=Paper=
{{Paper
|id=Vol-2563/aics_28
|storemode=property
|title=Identifying Extra-terrestrial Intelligence Using Machine Learning
|pdfUrl=https://ceur-ws.org/Vol-2563/aics_28.pdf
|volume=Vol-2563
|authors=Małgorzata Gutowska,Michael Scriney,Andrew McCarren
|dblpUrl=https://dblp.org/rec/conf/aics/GutowskaSM19
}}
==Identifying Extra-terrestrial Intelligence Using Machine Learning==
https://ceur-ws.org/Vol-2563/aics_28.pdf
Identifying extra-terrestrial intelligence using
machine learning
Malgorzata Gutowska1 , Michael Scriney2 , and Andrew McCarren3
1
Dublin City University, Dublin 9, Ireland
malgorzata.gutowska2@mail.dcu.ie
2
Insight Centre for Data Analytics, Dublin City University, Dublin 9, Ireland
michael.scriney@insight-centre.org
3
VistaMilk Research Centre, School of Computing, Dublin City University, Ireland
andrew.mccarren@dcu.ie
Abstract. Since the date of establishment of the SETI Institute, its sci-
entists have used various approaches in their search for extra-terrestrial
intelligence (SETI). A novel idea involved image categorisation tech-
niques in classifying radio signals represented by 2D spectrograms. The
dataset of simulated radio signals, created for classification purposes have
been used in this work to train models based on neural network archi-
tectures. It is shown in this paper that combining three different models,
trained on features obtained by various techniques, has a positive im-
pact on model accuracy and performance. Features learned by a convo-
lutional neural network (CNN), bottleneck features from existing models
and manually extracted features from the spectrograms comprised the
three feature sets used as training data for the combined model. It was
also shown that combining different methods of spectrogram generation
resulted in improving the accuracy of the final model.
Keywords: Convolutional Neural Network · Image Processing · Spec-
trograms.
1 Introduction
Scientists from the SETI Institute (Search for Extra-terrestrial Intelligence) mon-
itor radio signals coming from multiple directions from space, searching for signs
of extra-terrestrial intelligence [20]. The artificial signals are thought likely to be
narrow-band, in contrast to natural radiation featuring wide spectra. Therefore,
the efforts to identify artificially generated electromagnetic waves are focused on
searching for narrow-band signals. The Allen Telescope Array (ATA) is an array
of 42 radio telescopes in northern California [24], one of whose main activities is
the search for ETI. This telescope allows for simultaneous observations of 3 very
small windows of the sky, which helps in highlighting potential signs of intelligent
origin [5]. Narrow band signals of frequency drifting linearly over time highlight
candidates of potential artificial origin [5]. The sensitive algorithm within the
specialised software at ATA detects various types of narrow-band signals. While
Copyright © 2019 for this paper by its authors. Use permitted under Creative Commons License
Attribution 4.0 International (CC BY 4.0).
2 M. Gutowska, et al.
the rate of not detecting potentially interesting signals is acceptably low, many
other similar signal types produce false positives, resulting in additional val-
ueless sets of observations [5]. Transformed into a form of 2D spectrograms,
narrow-band radio signals with a drifting frequency can be easily distinguished
from other signal types by a human due to their shape. This characteristic can
be utilised in an automated fashion with an adoption of rapidly developing im-
age recognition techniques such as convolutional neural networks (CNN). SETI
scientists simulated 140,000 signals and used these as training data for image
classification algorithms. A competition, which was held in July 2017, produced
a range of classification models, of which the best achieved results of 95% accu-
racy [6]. A subset of the simulated and labelled data (35,000 signals) was made
available after the competition was over [7]. The competition winners used high
powered computing provided by IBM [5]. The purpose of this research is to find
out whether a satisfactory classification of SETI signals can be achieved using
convolutional neural networks and other machine learning techniques on limited
resources accessible to citizen scientists. The aim of this work was to reproduce
or improve the results of top scoring models using Colaboratory platform with
GPU processing and 18 GB RAM.
2 Background
The ATA radio telescopes record analogue voltage signals representing electro-
magnetic field oscillations reaching individual antennas. Before the signals are
stored, they are demodulated down from GHz range to lower frequencies and
digitised, which results in a stream of complex-valued time-series data. The digi-
tisation process produces 104 million complex samples per second, corresponding
to about 70MHz bandwidth window around central frequency from the range 1-
10 GHz, which is the observation range for ATA [5].
2.1 Spectrograms
Usage of frequency domain for signals oscillating in time, in many use cases gives
a more meaningful illustration of signal’s characteristics than a time domain
representation. A transition from time to frequency domain can be achieved by
Fourier Transform (FT), based on the assumption that each periodic signal can
be decomposed to a sum of sines and cosines:
∞ ∞
1 X X
S(t) = a0 + an cos nωt + bn sin nωt (1)
2 n=1 n=1
or, if represented in a complex space, it can be expressed as
∞
X
S(t) = cn ejnωt (2)
−∞
Identifying extra-terrestrial intelligence using machine learning 3
where S(t) is the approximation of signal s(t) in the time domain and ω is signal
frequency. The coefficients an and bn or cn describe a signal in the frequency
domain and they are expressed as:
Z Z
2 2
an = s(t) cos nωt dt, bn = s(t) sin nωt dt (3)
T T
and Z
1
cn = s(t)e−jnωt dt (4)
T
where T is an oscillation period. The link between sine, cosine and expo-
nential functions is provided by the Euler formulas [15]. The result from the
FT is the relationship between a signal’s power and its frequency. By including
additional time dependency and representing the power by pixels brightness, it
is possible to produce 2D spectrograms, which are the types of images classified
in this work.
2.2 Observed signals
Radio signals carrying a large amount of energy at a narrow frequency range are
often called “narrowband”, as they feature narrow bandwidth. In the observed
signals, the carrier (central) frequency usually changes with time. This frequency
drift (Doppler’s drift) is caused by the Earth’s rotation and/or the movement
or acceleration of the source. These signals are of the greatest interest to SETI
scientists. Narrow-band signals with a constant frequency (without drift) are of
less interest, because they are more likely to come either from terrestrial broad-
casting stations or geostationary satellites [19]. SonATA, the software managing
the signals recording and pre-processing at ATA, is also responsible for deter-
mining the candidates for intelligent signals. A signal is flagged as a candidate, if
it demonstrates characteristics of a narrow-band signal; in such cases follow-up
observations are conducted in order to assess the possibility that the signal is
of extra-terrestrial origin. The relatively high rate of false positives, which is a
weak point of the currently used detection algorithm of SonATA, can easily lead
to wasting large amounts of observation time [5].
2.3 Signals simulation
A need for a multi-categorical classifier brought an idea of training a neural-
network-based model on spectrogram images. Simulations of radio signals pro-
vide a controlled dataset of labelled images. SETI scientists, based on their
domain expertise, have simulated 6 categories, which reflect the most commonly
observed signal groups, plus an additional category of background noise. Two ex-
amples of spectrograms of the simulated signals and their labels are presented in
Figure 1. Pictures of all categories are shown in [9]. The full list contained seven
categories: narrowband, narrowbanddrd, squiggle, squarepulsednarrowband,
squigglesquarepulsednarrowband, brightpixel, noise. Simulation process
4 M. Gutowska, et al.
(a) narrowband (b) squiggle
Fig. 1: Spectrograms representing examples of simulated signals of two categories.
The scale used on simulated images is arbitrary and does not reflect any physical
units.
and parameters of individual categories have been presented in detail by [5].
One of the goals of the simulation was to generate signals of a wide range of
difficulties in distinguishing between individual categories.
2.4 CNN architectures
In this paper we propose a technique that can be used on low performing home
computers and give compariable results to the winning competitors. Across
recent years, innovative CNN architectures such as VGG, ResNet, Inception,
DenseNet or NASNet [18] were each setting a new state-of-the-art classification
on datasets such as ImageNet [8] or CIFAR [14]. ResNet’s innovation was an
introduction of identity connections and optimisation of the residual mapping
instead of the original mapping [11]. By developing the residual learning it was
possible to train extremely deep networks without gradient vanishing and ac-
curacy degradation. Another innovation – DenseNet – introduced connections
between all convolutional layers [12].
3 Methods
The following section describes the methods utilised in this work and the data
pre-processing steps, which are as follows:
– obtaining spectrograms from time-series data
– spectrograms enhancement
– exploration of existing Convolutional Neural Networks (CNN)
– creation of manual features
– building the combined architecture
3.1 From time-series data to spectrograms
The simulated data consisted of a series of complex numbers in the time domain.
Complex numbers are often chosen for mathematical convenience, when it is
required to perform operations on amplitude and frequency.
Identifying extra-terrestrial intelligence using machine learning 5
For example, a signal from a real space written as x(t) = A1 cos ωt in a
complex space can be expressed as x(t) = A2 ejωt and then by using the Euler
formula (§2), it can be written as a sum of Real and Imaginary components:
x(t) = A2 cos ωt + A2 j sin ωt = a(t) + jb(t) The raw data used in this work
consisted of numbers of the above pattern.
The following steps were performed to obtain spectrograms from the simu-
lated data:
1. Time-series data has been split into time intervals. The length of the interval
would later reflect a time unit on the spectrogram’s vertical axis.
2. The time segments have been stacked horizontally to create a 2D array.
3. The array has been processed with Hann window [10]. This technique is used
to prevent against spectral leakage and generation of artificial frequencies.
This step was performed only in some cases, as described later.
4. Fourier Transform (FT) has been performed on each time segment and a
square of its modulus has been taken.
5. A logarithm has been applied on the FT results. This operation brings higher
granularity into the lower ranges of a colour histogram. This step was not
done in all of the cases, which will be discussed later.
6. The resulting values have been normalised to the range 0-255, for the array
to be plotted as an image, where brighter colours correspond to higher signal
power.
3.2 Enhancing the spectrograms
The maximum image size attainable from the simulated data was 196,608 pixels,
reflecting a resolution of 384x512. However, for the majority of use cases lower
resolution images (128x256) have been created.
Monochromatic pictures have only one colour channel, whereas colour pic-
tures can have 2-4 channels; e.g. the RGB colour space uses 3 channels where
each colour – red, green and blue are described separately on a scale from 0 to
255. Convolutional networks are therefore well suited to operate on colour images
by convolving 3-dimensional filters simultaneously through all colour layers and
forming independent weight tensors. Even though spectrograms are monochro-
matic, in this research separate channels have been used to carry the results of
different techniques of image generation. The following methods were explored
for this purpose:
– a logarithm has been applied over the FT (Figure 2a),
– the logarithm has been omitted (Figure 2b),
– Hann window has been applied before the FT and the logarithm was taken
from the FT result (Figure 2c),
– Radon transform [16] has been applied over the spectrogram obtained as in
the first point.
6 M. Gutowska, et al.
(a) Spectrogram obtained (b) Spectrogram obtained (c) Spectrogram obtained
using logarithm over the without applying the log- using the Hann window
Fourier Transform out- arithm; no Hann function before Fourier Transform
put; no Hann window used. and the logarithm over
used. the FT output.
Fig. 2: Spectrogram enhancing. The spectrograms obtained by presented meth-
ods have been combined into multichannel images.
While the Radon transform can be useful in detecting straight lines within
a noisy background [17], it was eventually omitted in the CNN training, due
to lack of sufficient distinction. The spectrograms presented in Figure 2 have
been merged into 3-channel pictures and used in CNN’s. Another image pre-
processing technique which has been explored was a noise reduction technique.
Two algorithms were utilised for this purpose; contrast enhancing custom func-
tion and the median-based denoising method from OpenCV library [2]. While
this process seemed to work well with many spectrograms, it failed in cases
where signal to noise ratio was very low. Overall it did not appear to improve
the native CNN’s pattern detection methods and thus was omitted in further
CNN trials. However, it was adopted in the manual feature extraction, which is
described later.
3.3 Exploring CNN architectures
Prototyping with CNN’s has been performed on 4 categories, with total size of
the data set being 4000. Further method exploration was performed on a set
of 7000 signals, including all 7 categories, whereas the full data set size used
here was 35,000 signals. The CNN architectures, which have been tested in this
research include:
– 3 and 4-convolutional-block networks of Convolutional + ReLU4 + Max
Pooling5 layers with variable filter sizes: 3 – 7,
– VGG16 and VGG19 [21]
– Xception [4]
– ResNeXt101 [25]
– InceptionResNetV2 [22]
– DenseNet201 (201 layers) [12]
4
ReLU is a non-linear activation function
5
A max value of 2x2 pixels area
Identifying extra-terrestrial intelligence using machine learning 7
Most of the above networks were used through transfer learning, rather than
training from random weights. The CNNs of 3 and 4 convolutional blocks, as
well as VGG16 have been fully trained from randomly distributed weights, yet
the accuracy from the VGG did not outperform the former one.
In addition to the above architectures a few variants of combined models
were tested, such as DenseNet or ResNeXt accompanied with a neural network
of fully connected layers trained on manually crafted features.
3.4 Manually crafted features
Apart from variants of CNN’s, a more traditional machine learning method was
also part of this research. A regular neural network was trained on a set of
features, which were manually crafted to reflect the main characteristics of all
classes. The following features were generated:
– the number of pixels in each of the 4 brightness ranges, such as: 0 – 62.5,
62.5 – 125, etc., marked as Q1 , – Q4 in Figure 3a,
– characteristics of pixel brightness distribution, such as mean, variance and
skewness,
– estimated width of the bounding box around the signal, as shown in Fig-
ure 3b; this was obtained using a custom function,
– estimated total height of individual bounding boxes, as in Figure 3b; this
was generated using a custom method,
– estimated length of the signal’s line using Canny edge detection algorithm
[3].
The width and height of a signal line were calculated by counting total number
of columns or rows containing pixels of a brightness over a predefined threshold.
(a) Colour (brightness) histogram of (b) Bounding boxes around signal la-
a sample spectrogram with annotated belled with width and height.
four quarters.
Fig. 3: Illustration of extraction of some of the features.
8 M. Gutowska, et al.
3.5 Combined architecture
Residual Networks (ResNet) [11] and Densely Connected Convolutional Net-
works (DenseNet) [12] were shown to perform exceptionally well with the spec-
trogram images from SETI [6]. In this work, an architecture of the DenseNet
(or ResNeXt) model combined with other models was created. The motivation
behind combining two or more models was a belief that different types of net-
works would be able to learn different characteristics of original images, therefore
combining them would lead to the “best of each” net result. Both versions of
combined architectures – with DenseNet and ResNeXt performed similarly, how-
ever, to achieve the same result, DenseNet required much less memory due to a
compact architecture featuring less parameters.
Densely Connected Convolutional Network. DenseNet had the highest
performance out of all the networks tested [12]. The innovative feature of this
architecture is passing the output from all preceding convolution layers to all
subsequent layers within a dense block. This way, low level features combined
with more complex features are simultaneously passed to the classification layer.
This is not without significance for this particular use case, where images contain
mostly simple shapes.
Final model implementation. The model developed in this work is shown
in Figure 4. Pretrained on the ImageNet dataset, the DenseNet201 model has
been used for extracting features out of the training dataset (bottleneck fea-
tures), by trimming the top part of its architecture and using only the convolu-
tional part. A separate 4-convolutional-block CNN was trained on the spectro-
gram data from random weights, which produced similarly structured bottleneck
features. Two fully connected layers were then added on top of each model and
the output was passed onto the common layer. A third element of the whole ar-
chitecture was a regular network with one hidden and one output layer trained
on manually extracted features. The output produced was passed again to the
common layer, which was comprised of the outputs of all three individual models.
The common layer combined the outputs together, giving each of them the same
number of features. Two fully connected layers were added on top of the common
layer, with the softmax activation, producing the final output of probabilities
for seven categories. In order to achieve the optimal batch size for training, the
split of 35,000 pictures into train, validation and test datasets was performed
with the ratio of 76:12:12 respectively. The data split have been performed once
and then the same subsets were fed into three models. The test data was used
at the evaluation stage only, after the training process had completed. During
training, images were loaded in batches in order to accommodate available mem-
ory resources. In order to overcome overfitting, dropout was applied onto three
branches of the combined model.
The gradient descent process was accomplished with the Adam optimiser [13],
with the initial learning rate 1e−3 decaying with subsequent epochs. Training
ran through 9 epochs, with each of them taking approximately 43s.
Identifying extra-terrestrial intelligence using machine learning 9
Fig. 4: Combined model architecture. CNN trained on the spectrograms, pre-
trained DenseNet [12] and regular network trained on extracted features have
been combined together to produce the final output.
4 Results and Discussion
Table 1 compares accuracy values achieved by various methods of this research
and Table 2 shows the winning SETI competition results with the best per-
forming method from this work. The performance was measured for both types
of images: monochromatic (including only one type of spectrogram) and multi-
channel (containing three versions of spectrograms). The triple model achieved
the following scores: accuracy – 87.9%, logarithmic loss – 0.3815. Detailed per-
formance scores are shown in Table 3.
From among the single models, the best accuracy has been achieved by the
CNN with 4 convolutional blocks, though the combined model outperformed all
single model architectures. While other combined architectures were explored,
the one combining DenseNet with manual features (MF) and the CNN-trained
features yielded the best results. The Table 3 shows the accuracy for all cate-
gories. The category with the highest F1 score was squiggle, which by visual
inspection is also the most obvious to recognise. The next best recognised signal
is brightpixel. While the precision of brightpixel detection is quite high, the
recall is not as good, as many actual brightpixel signals were not distinguish-
10 M. Gutowska, et al.
Table 1: Performance of models developed in this work (individual and merged).
Model Images Dataset Accuracy
DenseNet mono 35K Signals 73.96%
DenseNet-MF mono 35K Signals 82.49%
DenseNet 3 chnl 35K Signals 76.80%
DenseNet-MF 3 chnl 35K Signals 84.48%
4-bl CNN 3 chnl 35K Signals 78.52%
4-bl CNN-DenseNet-MF 3 chnl 35K Signals 87.86%
Table 2: Comparison between the best performing method in this study to com-
petition winners.
Team Rank Dataset Accuracy LogLoss
Effsubsee [23] 1 140K signals 94.6% 0.1881
Signet [1] 2 140K signals 94.7% 0.2263
Snb1 3 140K signals 87.5% 0.3847
4-bl CNN-DenseNet-MF 3 35K Signals 87.9% 0.3815
Table 3: Final model performance scores
Category Precision Recall F1
brightpixel 0.94 0.88 0.91
narrowband 0.84 0.89 0.87
narrowbanddrd 0.87 0.80 0.83
noise 0.82 0.98 0.89
squarepulsednarrowband 0.86 0.82 0.84
squiggle 0.92 0.95 0.93
squigglesquarepulsednarrowband 0.92 0.83 0.87
able from noise signals. The noise classification has relatively low precision and
in addition to brightpixel, many other signals were misclassified as noise. The
lowest confusion with noise appears for squiggle and narrowband. As the cor-
rect classification of the narrowband category is of the greatest interest to SETI
researchers [5], this can be perceived as an advantage of this model. The poor-
est classification occurs in narrowbanddrd signals, which are often confused with
narrowband. Indeed, in many cases from the narrowbanddrd class, the frequency
drift was so small that it would be impossible to see a difference when visually
inspecting both types of signal. A fact worth noting is that the model presented
in this work achieved better overall results in noise detection than both competi-
tion winning models, which reported the F1 score values as: 88.54% and 88.04%
[6], significantly lower than the average accuracy of both models: 94.61% and
94.74% respectively. The model presented in this paper achieved 89%, which is
above the average accuracy of the model itself. The reason behind this is the
inclusion of manually extracted features, that added an ability to separate the
noise category relatively well from other categories. Overall, the model presented
Identifying extra-terrestrial intelligence using machine learning 11
in this work did not outperform already existing models, as it would have placed
slightly higher than 3rd place. However, it should be noted that the winning
team utilised the computational and storage resources provided by IBM for this
purpose [5]. The dataset size used in this research was also significantly smaller,
as the full dataset generated for this purpose had 140,000 images, while the one
used in this work had 35,000. By using the combined model it was possible to
achieve the reported accuracy during training taking 6 minutes and 27s. The
training ran through 9 epochs and stopped afterwards, when the validation loss
stopped decreasing. This is not the total time however, as training of the custom
CNN took 40 minutes, similar to feature extraction from DenseNet. The longest
time taken was to calculate the manual features for all images, especially, the
width and height of the bounding box, which altogether took around 20 hours.
The main reason for this process taking so much time was usage of non-optimised
custom functions. In summary, given the resources used, rather short training
time, and the fact that images of reduced resolution were used in both convo-
lutional networks, a relatively good accuracy was achieved thanks to additional
innovations introduced in this work.
5 Conclusions and Future Work
In this work a means of improving classification accuracy (in relation to other
examined methods) by both enhancing the source data and consolidating NN
architectures was presented. Three methods were integrated by creating spec-
trograms from raw time-series data into 3 image channels and the inclusion of
manually extracted features from the spectrograms. In addition, the consolida-
tion of three neural network architectures into a single model was implemented.
By using these methods an accuracy of 87.86% with F1 scores ranging from 0.83
to 0.93 was obtained. This research shows that including manually extracted fea-
tures helps in classifying specific categories, such as the SETI category noise.
Finally, this work showed favourable results in comparison to competitor meth-
ods implemented on more advanced architecture environments. Further work in
this area could be a promising direction for additional model enhancement. Fu-
ture work seeks to explore the addition of a wider range of manual features, as
well as other denoising techniques.
References
1. Bastian, B.: Ml4seti (2017), https://github.com/sagelywizard/ml4seti?source=
post page
2. Bradski, G.: The OpenCV Library. Dr. Dobb’s Journal of Software Tools (2000)
3. Canny, J.: A computational approach to edge detection. In: Fischler, M.A.,
Firschein, O. (eds.) Readings in Computer Vision, pp. 184 – 203. Morgan Kauf-
mann, San Francisco (CA) (1987). https://doi.org/10.1016/B978-0-08-051581-
6.50024-6
4. Chollet, F.: Xception: Deep learning with depthwise separable convolutions. CoRR
abs/1610.02357 (2016), http://arxiv.org/abs/1610.02357
12 M. Gutowska, et al.
5. Cox, G.A., Egly, S., Harp, G.R., Richards, J., Vinodababu, S., Voien, J.: Clas-
sification of simulated radio signals using wide residual networks for use in the
search for extra-terrestrial intelligence. CoRR abs/1803.08624 (2018), http:
//arxiv.org/abs/1803.08624
6. Cox, G.A.: Using artificial intelligence to search for extraterrestrial intelli-
gence (2017), https://medium.com/codait/using-artificial-intelligence-to-search-
for-extraterrestrial-intelligence-ec19169e01af
7. Cox, G.A.: Welcome to the seti institute code challenge! (2018), https://
github.com/setiQuest/ML4SETI/blob/master/tutorials/Step 1 Get Data.ipynb
8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A Large-
Scale Hierarchical Image Database. In: CVPR09 (2009)
9. Gutowska, M., McCarren, A.: The search for extra-terrestrial intelligence: classifi-
cation of radio signals with machine learning. Insight Technical Report 1 (Septem-
ber 2019), https://doras.dcu.ie/23780/1/2019-mcm-gutowsm2-Final Practicum
Paper.pdf
10. Harris, F.J.: On the use of windows for harmonic analysis with the discrete fourier
transform. In: Proc. IEEE. pp. 51–83 (1978)
11. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition.
CoRR abs/1512.03385 (2015), http://arxiv.org/abs/1512.03385
12. Huang, G., Liu, Z., Weinberger, K.Q.: Densely connected convolutional networks.
CoRR abs/1608.06993 (2016), http://arxiv.org/abs/1608.06993
13. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint
arXiv:1412.6980 (2014)
14. Krizhevsky, A.: Learning multiple layers of features from tiny images. University
of Toronto (05 2012)
15. Moskowitz, M.A.: A course in complex analysis in one variable. World Scientific
(2002)
16. Murphy, L.M.: Linear feature detection and enhancement in noisy images via
the radon transform. Pattern Recognition Letters 4(4), 279 – 284 (1986).
https://doi.org/10.1016/0167-8655(86)90009-7
17. Nikitin, I.: Statistical analysis of narrow-band signals at setilive.org. arXiv e-prints
arXiv:1502.04887 (Feb 2015)
18. Pan, S.J., Yang, Q.: A survey on transfer learning. IEEE Transac-
tions on Knowledge and Data Engineering 22(10), 1345–1359 (Oct 2010).
https://doi.org/10.1109/TKDE.2009.191
19. Richards, J.: Seti institute (2016), https://old.seti.org/users/jrichards
20. SETI: Seti institute (2019), https://seti.org
21. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale
image recognition. CoRR abs/1409.1556 (2014)
22. Szegedy, C., Ioffe, S., Vanhoucke, V.: Inception-v4, inception-resnet and the im-
pact of residual connections on learning. CoRR abs/1602.07261 (2016), http:
//arxiv.org/abs/1602.07261
23. Vinodababu, S.: Wide-residual-nets-for-seti (2017), https://github.com/sgrvinod/
Wide-Residual-Nets-for-SETI?source=post page
24. Welch, J., et al.: The Allen Telescope Array: The First Widefield, Panchromatic,
Snapshot Radio Camera for Radio Astronomy and SETI. IEEE Proceedings 97,
1438–1447 (Aug 2009). https://doi.org/10.1109/JPROC.2009.2017103
25. Xie, S., Girshick, R.B., Dollár, P., Tu, Z., He, K.: Aggregated residual transforma-
tions for deep neural networks. CoRR abs/1611.05431 (2016), http://arxiv.org/
abs/1611.05431