activity
Fig. 4: Visualization of the separation of activities recognition,” in 2017 IEEE International Conference on Big Data (Big
Data), pp. 522–529, IEEE, 2017.
[24] M. Baccouche, F. Mamalet, C. Wolf, C. Garcia, and A. Baskurt, “Se-
quential deep learning for human action recognition,” in International
[2] F. Montalto, C. Guerra, V. Bianchi, I. De Munari, and P. Ciampolini, workshop on human behavior understanding, pp. 29–39, Springer, 2011.
“Musa: Wearable multi sensor assistant for human activity recogni- [25] K. Cho, B. van Merrienboer, Ç. Gülçehre, F. Bougares, H. Schwenk,
tion and indoor localization,” in Ambient Assisted Living, pp. 81–92, and Y. Bengio, “Learning phrase representations using RNN encoder-
Springer, 2015. decoder for statistical machine translation,” CoRR, vol. abs/1406.1078,
[3] D. Van Krevelen and R. Poelman, “Augmented reality: Technologies, 2014.
applications, and limitations,” Vrije Univ. Amsterdam, Dep. Comput. Sci, [26] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,”
2007. CoRR, vol. abs/1412.6980, 2015.
[4] Y. Bengio et al., “Learning deep architectures for ai,” Foundations and [27] D. Bahdanau, K. Cho, and Y. Bengio, “Neural machine translation by
trends® in Machine Learning, vol. 2, no. 1, pp. 1–127, 2009. jointly learning to align and translate,” CoRR, vol. abs/1409.0473, 2015.
[5] L. Deng, “A tutorial survey of architectures, algorithms, and applications [28] J. K. Chorowski, D. Bahdanau, D. Serdyuk, K. Cho, and Y. Bengio,
for deep learning,” APSIPA Transactions on Signal and Information “Attention-based models for speech recognition,” in Advances in neural
Processing, vol. 3, 2014. information processing systems, pp. 577–585, 2015.
[6] C. A. Ronao and S.-B. Cho, “Human activity recognition with smart- [29] Z. Yang, D. Yang, C. Dyer, X. He, A. J. Smola, and E. H. Hovy,
phone sensors using deep learning neural networks,” Expert Syst. Appl., “Hierarchical attention networks for document classification,” in HLT-
vol. 59, pp. 235–244, Oct. 2016. NAACL, 2016.
[7] D. Arifoglu and A. Bouchachia, “Activity recognition and abnormal [30] C. A. Raffel and D. P. W. Ellis, “Feed-forward networks with
behaviour detection with recurrent neural networks,” Procedia Computer attention can solve some long-term memory problems,” CoRR,
Science, vol. 110, pp. 86–93, 2017. vol. abs/1512.08756, 2015.
[8] T. Plötz, N. Y. Hammerla, and P. L. Olivier, “Feature learning for activity [31] V. S. Murahari and T. Plötz, “On attention models for human activity
recognition in ubiquitous computing,” in Twenty-Second International recognition,” in Proceedings of the 2018 ACM International Symposium
Joint Conference on Artificial Intelligence, 2011. on Wearable Computers, pp. 100–103, ACM, 2018.
[32] N. X. Vinh, S. Zhou, J. Chan, and J. Bailey, “Can high-order depen-
[9] C. Vollmer, H.-M. Gross, and J. P. Eggert, “Learning features for
dencies improve mutual information based feature selection?,” Pattern
activity recognition with shift-invariant sparse coding,” in International
Recognition, vol. 53, pp. 46–58, 2016.
conference on artificial neural networks, pp. 367–374, Springer, 2013.
[33] S. Sharmin, M. Shoyaib, A. A. Ali, M. A. H. Khan, and O. Chae,
[10] S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural “Simultaneous feature selection and discretization based on mutual
Comput., vol. 9, pp. 1735–1780, Nov. 1997. information,” Pattern Recognition, vol. 91, pp. 162–174, 2019.
[11] Y. Meng and A. Rumshisky, “Context-aware neural model for temporal [34] D. Anguita, A. Ghio, L. Oneto, X. Parra, and J. L Reyes-Ortiz, “A public
information extraction,” in Proceedings of the 56th Annual Meeting of domain dataset for human activity recognition using smartphones,” 01
the Association for Computational Linguistics (Volume 1: Long Papers), 2013.
2018. [35] M. D. Zeiler, “Adadelta: an adaptive learning rate method,” arXiv
[12] W. Yin, K. Kann, M. Yu, and H. Schütze, “Comparative study of cnn and preprint arXiv:1212.5701, 2012.
rnn for natural language processing,” arXiv preprint arXiv:1702.01923, [36] C. X. Ling, J. Huang, and H. Zhang, “Auc: a better measure than
2017. accuracy in comparing learning algorithms,” in Conference of the
[13] Y. Bengio, P. Frasconi, and P. Simard, “The problem of learning canadian society for computational studies of intelligence, pp. 329–341,
long-term dependencies in recurrent networks,” in IEEE international Springer, 2003.
conference on neural networks, pp. 1183–1188, IEEE, 1993. [37] L. van der Maaten and G. Hinton, “Visualizing data using t-SNE,”
[14] J. Chung, C. Gulcehre, K. Cho, and Y. Bengio, “Empirical evaluation of Journal of Machine Learning Research, vol. 9, pp. 2579–2605, 2008.
gated recurrent neural networks on sequence modeling,” arXiv preprint
arXiv:1412.3555, 2014.
[15] L. Bao and S. S. Intille, “Activity recognition from user-annotated
acceleration data,” in International conference on pervasive computing,
pp. 1–17, Springer, 2004.
[16] W. Wu, S. Dasgupta, E. E. Ramirez, C. Peterson, and G. J. Norman,
“Classification accuracies of physical activities using smartphone motion
sensors,” Journal of medical Internet research, vol. 14, no. 5, 2012.
[17] A. M. Khan, “Recognizing physical activities using wii remote,” Inter-
national Journal of Information and Education Technology, vol. 3, no. 1,
p. 60, 2013.
[18] D. Anguita, A. Ghio, L. Oneto, X. Parra, and J. L. Reyes-Ortiz, “Human
activity recognition on smartphones using a multiclass hardware-friendly
support vector machine,” in International workshop on ambient assisted
living, pp. 216–223, Springer, 2012.
[19] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification