Conference PaperPDF Available

Abstract and Figures

This work presents an image classification approach to EEG brainwave classification. The proposed method is based on the representation of temporal and statistical features as a 2D image, which is then classified using a deep Convolutional Neural Network. A three-class mental state problem is investigated, in which subjects experience either relaxation, concentration, or neutral states. Using publicly available EEG data from a Muse Electroencephalography headband, a large number of features describing the wave are extracted, and subsequently reduced to 256 based on the Information Gain measure. These 256 features are then normalised and reshaped into a 16 × 16 grid, which can be expressed as a grayscale image. A deep Convolutional Neural Network is then trained on this data in order to classify the mental state of subjects. The proposed method obtained an out-of-sample classification accuracy of 89.38%, which is competitive with the 87.16% of the current best method from a previous work.
Content may be subject to copyright.
Classification of EEG Signals Based on Image
Representation of Statistical Features
Jodie Ashford*, Jordan J. Bird*, Felipe Campelo, and Diego R. Faria
School of Engineering and Applied Science
Aston University, Birmingham, UK
{ashfojsm, birdj1, f.campelo, d.faria}
Abstract. This work presents an image classification approach to EEG
brainwave classification. The proposed method is based on the repre-
sentation of temporal and statistical features as a 2D image, which is
then classified using a deep Convolutional Neural Network. A three-class
mental state problem is investigated, in which subjects experience ei-
ther relaxation, concentration, or neutral states. Using publicly avail-
able EEG data from a Muse Electroencephalography headband, a large
number of features describing the wave are extracted, and subsequently
reduced to 256 based on the Information Gain measure. These 256 fea-
tures are then normalised and reshaped into a 16 ×16 grid, which can
be expressed as a grayscale image. A deep Convolutional Neural Net-
work is then trained on this data in order to classify the mental state of
subjects. The proposed method obtained an out-of-sample classification
accuracy of 89.38%, which is competitive with the 87.16% of the current
best method from a previous work.
Keywords: Machine Learning, Convolutional Neural Networks, Image
Recognition, Mental State Classification, Electroencephalography
1 Introduction
Human-machine interaction is often considered a mirror of the human experi-
ence; sound and visuals constitute voice recognition, human activity classifica-
tion, facial recognition, sentiment analysis and so on. Though, with the avail-
ability of sensors to gather data that the human body cannot, interaction with
machines can often exceed the abilities of the natural human experience. An
example of this is the consideration of electroencephalographic brainwaves. The
brain, based on what a person is thinking, feeling, or doing, has a unique pat-
tern of electrical activity that emerges as a consequence of the aggregate firing
patterns of billions of individual neurones [1, 2]. These electrical signals can, in
principle, be detected and processed to infer the state of the brain and, by ex-
tension, the mental state of a given subject. Besides clinical applications, this
possibility is also useful, e.g., for brain-machine interfacing.
* JA and JJB are co-first authors.
2 Ashford, Bird et al.
More effective methods of feature extraction and classification are of utmost
importance in brain-machine interaction, since better performing models can in-
terpret human brain activity with higher accuracy. Previous works [3, 4] suggest
that static statistical descriptions of brainwaves present the information in the
signals in a more machine learning-friendly shape than the raw waves themselves,
even when temporally-aware machine learning methods are employed.
This work focuses on the process of feature extraction, selection and format-
ting in order to achieve improved classification accuracy of EEG signals. More
specifically, the main contribution is a framework to perform classification of
these signals, based on (i) the extraction of a large number of static statistical
features of the data, followed by (ii) automated feature selection and (iii) repre-
sentation of the selected attributes as a 2D matrix. The resulting matrices are
(iv) interpreted as grayscale images, which allows the leveraging of the state-of-
the-art performance of convolutional neural networks [5, 6] as image classifiers.
The remainder of this paper is organised as follows. A brief presentation of
the background concepts related to the present work is provided in Section 2,
followed by the description of the proposed approach in Section 3. The results
obtained by the proposed method are discussed in Section 4. Finally, conclusions
and suggestions of further investigations are provided in Section 5.
2 Background
Electroencephalography (EEG) is a technique used to measure the electrical
activity of a brain. The human brain contains billions of neurones, which each
exhibit electrical activity in the form of nervous impulses [7, p. 31]. The electrical
signal produced by a single neurone is difficult to detect, but the combined signal
from the action of many neurones together can be measured using EEG [8, p. 4].
Typically, EEG involves placing electrodes onto the scalp of the subject.
These electrodes measure the voltage fluctuations generated by thousands of
active neurones in the brain. These signals are then digitised and amplified [9,
10]. Possibly the main advantages of this method of measuring brain activity are
that it is a non-invasive and inexpensive technique. Even less invasive techniques,
such as the Muse headband, have extended the utility of EEG beyond medical
examination alone, at the expense of sensitivity. Unlike imaging techniques such
as MRI, EEG can measure fluctuations in electrical activity on the scale of
milliseconds, which makes it an incredibly powerful tool for measuring real-time
brain activity in response to stimulus [10].
The ability to infer human mental states is as important in human-machine
interaction as a form of Affective Computing[11] as it is in natural human inter-
action. In the past, such techniques have used attributes available to humans:
speech, gestures, facial expressions, etc. [12, 13]. With the increasing develop-
ment of non-invasive EEG technology, researchers can take advantage of sensors
available only to machines to attempt to classify human emotions directly from
the brain. Such analysis is less dependent on environmental factors or differences
Classification of EEG Signals Based on Image Representation 3
in somatic expression between individuals. It also offers a more seamless avenue
of human-machine communication.
2.1 Related Work
Non-invasive EEG headsets have been used in previous works to analyse mental
states. In a related example, data from the Muse headset was was found to be
useful in the evaluation of the enjoyment levels of subjects playing two different
video games [14]. The findings aligned with the current understanding of how
waves detected by EEG (in this case, frontal theta frequencies) map to enjoy-
ment. This is an example of how non-invasive techniques can provide uses of EEG
outside of the medical setting, and provide data for emotional classification.
Previous works have also shown the excellent performance of convolutional
neural networks (CNN) in EEG-based mental state classification. In 2017, us-
ing the DEAP dataset [5], EEG signals were classified using both deep (DNN)
and convolutional (CNN) neural networks [6]. Two different classifications were
performed: one for valence and one for arousal, classifying each as either high
or low. The DNN achieved 75.78% accuracy for valence and 73.28% for arousal,
while the CNN achieved 81.41% and 73.35% respectively [6].
Projection of EEG data onto a “visual” space is a fairly recent approach, with
relatively little work as of yet performed into its exploration. Most of the relevant
literature in this area [15, 16] relates to mapping the signal readings of electrodes
to a spatial representation of the brain itself, interpolating intermediary points
based on values from the nearest electrodes. Alternatively, some limited but
successful work has explored the CNN classification of visual spectrograms pro-
duced by the signals[17]. Spectrograms produced by Limited Field Potentials
have also found varying levels of success in classifying biological signals from rat
brains[18, 19]. In these works, a limited set of five features were extracted and
machine learning approaches (decision trees, discriminant analysis, support vec-
tor machines, and nearest neighbour classifiers) were used to recognise patterns,
producing results with accuracies ranging from 95.8% to 98.8%. These solutions,
though effective, rarely consider statistical processing of the waves as a way to
extract relevant data from the complex waveforms generated by EEG. Visual
pixel-wise approaches and subsequent CNN applications have been successfully
implemented in other biological domains such as image segmentation of electron
microscopy images [20], with promising results for a variety of applications.
The solution suggested in this study, on the other hand, is based on extract-
ing statistical features from EEG signal waves and maps them onto static 2D
matrices, which are then represented as images and used for the classification of
mental states using a convolutional neural network. This proposed methodology
is detailed in the following section.
3 Proposed Approach
Firstly, an available training set of EEG signals is preprocessed. The data is
assumed to contain the time series related to one or more electrodes, within a
4 Ashford, Bird et al.
given experimental time frame, labelled in terms of three distinct mental states
(relaxed, concentrating, and neutral) that the subjects were keeping during data
collection [4]. From these signals a number of statistical features are extracted [3,
4], resulting in a high dimensional attribute space - in the case of this work, 1274
features are generated for each time window, as detailed in Section 3.1. To focus
only on the most relevant ones for the classification process, feature selection
is applied to the resulting features. Here, the 162= 256 most descriptive ones,
based on the estimated information gain [21], are selected.
Finally, the selected features are converted into a 16 ×16 grid of numerical
values normalised to the [0,1] range, which can be represented as a grayscale
image. Figure 1 shows a number of samples of relaxed, neutral and concentrating
brainwave data, using this particular image representation.
(a) Ten samples of relaxed brainwave data represented as 16 ×16-pixel images
(b) Ten samples of concentrating brainwave data represented as 16 ×16-pixel images
(c) Ten samples of neutral brainwave data represented as 16 ×16-pixel images
Fig. 1: Examples of image representations for each of the three mental states
considered in this work.
The resulting set of images is then used to train a convolutional neural net-
work (CNN) [22] as a classifier of the three mental states investigated in this
particular work. The details of the CNN are provided in Section 3.2.
3.1 Feature Extraction
Due to the temporal, auto-correlated nature of the EEG waves, single-point fea-
tures cannot generally provide enough information for good rules to be generated
by machine learning models. In this work we follow the approach of extracting
statistical features based on sliding time windows [3, 4]. More specifically, the
EEG signal is divided into a sequence of windows of length one second, with
consecutive windows overlapping by 0.5 seconds, e.g., [(0s1s), [0.5s1.5s),
[1s2s), . . . ]).
Classification of EEG Signals Based on Image Representation 5
Assume that each 1-second time window contains a sequence x= [x1, . . . , xN]
composed of Nsamples. Also let xh1and xh2denote the first and second halves
of the window, and xq1,xq2,xq3,xq4denote the four quarter-windows obtained
by dividing the window into four (roughly) equal-sized parts, each composed of
approximately N/4 samples.1
In this work the following statistical features were generated for each time
Considering the full time window:
The sample mean and sample standard deviation of each signal (8 fea-
The sample skewness and sample kurtosis of each signal [23] (8 features).
The maximum and minimum value of each signal (8 features).
The sample variances of each signal, plus the sample covariances of all
signal pairs [24] (10 features).
The eigenvalues of the covariance matrix [25] (4 features).
The upper triangular elements of the matrix logarithm of the covariance
matrix [26]. (10 features)
The magnitude of the frequency components of each signal, obtained
using a Fast Fourier Transform (FFT) [27] (300 features).
The frequency values of the ten most energetic components of the FFT,
for each signal (40 features).
Considering the two half-windows:
The change in the sample means and in the sample standard deviations
between the first and second half-windows, for all signals (8 features).
The change in the maximum and minimum values between the first and
second half-windows, for all signals (8 features).
Considering the quarter-windows:
The sample mean of each each quarter-window, plus all paired differ-
ences of sample means between the quarter-windows, for all signals (56
The maximum (minimum) values of each quarter-window, plus all paired
differences of maximum (minimum) values between the quarter-windows,
for all signals (112 features).
Regarding the representation of the signals in the frequency domain using
FFT [27], two specific aspects were taken into account: first, the DC-component
of the signals was filtered out prior to the application of the FFT, so the zero-
frequency component was always set as zero. This was done to prevent the offset
to completely dominate the power spectrum, even though it carries no relevant
information for the classification task. The second aspect is that frequencies in
the range of (50 ±1) Hz were also filtered out, to remove any contamination
from the AC electrical distribution frequency, which could also skew the power
spectrum of our signals.
1In this work we standardised the number of samples within each window to N= 150.
This means that quarter-windows have either n= 37 or n= 38 observations.
6 Ashford, Bird et al.
Each window receives as features the vector of quantities computed above
for both itself and the window that immediately precedes it (1-lag window).
Features from the 1-lag window that were clearly redundant due to the half-
window overlaps were removed prior to the composition of the feature vector,
namely the sample means, maximum and minimum values of xq3and xq4, as well
as their respective differences. In the end a total of 989 features were generated
for each time window (except the first, which was only used as the 1-lag for the
second one).
After the statistical features were extracted the resulting dataset was com-
posed of 2479 data objects, each represented by its corresponding 989 feature
values plus a single class label. Feature selection was then performed based
on the Information Gain of each feature, and the total number of features
was reduced to 256 (plus class label). Due to privacy considerations the raw
EEG data cannot be released, but the processed dataset is publicly available
at as a UTF-8 encoded
CSV with approximately 6MB.
3.2 Convolutional Neural Network
Convolutional neural networks (CNN) [28] are a specialised kind of neural net-
work for processing data that has a known grid-like topology, which makes them
particularly suitable for dealing with data represented as time series or images
[22]. The main distinguishing feature of these networks is the use of a convolu-
tion [29] instead of simple matrix multiplication in at least one of their layers
[22]. Convolutional neural networks are generally very effective at image classi-
fication tasks [30–32], which motivates their use here. For more details on these
networks, please refer to Ian Goodfellow et al.’s book on the subject [22].
In this particular work we have opted for using the CNN implementation
available in the Keras Deep Learning Python library [33]. The network was
trained on an Nvidia GTX1060 (1280 CUDA Cores, 6GB 8Gbps GDDR5 VRAM).
The topology and hyperparameters of the convolutional neural network were de-
fined based on preliminary, trial-and-error experimentation. Table 1 shows the
resulting model for the classification of brainwave images.
Other design choices that were arbitrarily set in this experiment are the use
of the ADAM optimiser [34] to train the network; and the use of a batch size of
100, trained for 400 epochs, with the loss calculated via categorical cross entropy
at a 70/30 validation split:
CE =
yo,c log(po,c),(1)
where Mis the number of class labels (in this case, 3), yis a binary indication
of a correct prediction (1 or 0), and pis the predicted probability of observation
oof class c. The entropy of each class within the testing split is calculated and
added for a final, overall result. In this case, this is the entropy of the three
classes of mental state - relaxed, neutral, and concentrating.
Classification of EEG Signals Based on Image Representation 7
Table 1: Network topology and parameters used. Please refer to the Keras doc-
umentation [33] for specific definitions.
Layer Output Params
Conv2d (ReLu) (0, 14, 14, 32) 320
Conv2d (ReLu) (0, 12, 12, 64) 18496
Max Pooling (0, 6, 6, 64) 0
Dropout (0.25) (0, 6, 6, 64) 0
Flatten (0, 2304) 0
Dense (ReLu) (0, 512) 1180160
Dropout (0.5) (0, 512) 0
Dense (Softmax) (0, 3) 1539
4 Results
In this section the results for the experiments are presented. The experiments
were performed three times, the difference between the three runs being random
seeds set at the start of the experiment. The overall final score always resulted
within 400 epochs. Accuracy and loss per-epoch are illustrated for the first run.
4.1 Results obtained
Figure 2 illustrates the accuracy and loss of the network, for both training and
testing data from the validation split. The overall out-of-sample accuracy of
the CNN in classifying the dataset was 89.38% (665/744 correct classifications,
CI0.95 = [86.94,91.50] %). As can be observed, the accuracy curve saturates af-
ter about 50 epochs, after which the loss starts increasing. This suggests that
computational resources are essentially wasted after this point, and more parsi-
monious training can be employed in the future.
Table 2: Comparison with Related Studies using the Same Dataset as this Exper-
iment. Column Accuracy also provides 95% confidence intervals for the accuracy.
Study Method Validation Focus Accuracy
This study Inf. Gain Selection,
CNN 70/30 Split Accuracy 89.38%
[3] OneR Selection,
Random Forest 10-fold Accuracy 87.2%
[35] Evol. Selection,
DEvoMLP 5-fold Accuracy,
Resource Usage
Table 2 contrasts the results obtained in this paper with previous works
dealing with the same mental state dataset. It is worth mentioning that only
8 Ashford, Bird et al.
Fig. 2: (a) Accuracy and (b) Loss of the CNN for Training and Testing Data
across 400 epochs.
Classification of EEG Signals Based on Image Representation 9
one of the compared experiments had the single goal of maximising accuracy,
while the other was also focused on minimising computational effort. Another
noteworthy point is that both previous works used cross-validation instead of a
split set in order to estimate accuracy. With these factors in mind, the approach
used in the present work has provided results that seem to be very competitive,
with a point estimate of the accuracy that is approximately 2.18% greater (in
absolute terms) than the one reported in the 2018 study. This difference is not,
however, statistically significant at the 95% confidence level (p= 0.129 using
the chi-squared test for equality of two binomial proportions [24]). Despite not
clearly outperforming the current state of the art, this result suggests that the
proposed approach of coupling an image-based representation of the data with
CNN-based classification may represent an effective new strategy for performing
EEG classification, with potential extensions to classification in the context of
general time series data.
5 Future Work and Conclusions
In this work, a new approach for classification of EEG signals has been presented,
based on the sequential application of statistical feature extraction and selection,
normalisation and subsequent projection of the selected features as small images,
and classification based on a convolutional neural network. The results obtained
for this method have been shown to be very competitive with the best known
results to date for the available dataset.
Possibly the most clear limitation of the present work is related to the ques-
tion of generality. Since a single dataset is used, it remains to be seen how well
the proposed methodology generalises not only to larger, possibly richer EEG
data, but also - and more interestingly - to other similar time series. In this
regard, further testing and statistical assessment of the proposed methodology
are fundamental next steps as this line of research progresses.
Due to the limited available resources, the experiment reported in this work
used a simple 70/30 data split instead of the more usual (but more computa-
tionally demanding) cross-validation, which should be used in future experiments
whenever possible so as to obtain better estimates of out-of-sample accuracy[36].
Two other aspects related to the issue of limited resources were present. The first
was the lack of a principled parameter tuning approach for both the structure and
other parameters of the network, which can be optimised using, e.g., iterated rac-
ing [37], hyperheuristics [38], or topology-specific tuning methods [39–41]. Even
under more constrained computational budgets, traditional design and analysis
of experiments approaches [42] can be useful in defining the best network for
this particular problem. The second issue is related with the selection of only
256 features to compose the image to be used in the training of the CNN. Future
work in this direction should concern the testing of varying image sizes in order
to better fine-tune the attribute selection process. In addition, further methods
of feature extraction should be investigated and compared, rather than focusing
solely on Information Gain as this study has done. The investigation of other
10 Ashford, Bird et al.
CNN architectures, which have shown much promise in other contexts [43], is
also an interesting point for further development.
Regardless of the possible improvements discussed above, we argue that the
proposed framework of projecting selected features onto a 2D matrix and subse-
quent image recognition through a Convolutional Neural Network already con-
stitutes a competitive approach for brainwave data classification. The results
obtained are promising, as compared to current scientific standards, and further
exploration is strongly suggested to advance the results beyond the preliminary
outcome presented in this paper.
1. R. Caton, “The electric currents of the brain,” American Journal of EEG Tech-
nology, vol. 10, no. 1, pp. 12–14, 1970.
2. R. R. Llin´as, “Intrinsic electrical properties of mammalian neurons and cns func-
tion: a historical perspective,” Frontiers in cellular neuroscience, vol. 8, p. 320,
3. J. J. Bird, L. J. Manso, E. P. Ribiero, A. Ekart, and D. R. Faria, “A study on mental
state classification using eeg-based brain-machine interface,” in 9th International
Conference on Intelligent Systems, IEEE, 2018.
4. J. J. Bird, A. Ekart, C. D. Buckingham, and D. R. Faria, “Mental emotional
sentiment classification with an eeg-based brain-machine interface,” in The Inter-
national Conference on Digital Image and Signal Processing (DISP’19), Springer,
5. S. Koelstra, C. Muhl, M. Soleymani, J.-S. Lee, A. Yazdani, T. Ebrahimi, T. Pun,
A. Nijholt, and I. Patras, “Deap: A database for emotion analysis; using physio-
logical signals,” IEEE transactions on affective computing, vol. 3, no. 1, pp. 18–31,
6. S. Tripathi, S. Acharya, R. D. Sharma, S. Mittal, and S. Bhattacharya, “Using
deep and convolutional neural networks for accurate emotion classification on deap
dataset.,” in Twenty-Ninth IAAI Conference, 2017.
7. D. Purves, G. Augustine, D. Fitzpatrick, W. Hall, A. LaMantia, J. McNamara,
and S. Williams, Neuroscience. Sinauer Associates, 2004.
8. J. W. Britton, L. C. Frey, J. Hopp, P. Korb, M. Koubeissi, W. Lievens, E. Pestana-
Knight, and E. L. St, Electroencephalography (EEG): An introductory text and
atlas of normal and abnormal findings in adults, children, and infants. American
Epilepsy Society, Chicago, 2016.
9. G. Buzs´aki, C. A. Anastassiou, and C. Koch, “The origin of extracellular fields
and currentseeg, ecog, lfp and spikes,” Nature reviews neuroscience, vol. 13, no. 6,
p. 407, 2012.
10. M. X. Cohen, Analyzing neural time series data: theory and practice. MIT press,
11. R. W. Picard, Affective computing. MIT press, 2000.
12. M. Pantic and L. J. Rothkrantz, “Toward an affect-sensitive multimodal human-
computer interaction,” Proceedings of the IEEE, vol. 91, no. 9, pp. 1370–1390,
13. P. V. Rouast, M. Adam, and R. Chiong, “Deep learning for human affect recogni-
tion: Insights and new developments,” IEEE Transactions on Affective Computing,
Classification of EEG Signals Based on Image Representation 11
14. M. Abujelala, C. Abellanoza, A. Sharma, and F. Makedon, “Brain-ee: Brain enjoy-
ment evaluation using commercial eeg headband,” in Proceedings of the 9th acm
international conference on pervasive technologies related to assistive environments,
p. 33, ACM, 2016.
15. P. A. Abhang and B. W. Gawali, “Correlation of eeg images and speech signals for
emotion analysis,” British Journal of Applied Science & Technology, vol. 10, no. 5,
pp. 1–13, 2015.
16. A. Gevins, M. E. Smith, L. McEvoy, and D. Yu, “High-resolution eeg mapping
of cortical activation related to working memory: effects of task difficulty, type of
processing, and practice.,” Cerebral cortex (New York, NY: 1991), vol. 7, no. 4,
pp. 374–385, 1997.
17. X. Zhang and D. Wu, “On the vulnerability of cnn classifiers in eeg-based bcis,”
IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 27,
no. 5, pp. 814–825, 2019.
18. X. Wang, M. Magno, L. Cavigelli, M. Mahmud, C. Cecchetto, S. Vassanelli, and
L. Benini, “Embedded classification of local field potentials recorded from rat barrel
cortex with implanted multi-electrode array,” in 2018 IEEE Biomedical Circuits
and Systems Conference (BioCAS), pp. 1–4, IEEE, 2018.
19. X. Wang, M. Magno, L. Cavigelli, M. Mahmud, C. Cecchetto, S. Vassanelli, and
L. Benini, “Rat cortical layers classification extracting evoked local field potential
images with implanted multi-electrode sensor,” in 2018 IEEE 20th International
Conference on e-Health Networking, Applications and Services (Healthcom), pp. 1–
6, IEEE, 2018.
20. M. Mahmud, M. S. Kaiser, A. Hussain, and S. Vassanelli, “Applications of deep
learning and reinforcement learning to biological data,” IEEE transactions on neu-
ral networks and learning systems, vol. 29, no. 6, pp. 2063–2079, 2018.
21. P.-N. Tan, Introduction to data mining. Pearson Education India, 2018.
22. I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning. MIT Press, 2016.
23. D. Zwillinger and S. Kokoska, CRC Standard Probability and Statistics Tables and
Formulae. Chapman & Hall, 2000.
24. D. C. Montgomery and G. C. Runger, Applied Statistics and Probability for Engi-
neers. John Wiley & Sons, 2010.
25. G. Strang, Linear algebra and its applications. Brooks Cole, 2006.
26. T. Y. Chiu, T. Leonard, and K.-W. Tsui, “The matrix-logarithmic covariance
model,” Journal of the American Statistical Association, vol. 91, no. 433, pp. 198–
210, 1996.
27. C. Van Loan, Computational frameworks for the fast Fourier transform, vol. 10.
Siam, 1992.
28. Y. LeCun, B. E. Boser, J. S. Denker, D. Henderson, R. E. Howard, W. E. Hub-
bard, and L. D. Jackel, “Handwritten digit recognition with a back-propagation
network,” in Advances in neural information processing systems, pp. 396–404, 1990.
29. A. V. Oppenheim, A. S. Willsky, and S. Nawab, Signals and Systems. Prentice
Hall, 1996.
30. D. Ciresan, U. Meier, and J. Schmidhuber, “Multi-column deep neural networks for
image classification,” in 2012 IEEE Conference on Computer Vision and Pattern
Recognition, pp. 3642–3649, 2012.
31. O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpa-
thy, A. Khosla, M. Bernstein, A. C. Berg, and L. Fei-Fei, “ImageNet large scale
visual recognition challenge,” International Journal of Computer Vision, vol. 115,
no. 3, pp. 211–252, 2015.
12 Ashford, Bird et al.
32. C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Van-
houcke, and A. Rabinovich, “Going deeper with convolutions,” in Proceedings of
the IEEE conference on computer vision and pattern recognition, pp. 1–9, 2015.
33. F. Chollet et al., “Keras.”, 2015.
34. D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv
e-prints, p. arXiv:1412.6980, Dec 2014.
35. J. J. Bird, D. R. Faria, L. J. Manso, A. Ekart, and C. D. Buckingham, “A deep
evolutionary approach to bioinspired classifier optimisation for brain-machine in-
teraction,” Complexity, vol. 2019, 2019.
36. R. Kohavi, “A study of cross-validation and bootstrap for accuracy estimation
and model selection,” in Proceedings of the 14th International Joint Conference on
Artificial Intelligence - Volume 2, IJCAI’95, pp. 1137–1143, 1995.
37. M. L´opez-Ib´nez, J. Dubois-Lacoste, L. P´erez C´aceres, T. St¨utzle, and M. Birat-
tari, “The irace package: Iterated racing for automatic algorithm configuration,”
Operations Research Perspectives, vol. 3, pp. 43–58, 2016.
38. E. K. Burke, M. Gendreau, M. Hyde, G. Kendall, G. Ochoa, E. ¨
Ozcan, and R. Qu,
“Hyper-heuristics: a survey of the state of the art,” Journal of the Operational
Research Society, vol. 64, no. 12, pp. 1695–1724, 2013.
39. A. Mart´ın, R. Lara-Cabrera, F. Fuentes-Hurtado, V. Naranjo, and D. Cama-
cho, “Evodeep: A new evolutionary approach for automatic deep neural networks
parametrisation,” Journal of Parallel and Distributed Computing, vol. 117, pp. 180–
191, 2018.
40. F. Assun¸cao, N. Louren¸co, P. Machado, and B. Ribeiro, “Denser: Deep evolutionary
network structured representation,” arXiv preprint arXiv:1801.01563, 2018.
41. J. J. Bird, A. Ekart, and D. R. Faria, “Evolutionary optimisation of fully connected
artificial neural network topology,” in SAI Computing Conference 2019, SAI, 2019.
42. D. C. Montgomery, Design and Analysis of Experiments. John Wiley & Sons, 8th
ed. ed., 2012.
43. S. Ji, W. Xu, M. Yang, and K. Yu, “3d convolutional neural networks for human ac-
tion recognition,” IEEE transactions on pattern analysis and machine intelligence,
vol. 35, no. 1, pp. 221–231, 2013.
... In this work, an experimental framework is presented in which evolutionary optimisation of neural network hyperparameters is applied in conjunction with a visual data preprocessing technique preliminarily explored in a previous work. During the previous study [5], a 2D CNN was succesfully applied to a 2D image representation of EEG features with a dimensionality reduction algorithm on a 4-channel EEG dataset. In this work, we explore visual data reshaping in 2 and 3 dimensions in order to form pixel image and voxel cube representations of statistical features extracted from electrical brain activity, through which 2D and 3D CNN convolve 'features of features'. ...
... Due to the temporal nature of the EEG waves, single point measures rarely harbor any useful classification accuracy and thus make weak datasets. In this work, statistical features are extracted through a sliding time-window approach [5,36,37] ( jordan-bird/eeg-feature-generation). ...
... In this work the following statistical features were generated for each time window via the process that can be observed in Algorithm 1 as in the previous aforementioned works, where y k = [y k1 , . . . , y kN ], within which K are vectors of paired observations [5,36,37]: ...
Full-text available
Objective The novelty of this study consists of the exploration of multiple new approaches of data pre-processing of brainwave signals, wherein statistical features are extracted and then formatted as visual images based on the order in which dimensionality reduction algorithms select them. This data is then treated as visual input for 2D and 3D CNNs which then further extract 'features of features'. Approach Statistical features derived from three electroencephalography datasets are presented in visual space and processed in 2D and 3D space as pixels and voxels respectively. Three datasets are benchmarked, mental attention states and emotional valences from the four TP9, AF7, AF8 and TP10 10-20 electrodes and an eye state data from 64 electrodes. 729 features are selected through three methods of selection in order to form 27x27 images and 9x9x9 cubes from the same datasets. CNNs engineered for the 2D and 3D preprocessing representations learn to convolve useful graphical features from the data. Main results: A 70/30 split method shows that the strongest methods for classification accuracy of feature selection are One Rule for attention state and Relative Entropy for emotional state both in 2D. In the eye state dataset 3D space is best, selected by Symmetrical Uncertainty. Finally, 10-fold cross validation is used to train best topologies. Final best 10-fold results are 97.03% for attention state (2D CNN), 98.4% for Emotional State (3D CNN), and 97.96% for Eye State (3D CNN). Significance: The findings of the framework presented by this work show that CNNs can successfully convolve useful features from a set of pre-computed statistical temporal features from raw EEG waves. The high performance of K-fold validated algorithms argue that the features learnt by the CNNs hold useful knowledge for classification in addition to the pre-computed features.
... Due to the computational costs of hyperparameter optimisation used in the MLP experiment when a CNN is considered, CNN topology is instead inspired by Ashford et al.'s work [69]. In this work, the topology observed in Table 1 was found to be best after empirical exploration. ...
... It was observed that learning was still occurring at epoch 30 (unlike in the MLPs in Experiment 1), and due to this, learning time was increased to 100 epochs. Table 3 It is important to note that previously, the One Rule Random Forest approach [60] gained 87.16% accuracy and the image representation and CNN approach [69] gained 89.38% accuracy on EEG data. Our network is competitive at 82.39% accuracy on the same dataset with no training whatsoever, using simply the weights from the EMG network. ...
Full-text available
In this work, we show the success of unsupervised transfer learning between Electroencephalographic (brainwave) classification and Electromyographic (muscular wave) domains with both MLP and CNN methods. To achieve this, signals are measured from both the brain and forearm muscles and EMG data is gathered from a 4-class gesture classification experiment via the Myo Armband, and a 3-class mental state EEG dataset is acquired via the Muse EEG Headband. A hyperheuristic multi-objective evolutionary search method is used to find the best network hyperparameters. We then use this optimised topology of deep neural network to classify both EMG and EEG signals, attaining results of 84.76% and 62.37% accuracy, respectively. Next, when pre-trained weights from the EMG classification model are used for initial distribution rather than random weight initialisation for EEG classification, 93.82%(+29.95) accuracy is reached. When EEG pre-trained weights are used for initial weight distribution for EMG, 85.12% (+0.36) accuracy is achieved. When the EMG network attempts to classify EEG, it outperforms the EEG network even without any training (+30.25% to 82.39% at epoch 0), and similarly the EEG network attempting to classify EMG data outperforms the EMG network (+2.38% at epoch 0). All transfer networks achieve higher pre-training abilities, curves, and asymptotes, indicating that knowledge transfer is possible between the two signal domains. In a second experiment with CNN transfer learning, the same datasets are projected as 2D images and the same learning process is carried out. In the CNN experiment, EMG to EEG transfer learning is found to be successful but not vice-versa, although EEG to EMG transfer learning did exhibit a higher starting classification accuracy. The significance of this work is due to the successful transfer of ability between models trained on two different biological signal domains, reducing the need for building more computationally complex models in future research.
Full-text available
This study suggests a new approach to EEG data classification by exploring the idea of using evolutionary computation to both select useful discriminative EEG features and optimise the topology of Artificial Neural Networks. An evolutionary algorithm is applied to select the most informative features from an initial set of 2550 EEG statistical features. Optimisation of a Multilayer Perceptron (MLP) is performed with an evolutionary approach before classification to estimate the best hyperparameters of the network. Deep learning and tuning with Long Short-Term Memory (LSTM) are also explored, and Adaptive Boosting of the two types of models is tested for each problem. Three experiments are provided for comparison using different classifiers: one for attention state classification, one for emotional sentiment classification, and a third experiment in which the goal is to guess the number a subject is thinking of. The obtained results show that an Adaptive Boosted LSTM can achieve an accuracy of 84.44%, 97.06%, and 9.94% on the attentional, emotional, and number datasets, respectively. An evolutionary-optimised MLP achieves results close to the Adaptive Boosted LSTM for the two first experiments and significantly higher for the number-guessing experiment with an Adaptive Boosted DEvo MLP reaching 31.35%, while being significantly quicker to train and classify. In particular, the accuracy of the nonboosted DEvo MLP was of 79.81%, 96.11%, and 27.07% in the same benchmarks. Two datasets for the experiments were gathered using a Muse EEG headband with four electrodes corresponding to TP9, AF7, AF8, and TP10 locations of the international EEG placement standard. The EEG MindBigData digits dataset was gathered from the TP9, FP1, FP2, and TP10 locations.
Conference Paper
Full-text available
This paper focuses on ultra-low power embedded classification of neural activities. The machine learning (ML) algorithm has been trained using evoked local field potentials (LFPs) recorded with an implanted 16×16 multi-electrode array (MEA) from the rat barrel cortex while stimulating the whisker. Experimental results demonstrate that ML can be successfully applied to noisy single-trial LFPs. We achieved up to 95.8% test accuracy in predicting the whisker deflection. The trained ML model is successfully implemented on a low-power embedded system with an average consumption of 2.6 mW.
Conference Paper
Full-text available
This paper explores single and ensemble methods to classify emotional experiences based on EEG brainwave data. A commercial MUSE EEG headband is used with a resolution of four (TP9, AF7, AF8, TP10) electrodes. Positive and negative emotional states are invoked using film clips with an obvious valence, and neutral resting data is also recorded with no stimuli involved, all for one minute per session. Statistical extraction of the alpha, beta, theta, delta and gamma brainwaves is performed to generate a large dataset that is then reduced to smaller datasets by feature selection using scores from OneR, Bayes Network, Information Gain, and Symmetrical Uncertainty. Of the set of 2548 features, a subset of 63 selected by their Information Gain values were found to be best when used with ensemble classifiers such as Random Forest. They attained an overall accuracy of around 97.89%, outperforming the current state of the art by 2.99 percentage points. The best single classifier was a deep neural network with an accuracy of 94.89%.
Conference Paper
Full-text available
This paper proposes an approach to selecting the amount of layers and neurons contained within Multilayer Perceptron hidden layers through a single-objective evolutionary approach with the goal of model accuracy. At each generation, a population of Neural Network architectures are created and ranked by their accuracy. The generated solutions are combined in a breeding process to create a larger population, and at each generation the weakest solutions are removed to retain the population size inspired by a Darwinian 'survival of the fittest'. Multiple datasets are tested, and results show that architectures can be successfully improved and derived through a hyper-heuristic evolutionary approach, in less than 10% of the exhaustive search time. The evolutionary approach was further optimised through population density increase as well as gradual solution max complexity increase throughout the simulation.
Conference Paper
Full-text available
This work aims to find discriminative EEG-based features and appropriate classification methods that can categorise brainwave patterns based on their level of activity or frequency for mental state recognition useful for human-machine interaction. By using the Muse headband with four EEG sensors (TP9, AF7, AF8, TP10), we categorised three possible states such as relaxing, neutral and concentrating based on a few states of mind defined by cognitive behavioural studies. We have created a dataset with five individuals and sessions lasting one minute for each class of mental state in order to train and test different methods. Given the proposed set of features extracted from the EEG headband five signals (alpha, beta, theta, delta, gamma), we have tested a combination of different features selection algorithms and classifier models to compare their performance in terms of recognition accuracy and number of features needed. Different tests such as 10-fold cross validation were performed. Results show that only 44 features from a set of over 2100 features are necessary when used with classical classifiers such as Bayesian Networks, Support Vector Machines and Random Forests, attaining an overall accuracy over 87%.
Full-text available
Deep Evolutionary Network Structured Representation (DENSER) is a novel approach to automatically design Artificial Neural Networks (ANNs) using Evolutionary Computation (EC). The algorithm not only searches for the best network topology (e.g., number of layers, type of layers), but also tunes hyper-parameters, such as, learning parameters or data augmentation parameters. The automatic design is achieved using a representation with two distinct levels, where the outer level encodes the general structure of the network, i.e., the sequence of layers, and the inner level encodes the parameters associated with each layer. The allowed layers and hyper-parameter value ranges are defined by means of a human-readable Context-Free Grammar. DENSER was used to evolve ANNs for two widely used image classification benchmarks obtaining an average accuracy result of up to 94.27% on the CIFAR-10 dataset, and of 78.75% on the CIFAR-100. To the best of our knowledge, our CIFAR-100 results are the highest performing models generated by methods that aim at the automatic design of Convolutional Neural Networks (CNNs), and is amongst the best for manually designed and fine-tuned CNNs .
Deep learning has been successfully used in numerous applications because of its outstanding performance and the ability to avoid manual feature engineering. One such application is electroencephalogram (EEG) based brain-computer interface (BCI), where multiple convolutional neural network (CNN) models have been proposed for EEG classification. However, it has been found that deep learning models can be easily fooled with adversarial examples, which are normal examples with small deliberate perturbations. This paper proposes an unsupervised fast gradient sign method (UFGSM) to attack three popular CNN classifiers in BCIs, and demonstrates its effectiveness. We also verify the transferability of adversarial examples in BCIs, which means we can perform attacks even without knowing the architecture and parameters of the target models, or the datasets they were trained on. To our knowledge, this is the first study on the vulnerability of CNN classifiers in EEG-based BCIs, and hopefully will trigger more attention on the security of BCI systems.
Automatic human affect recognition is a key step towards more natural human-computer interaction. Recent trends include recognition in the wild using a fusion of audiovisual and physiological sensors, a challenging setting for conventional machine learning algorithms. Since 2010, novel deep learning algorithms have been applied increasingly in this field. In this paper, we review the literature on human affect recognition between 2010 and 2017, with a special focus on approaches using deep neural networks. By classifying a total of 950 studies according to their usage of shallow or deep architectures, we are able to show a trend towards deep learning. Reviewing a subset of 233 studies that employ deep neural networks, we comprehensively quantify their applications in this field. We find that deep learning is used for learning of (i) spatial feature representations, (ii) temporal feature representations, and (iii) joint feature representations for multimodal sensor data. Exemplary state-of-the-art architectures illustrate the progress. Our findings show the role deep architectures will play in human affect recognition, and can serve as a reference point for researchers working on related applications.