Astronomy & Astrophysics manuscript no.

33800corr c ESO 2018

December 6, 2018

Exploring galaxy evolution with generative models

Kevin Schawinski1 , M. Dennis Turp1, , and Ce Zhang2

1
Institute for Particle Physics and Astrophysics, Department of Physics, ETH Zurich, Wolfgang-Pauli-Strasse 27, CH-8093, Zürich,
Switzerland
e-mail: [Link]@[Link], dturp@[Link]
2
Systems Group, Department of Computer Science, ETH Zurich, Universitätstrasse 6, CH-8006, Zürich, Switzerland
e-mail: [Link]@[Link]
Received July 9, 2018; accepted August 6, 2018
arXiv:1812.01114v2 [[Link]] 5 Dec 2018

ABSTRACT

Context. Generative models open up the possibility to interrogate scientific data in a more data-driven way.
Aims. We propose a method that uses generative models to explore hypotheses in astrophysics and other areas. We use a neural
network to show how we can independently manipulate physical attributes by encoding objects in latent space.
Methods. By learning a latent space representation of the data, we can use this network to forward model and explore hypotheses in
a data-driven way. We train a neural network to generate artificial data to test hypotheses for the underlying physical processes.
Results. We demonstrate this process using a well-studied process in astrophysics, the quenching of star formation in galaxies as they
move from low- to high-density environments. This approach can help explore astrophysical and other phenomena in a way that is
different from current methods based on simulations and observations.
Key words. Methods: data analysis – Methods: statistical – Galaxies: evolution

1. Introduction
Many objects of interest in astrophysics appear effectively static
Discriminator
ỹ=0
as the relevant characteristic timescales are far beyond human Dis(E(x))

lifetimes. For this reason, there are generally two approaches Encoder
E(x)
E(X)
researchers take to understand the formation and evolution of
Decoder
objects such as galaxies and quasars: they either take observa- D(E(x),y)

tions and fit models to the data, or they propose some underlying log(ssfr)=-9.53 Binarize y

physical model and implement it in a simulation. Observations

are limited by the underlying processes we can infer from them, Original Data Latent space Output
and simulations make assumptions for the processes modeled
and are often computationally very expensive.
On the other hand, an emerging technique (Lample et al.
2017; Antipov et al. 2017; Mathieu et al. 2016; Bowman et al.
2016; Kingma & Welling 2013; Rezende et al. 2014) has re-
cently been developed by the machine-learning community that
can perform tasks such as taking a photo of a human being as
input, and manipulating a certain attribute (e.g. age) and gen-
erating a new photo of the same person (See Figure 1). In this
paper, we ask the following question: Can this method be used
as a way to explore scientific hypotheses in a purely data-driven Fig. 1. Fader network architecture: Original galaxy images are input to
way? Can we achieve a similar physical understanding by using an encoder E(x) which performs a mapping to a latent space of fixed
such a method? dimension. The associated physical property is binarized into a label y.
The parametres E(x) and y are input to a decoder D(E(x), y) which tries
What we propose here is that we can take the data and inter- to reconstruct the original input image. The discriminator Dis(E(x))
rogate the network trained on the data to explore plausible hy- tries to predict the label y from the latent code E(x). Below, we show
potheses. We start with a neural network trained on a set of ob- two examples of changing a single attribute in latent space using a fader
jects associated with a given measured physical property. Once network: the aging of a human face learned from age labels (using a
trained, we can then use the network in two ways: we can encode pretrained model; Lample et al. 2017) , and the lowering of the sSFR of
a real object to a latent space, that is, obtain its representation in a galaxy using sSFR labels.
latent space by its latent space vector. We can also go the other
way and construct a latent space vector and have the network de-
code the corresponding object into real space. This latent space formation well, it needs to learn the most salient features of a
contains a model of all the salient features of the objects the net- group of objects. By changing the latent space vector entries, we
work is trained on, and so for the network to perform this trans- can walk in latent space and so transform objects from one state
Article number, page 1 of 4
 A&A proofs: manuscript no. 33800corr

Org Rec Field Satellite

Fig. 2. Hypothesis of taking field galaxies and turning them into satellites. The Org column shows the original galaxy image, and the Rec column
shows the best reconstruction of the original image. The remainder of each row shows the transformation from field galaxy to satellite as predicted
by our environment Fader network. This figure represents our hypothesis-generation step as we see that when galaxies become satellites, they
become redder, and their structure changes to become more centrally concentrated.

to another (Figure 1)(Lample et al. 2017; Antipov et al. 2017; approach is limited by the available labels; a process could de-
Mathieu et al. 2016; Bowman et al. 2016; Kingma & Welling pend on parameters traced by multiple labels, or by parameters
2013; Rezende et al. 2014). not captured by any of the available labels.
Using this structure, we can isolate the action of a parameter
and observe its effect on the data by varying it independently of
other properties. Suppose we start with a population of objects 2. Method
A which we suspect evolves into population B. Now we can ask We use the Fader network (Lample et al. 2017) architecture to
two questions: demonstrate how we can use this approach to test hypotheses
in real astrophysical settings. The Fader network (Figure 1) is
1. What are the changes that we observe when an object a ∈ A based on an encoder-decoder structure with a domain adversar-
evolves into an object b ∈ B? ial aspect that allows us to learn and manipulate images based
2. What are the physical parameters that can explain these on physical properties which have to be converted into binary
changes? labels. The key is that the Fader network is able to learn and vi-
sualise differences between two data distributions. Images x are
Using our proposed method we can address both questions in a input to a neural network called encoder E(x). The encoder E(x)
data-driven way: we first train a network on populations A and B performs a mapping from the space of images to a latent rep-
to visualize the differences between those two data distributions, resentation of fixed dimension. The decoder D(E(x), y), also a
that is, we can transform individual objects a ∈ A ⇒ T (a) ∈ B. neural network, tries to reconstruct the image x given its latent
Comparing a and T (a) we can use our domain knowledge and representation E(x) and its label y (e.g. binary sSFR, dust). Thus
hypothesize possible physical parameters that can explain the the latent space is split into two parts. One part which should
changes that we see. With our set of possible parameters, x, y, z, contain all the information on the labels and another part which
we can train the network to perform transformations X, Y, Z should contain all the salient information needed to reconstruct
based on these parameters. We then apply these transforms and the object. To perform this disentanglement, another neural net-
compare X(A), Y(A), Z(A) to B and use some statistical measure work called discriminator Dis(E(x)) is trained to predict the la-
to tell us how different they are from B. If, for example, trans- bel y from the latent code E(x). Given pairs of images and binary
form X(A) is the closest to B it is most likely to be a good ex- labels {x, y} the Fader network tries to minimise two objectives:
planation for, though not a proof of, how A evolves into B. If
none of the transforms are sufficient, for example, the distance 1X
between the data distributions is too large, then either all trans- Lae = − kD(E(x), y) − xk22 − λE log(P(1 − y|E(x))), (1)
m
forms are not a good explanation, or alternatively the training
data or network structure was not sufficient to learn X, Y, Z well
enough to test it, a possibility which affects all simulation-based 1X
hypotheses. The ability to generate and test hypotheses using this Ldis = − log(P(y|E(x))). (2)
m
Article number, page 2 of 4
 Kevin Schawinski et al.: Exploring galaxy evolution with generative models

The first term in Lae is the reconstruction loss and measures

how well the auto-encoder can reconstruct the original input.
The second term is the domain adversarial component and by
minimising it should become impossible to predict the physi-
cal property/label y from the latent code E(x). At the same time
the discriminator becomes better at predicting the physical prop-
erty/label from the latent code E(x) by minimizing Ldis . This ad- 60
versarial interplay between the two loss functions is what allows blue field
50 increased dust
the fader network to disentangle salient and label information in
the decoding process. During inference, the property/label infor- 40 lowered ssfr
red satellite

Number
mation can be continuous and changing it will resemble what the 30
original image would look like with the changed physical prop-
erty/label. 20
10
3. Specific scientific application 0
0.5 0.0 0.5 1.0 1.5
As a demonstration, we choose the question of satellite quench- Classifier Prediction
ing in galaxy formation. This problem has several advantages: it
is relatively well understood, it relies on changes in the imaging Fig. 3. (top) The original blue field galaxy images (left), and the real
red satellites (right). We use the two Fader networks to increase the
data and associated attributes, which are easy to visualize, and it
dust (left centre) and lower the sSFR (right centre). The galaxies with
has been probed both by observations and simulations in the as- lowered sSFRs are qualitatively closer to the real red satellites. (bottom)
trophysics literature. When a galaxy enters a high-density envi- The same information from the top panel shown quantitatively using a
ronment such as a group or cluster, its specific star formation rate ridge classifier. We show the classifier prediction distributions of the
(sSFR) is likely to drop. This process is known as environment real blue field and red satellite distributions (blue and red, respectively),
quenching and represents a subset of the overall quenching pro- and the results of the two Fader transforms of the real blue field galaxies.
cess in galaxies (Strateva et al. 2001; Bell et al. 2003; Faber et al. The increased dust population (yellow) is very different from the real
2007; Martin et al. 2007; Schawinski et al. 2014) . This quench- red satellites, while the lowered sSFR Fader population (grey) is very
ing is associated with a number of structural changes: chiefly, close. This suggests that lowering the sSFR is a better explanation for
the increasing prominence of a central bulge. These effects have the formation of red satellite galaxies.
been well studied in the observations (Gunn & Gott 1972; Oem-
ler 1974; Davis & Geller 1976; Kauffmann et al. 2003; Bamford label 0 label 1
log Mhalo < 11.8 log Mhalo > 12.5
et al. 2009; Peng et al. 2010; Woo et al. 2013) and explored us- Environment
R pro jL = 0.0 kpc 1.0 kpc < R pro jL < 500 kpc
ing numerical simulations (Moore et al. 1996, 1998; Tonnesen sSFR log sS FR > −10.5 log sS FR < −11.7
& Bryan 2009; De Lucia et al. 2012). We view the results from Dust ebvmgas < 0.15 ebvmgas > 0.35
both observational studies and simulations as a baseline to test Table 1. Overview of the selection criteria used to learn the environ-
whether our data-driven approach can lead us to similar results. ment, sSFR, and dust transform. Label 0 refers to field galaxies and
If yes, then our approach has comparable utility for exploring label 1 refers to satellite galaxies.
astrophysical phenomena.

4. Experiment Oh et al. (2011) using a sample of 1,452 SDSS galaxies in the

same redshift range (exact criteria Table 1).
Following the general outline proposed earlier, we take a popu-
That way we have learned two transforms, one which shuts
lation of 26,706 galaxies in the redshift range 0.02 < z < 0.05
down the SFR of a galaxy, and another one which increases the
with stellar masses log Mstellar > 10.0 from the Sloan Digital Sky
amount of dust. We verify that both sSFR and dust transform are
Survey (York et al. 2000; Kauffmann et al. 2003; Brinchmann
learned correctly by training a ridge classifier to predict the sSFR
et al. 2004; Alam et al. 2015). We train the Fader network on the
and dust physical property. The ridge classifier is a standard and
environment by using samples of galaxies in the field and satel-
widely used linear classifier. We train it on labelled images so
lites in groups and clusters (exact criteria Table 1) (Yang et al.
that it predicts the correct label given the image. We adjust the
2007). The underlying assumption is that field galaxies turn into
regularization strength to maximise the accuracy of the classi-
satellite galaxies. We then take some field galaxies and use the
fier. We see that both the sSFR and the dust distribution become
trained network to show us what they would look like if they
shifted to the right, that is, lowered sSFR and increased dust.
became satellites (Figure 2).
We can now test these hypotheses by selecting a sample of
Looking at Figure 2, we note that as field galaxies become
1,476 blue field galaxies (u − r < 1.9) and 1,476 red satellite
satellite galaxies, they become redder, and their bulges become
galaxies (u − r > 2.6). We then train a ridge classifier to predict
more prominent. Using our domain knowledge, we can hypoth-
whether a galaxy is a blue field or a red satellite. We apply both
esize two possible parameters which can explain the change in
transformations to a test set of 148 blue field galaxies, where
colour:
we select the label parameter y ∈ [0, 1] such that the median of
1. Shut down of SFR (quenching), or the transformed distribution lies closest to the median of the red
2. increased amount of dust. satellite distribution. This results in y∗sS FR = 0.64 and y∗dust =
1.00.
To test these hypotheses, we train the Fader network with We find that the classifier has difficulty in differentiating be-
sSFR measurements from Kauffmann et al. (2003) using a sam- tween the artificially lowered sSFR sample and the set of red
ple of 11,240 SDSS galaxies, and with dust measurements from satellites. The dust transform on the other hand is clearly not
Article number, page 3 of 4
 A&A proofs: manuscript no. 33800corr

able to explain the colour transformation (Figure 3). This sup- Gunn, J. E. & Gott, III, J. R. 1972, ApJ, 176, 1
ports the hypothesis that it is a change in sSFR which changes the Kauffmann, G., Heckman, T. M., White, S. D. M., et al. 2003, MNRAS, 341, 33
colours and morphologies of galaxies as they enter high-density Kingma, D. P. & Welling, M. 2013, arXiv preprint arXiv:1312.6114
Lample, G., Zeghidour, N., Usunier, N., et al. 2017, ArXiv e-prints
environments, in concordance with what we know from both ob- [arXiv:1706.00409]
servations and simulations. Martin, D. C., Wyder, T. K., Schiminovich, D., et al. 2007, ApJS, 173, 342
What we have shown is that an astrophysicist can address Mathieu, M. F., Zhao, J. J., Zhao, J., et al. 2016, in Advances in Neural Infor-
a complex problem such as galaxy quenching, and make rapid mation Processing Systems 29, ed. D. D. Lee, M. Sugiyama, U. V. Luxburg,
I. Guyon, & R. Garnett (Curran Associates, Inc.), 5040–5048
progress in testing hypotheses using our method. Not only Moore, B., Katz, N., Lake, G., Dressler, A., & Oemler, A. 1996, Nature, 379,
is training a neural network much faster and computationally 613
less expensive than running a hydrodynamical simulation, it Moore, B., Lake, G., & Katz, N. 1998, ApJ, 495, 139
also does not rely on strong assumptions about the underlying Oemler, Jr., A. 1974, ApJ, 194, 1
physics, or suffer from limitations arising from coarse resolu- Oh, K., Sarzi, M., Schawinski, K., & Yi, S. K. 2011, ApJS, 195, 13
Peng, Y.-j., Lilly, S. J., Kovač, K., et al. 2010, ApJ, 721, 193
tion. Similarly, our approach takes much greater advantage of Rezende, D. J., Mohamed, S., & Wierstra, D. 2014, arXiv preprint
the data than conventional model fitting to observations. Never- arXiv:1401.4082
theless, we were able to reach similarly robust conclusions to the Schawinski, K., Urry, C. M., Simmons, B. D., et al. 2014, MNRAS, 440, 889
baseline observational and simulation-based studies of the sub- Strateva, I., Ivezić, Ž., Knapp, G. R., et al. 2001, AJ, 122, 1861
Tonnesen, S. & Bryan, G. L. 2009, ApJ, 694, 789
ject. Woo, J., Dekel, A., Faber, S. M., et al. 2013, MNRAS, 428, 3306
Yang, X., Mo, H. J., van den Bosch, F. C., et al. 2007, ApJ, 671, 153
York, D. G., Adelman, J., Anderson, Jr., J. E., et al. 2000, AJ, 120, 1579
5. Limitations & outlook
Our approach also has some limitations. First, we ‘solved’ an
already fairly well-understood problem. We chose this because
if we had used it to approach an unsolved problem and claimed
new physical insight, it might not have been clear that either our
method or the insight offered was reliable. This does not mean
that our approach is necessarily limited to such well-understood
problems. Second, we stress that our approach can only help us
test hypotheses, not prove them in a mathematical sense. Third,
as with any scientific observations, there is scope for confusion
between real physical effects and deficiencies and biases in the
training data and the network architecture. Finally, we highlight
that this approach is not fully automatic, and domain knowledge
by the user is still required.
Nevertheless, we believe our approach of using generative
models like the Fader network to forward model physical pro-
cesses and test hypotheses in a data-driven way has significant
potential in astrophysics and other fields. Its central advantage is
its data-driven nature which makes no assumptions on the under-
lying physics. As we have shown, human insight is still required
for high-level interpretation.
Acknowledgements. K.S. acknowledges support from Swiss National Science
Foundation Grants PP00P2_138979 and PP00P2_166159 and the ETH Zurich
Department of Physics. C.Z. and the DS3Lab gratefully acknowledge the sup-
port from the Swiss National Science Foundation NRP 75 407540_167266, IBM
Zurich, Mercedes-Benz Research & Development North America, Oracle Labs,
Swisscom, Zurich Insurance, Chinese Scholarship Council, the Department of
Computer Science at ETH Zurich, and the cloud computation resources from
Microsoft Azure for Research award program. The SDSS data used, the Jupyter
notebooks with the Fader network implementation, and instructions for how to
run them can be found at [Link]

References
Alam, S., Albareti, F. D., Allende Prieto, C., et al. 2015, ApJS, 219, 12
Antipov, G., Baccouche, M., & Dugelay, J.-L. 2017, in ICIP 2017, IEEE Inter-
national Conference on Image Processing, 17-20 September 2017, Beijing,
China / Also on ArXiv, Beijing, CHINA
Bamford, S. P., Nichol, R. C., Baldry, I. K., et al. 2009, MNRAS, 393, 1324
Bell, E. F., McIntosh, D. H., Katz, N., & Weinberg, M. D. 2003, ApJS, 149, 289
Bowman, S. R., Vilnis, L., Vinyals, O., et al. 2016, in Proceedings of the 20th
SIGNLL Conference on Computational Natural Language Learning, CoNLL
2016, Berlin, Germany, August 11-12, 2016, 10–21
Brinchmann, J., Charlot, S., White, S. D. M., et al. 2004, MNRAS, 351, 1151
Davis, M. & Geller, M. J. 1976, ApJ, 208, 13
De Lucia, G., Weinmann, S., Poggianti, B. M., Aragón-Salamanca, A., & Zarit-
sky, D. 2012, MNRAS, 423, 1277
Faber, S. M., Willmer, C. N. A., Wolf, C., et al. 2007, ApJ, 665, 265

Article number, page 4 of 4