Exploring the Latent Space of Autoencoders with

Interventional Assays

Felix Leeb∗† Stefan Bauer † Michel Besserve † Bernhard Schölkopf †

arXiv:2106.16091v4 [cs.LG] 11 Jan 2023

Abstract
Autoencoders exhibit impressive abilities to embed the data manifold into a low-
dimensional latent space, making them a staple of representation learning methods.
However, without explicit supervision, which is often unavailable, the representa-
tion is usually uninterpretable, making analysis and principled progress challenging.
We propose a framework, called latent responses, which exploits the locally contrac-
tive behavior exhibited by variational autoencoders to explore the learned manifold.
More specifically, we develop tools to probe the representation using interventions
in the latent space to quantify the relationships between latent variables. We extend
the notion of disentanglement to take the learned generative process into account
and consequently avoid the limitations of existing metrics that may rely on spuri-
ous correlations. Our analyses underscore the importance of studying the causal
structure of the representation to improve performance on downstream tasks such
as generation, interpolation, and inference of the factors of variation.

1 Introduction
Autoencoders (AEs) [1, 2] and its modern variants like the widely used variational autoencoders
(VAEs) [3], are a powerful paradigm for self-supervised representation learning for generative
modeling [4], compression [5], anomaly detection [6] or natural language processing [7]. Since
autoencoders can learn low dimensional representations without requiring labeled data, they are
particularly useful for computer vision tasks where samples can be very high dimensional making
processing, transmitting, and search prohibitively expensive. Here VAEs have shown impressive
results, often achieving state-of-the-art results compared to other paradigms [8, 9, 10, 11].
The striking performance coupled with the relatively flexible approach has prompted an explosion
of variants to learn a representation with some structure that is particularly conducive to a given
task [12, 13, 14]. In addition to a meaningful lower-dimensional representations of our world [15],
the focus may be to improve generalization [16, 17, 18], increase interpretability by disentangling the
underlying mechanisms [19, 20, 21, 22], or even to enable causal reasoning [23, 24].
In designing ever more intricate training objectives to learn more specialized structures as part of
complicated model pipelines, it becomes increasingly important to gain a better understanding of
what the representation actually looks like to more quickly identify and resolve any weaknesses of
a proposed method. Here the manifold learning community provides a principled formulation to
analyze and control the geometry of the data manifold learned by the representation [25, 26, 27, 28,
29, 30, 31, 32, 33]. Developing tools to better understand the structure of the representation is not
only useful as a diagnostic to identify avenues for improving modelling and sampling [34, 35], but it
also has crucial importance for fairness [36, 37] and safety [18, 38, 39].
Our focus here is on taking advantage of common properties of autoencoders to gain a deeper
understanding of the structure of the representation. We summarize our contributions as follows:
∗
Email: [email protected]
†
Max Planck Institute for Intelligent Systems, Tübingen, Germany

36th Conference on Neural Information Processing Systems (NeurIPS 2022).

 • We propose a framework, called latent responses, which exploits the locally contractive
behavior of autoencoders to distinguish the informative components from the noise in the latent
space and to identify the relationships between latent variables.
• We develop tools to analyze how the data manifold is embedded in the latent space by estimating
the extrinsic curvature which also enables semantically meaningful interpolations.
• Where true labels are available, we use conditioned latent responses to assess how each true
factor of variation is encoded in the representation and introduce the Causal Disentanglement
Score to quantify how disentangled the learned generative process is.
We release our code at https://github.com/felixludos/latent-responses.

2 Background
Representation learning begins with a set of N observation samples x ∈ X ⊆ RD which originate
from some unknown stochastic generative process with distribution x ∼ p(X) and support X . This
data manifold X is embedded into a low-dimensional (d << D) latent space Rd , and is modelled by
the support Z of an encoder f : X 7→ Z with the decoder learning the inverse mapping g : Z 7→ X̂
where after training X̂ ≈ X .

Variational
R Autoencoders (VAEs) [3, 40] are a framework for optimizing a latent variable model
p(X) ≈ Z p(X | Z; θ)p(Z)dZ with parameters θ, typically with a fixed prior p(Z) = N (Z; 0, I),
using amortized stochastic variational inference. A variational distribution q(Z | X; φ) with parame-
ters φ approximates the intractable posterior p(Z | X). The encoder and decoder are parameterized
such that q(Z | X = x; φ) = N (Z; f φ (x), σ φ (x)), and E [p(X | Z = z; θ)] = g θ (z) where f φ (x)
and σ φ (x), and g θ (z) are neural networks which are jointly optimized using the reparameteriza-
tion trick [3] to maximize the ELBO (Evidence Lower BOund) which is a lower bound to the log
likelihood:
log p(X; θ) ≥ Eq(Z | X;φ) [log p(X | Z; θ)] − DKL (q(Z | X; φ)kp(Z)) = LELBO
θ,φ (X) . (1)

Note that in practice, p(X) is unknown, and we only have access samples {x(i) }N i , so p(X) is
PN
approximated by π(X = x) = N1 i=1 δ(x − x(i) ). The first term in the objective corresponds to
a reconstruction loss, while the second can be interpreted as a regularization term to encourage the
posterior to match the prior.

2.1 Related Work

Representation geometry A closely related approach is manifold learning which aims to exploit
the geometry of the data manifold, usually by regularizing the geometry of the representation by
estimating the intrinsic curvature of the data manifold [26, 27, 28, 29, 30], or by improving sampling
and interpolation using the the Riemmanian metric [31, 32, 25, 33]. In comparison, our response
maps estimate the extrinsic curvature, which focuses on the specific embedding, rather than being an
intrinsic property of the dataset (see appendix A.4).

Interpretability and Disentanglement Another general approach is to gain a better understanding

of the representation by making it more interpretable, by, for example, disentangling the true factors
of variation [19, 22, 41, 42, 43, 44]. While our analysis is more similar to approaches that improve
the representation by focusing on the structure of the representation, for example by learning extra
post-hoc models in the latent space [45, 46, 47], we develop a metric to evaluate how disentangled
the learned generator is, rather than just the encoder.

Autoencoder Consistency VAEs have been investigated from an information theoretic viewpoint
[48, 49] and with respect to training problems like posterior collapse [50] or the holes problem
[51] to better understand common failure modes. Similarly, mismatches between the encoder and
decoder [52, 53], have spurred research into increasing the self-consistency of autoencoders [54,
55, 56, 57, 58]. Our latent response framework relies on a very similar approach and formulation,
but crucially, thus far these methods focus on regularizing the training of the representation to
impose certain desired structure, while we focus on analyzing the structure that is learned rather than
modifying the training objective, making our tools directly applicable to virtually all VAEs.

2
3 Latent Responses
On a high level, to explore the structure of how the data manifold is embedded in the latent space, it
is necessary to separate the semantic information from the noise in the latent space. So our goal is to
decompose the latent variables Z into an endogenous S and exogenous U component, which, for
simplicity, we choose to relate to one another as shown in equation 2.
Z =S+U (2)
Conceptually, S should capture the semantic information necessary to to reconstruct the sample X,
while U is a local noise model in the latent space for a given observation X which does not mean-
ingfully affect the semantics. In the context of VAEs with a Gaussian prior, we propose equations 3
and 4 which recovers the familiar posterior for VAEs q(Z | X = x; φ) = q(S | X; φ)q(U | X; φ) =
N (f φ (x), σ φ (x)).
q(S = s | X = x; φ) := N (S = s; f φ (x), 0) = δ(f φ (x) − s) (3)
φ
q(U = u | X = x; φ) := N (U = u; 0, σ (x)) (4)
One implication of separating the deterministic and stochastic part of the encoder is a new perspective
on the training signal for the decoder from the VAE objective. Substituting our definitions for S and
U into the reconstruction loss term (shown in equation 5) reveals the decoder is trained to map all
samples from the posterior to match the same observation sample x(i) . As a result, the decoder learns
to filter out any exogenous component u from z around s, making the latent space locally contractive
around S to the extent of U . Although we observe this as a byproduct of the VAE objective, this
contractive behavior is even observed in unregularized autoencoders to some extent [59] suggesting
this may be a more fundamental feature of the inductive biases in deep autoencoders.
h i
Eu∼q(U | X=x(i) ;φ) log p(X = x(i) | s(i) + u; θ) (5)

where s(i) = f φ (x(i) ) is the (deterministic) latent code corresponding to the observation x(i) .
Starting from some latent sample z ∼ p(Z), we would like to separate the constituent exogenous
u and endogenous s components. If we had the matching x such that z ∼ q(Z | X = x; φ), then
the separation would be trivial since, by definition s = f φ (x) and u = z − s. However, since the
VAE is optimized to reconstruct observations from the latent space, we approximate the missing x
with x̂ = g θ (z). Subsequently, we encode the generated sample x̂ to infer an approximation of s,
ŝ = f φ (x̂). Now by expanding g θ around z and f φ around x to the first order, we are left with three
terms shown in equation 6 (neglecting higher order terms, full derivation in appendix A.2).

ŝ = s + Jf φ (x)(g θ (s) − x) + Jf φ (x)Jgθ (s)u +O 2 + O u2

   
(6)
1 2 3

where  = x̂ − x is the reconstruction error, Jf φ (x) is the jacobian of f φ evaluated at x and Jgθ (s)
is the jacobian of g θ evaluated at s.
Term 1 aligns with our conceptual interpretation of s as the semantic information should remain
invariant to any noise information also contained in z. Meanwhile the second and third terms
correspond to two different sources of error to this interpretation we must potentially take into
account. The term 2 derives from the encoder struggling to encode samples that were not seen
during training, since π(X) 6= p(X̂ | Z; θ)p(Z). However, provided the reconstructions sufficiently
faithful across latent space (i.e. g θ (s)−x is small), this term may be ignored. This error can be further
mitigated by training the encoder to be more robust with respect to the input (diminishing Jf φ (x))
with mild additive noise on the input observations or additional regularization terms [54, 52, 58].
Finally, term 3 in equation 6 originates from the decoder having to filter out the stochastic exogenous
information from z when decoding. As discussed above in equation 5, the VAE objective already
directly minimizes this term by training the decoder output to be invariant to noise samples u, thus
diminishing Jgθ (s).
The bottom line is, as long as the decoder can filter out the exogenous information u from latent sam-
ples z and the encoder can recognize the resulting generated samples x̂, the endogenous information

3
s is preserved. We call this process of decoding and reencoding latent samples, hφθ = f φ ◦ g θ the
latent response function. Crucially, the latent response function allows us to extract the semantic
information from the latent space without knowing how the information is encoded, so although we
can identify the structure of the representation, the representation is not necessarily interpretable
without ground truth label information. A statistical treatment of this phenomenon is discussed in
the appendix, however provided the error terms are sufficiently small the corresponding response
distribution r(Ẑ | Z; θ, φ) is shown in equation 7.
r(Ẑ | Z; θ, φ) = Ep(X̂ | Z;θ) q(Ẑ | X̂; φ) (7)
One subtlety of our interpretation remains to be addressed: how ambiguities due to overlapping pos-
terior distributions are resolved by latent responses. To match the prior, there will be overlap between
posterior distributions which correspond to semantically different samples. Statistically speaking,
these ambiguities are captured by the higher moments of r(Ẑ | Z; θ, φ). Despite Er(Ẑ | Z;θ,φ) [Ŝ] ≈ S,
the variance can be interpreted to relate to the uncertainty of the encoder in the inference of the
endogenous variable from the generated sample, suggesting a potential signal for anomaly detec-
tion [55]. From another perspective, a strong deviation between Ŝ and S implies there is some
inconsistency between the encoder and decoder, which identifies the "holes" in the latent space [51],
where the encoder has trouble recognizing the samples generated by the decoder.

3.1 Interventions intervention

response

To control the learned generative process, we

need to know how to manipulate specific se-
mantic information in the representation. We
conceptualize these manipulations as interven-
tions where a chosen latent variable is modified
while all others remain unchanged. For exam-
ple, given latent sample z, ∆(zj ←z̃j ) (z) refers Figure 1: Encoded data points are shown as black
to the interventional sample where the jth latent points with the surrounding latent manifold shaded
variable is resampled from the marginal of the blue. An intervention depicted as the orange ar-
aggregate posterior z̃j ∼ q(Zj ; φ). row might replace the semantic information along
the horizontal latent dimension with that of the
When intervening on latent variables Z, there sample with a yellow border. In this case, the in-
are two possible outcomes: either the resulting tervention results in the latent sample leaving the
generated sample is affected significantly, that is latent manifold. Now the latent response function
to say, the semantic information S is affected by approximately filters out the exogenous noise to ef-
the intervention, or there is no significant change fective project the sample back onto the manifold.
in which case only the noise was affected. In Note that this projection (green arrow) changes
the first case, we identify a specific kind of in- both the horizontal and vertical dimensions, from
tervention to manipulate the learned generative which we infer there is a non-trivial relationship
process. between the horizontal and vertical.

Latent Response Matrix This brings us to the first practical tool we introduce based on latent
responses, which aims to describe to what extent the latent variables causally affect one another
with respect to the learned generative process. Each element in the matrix M ∈ Rd∗d quantifies the
degree to which an intervention in latent variable j causes a response in latent variable k as seen in
equation 8, where hθφ k refers to the kth variable of the latent response (see figure 1).
1 h i
Mjk2
= Ez∼p(Z);z̃j ∼p(Zj ) |hφθ k (∆ (zj ←z̃j )
(z)) − hφθ
k (z)|2
(8)
2
Along the diagonal, Mjj can be interpreted as quantifying the extent to which an intervention along
the jth latent variable is detectable at all. As the value approaches 0, changes in the latent variable do
not affect the generated sample in any way detectable by the encoder. For VAEs, this is frequently
due to posterior collapse [60, 61, 50, 62]. On the other hand, if the latent variable is maximally
informative and the intervention z̃j is fully recoverable it implies negligible exogenous noise, so
then hφθ φθ
j (z̃) ≈ z̃j and hj (z) ≈ zj , and Mjj approaches 1, since both z̃j and zj are sampled
independently from a standard normal, and the elements are normalized with a factor of 12 .

4
Perhaps even more interesting, the off-diagonal elements of Mjk show to what extent an intervention
on variable j affects variable k. Consequently, the latent response matrix can be interpreted as a
weighted adjacency matrix of a directed graph of the learned structural causal graph. Note that there
may be cycles in this graph if there is some non-trivial relationship between latent variables. For
example, the model could embed a periodic variable into two latent dimensions, such as j1 and j2 , to
keep the representation continuous. In that case, we would expect an intervention along dimension
j1 to elicit a similar response in j2 as the response of j1 from an intervention on j2 . Consequently,
Mj1 j2 ≈ Mj2 j1 > 0 suggests j1 and j2 should be treated jointly as a single latent variable. Thus, the
latent response matrix not only describes the causal structure of the learned generative process, but it
can also identify more complex relationships between latent variables for further analysis.

Conditioned Response Matrix While the latent response matrix Mjk lets us quantitatively com-
pare how much each latent dimension affects another, without manual inspection or label information,
the correspondence between the learned causal variables the true causal factors is, in general, un-
known [19].
However, when we have access to the true generative process, or at least the ground truth labels
∗
y (i) ∈ Y ⊆ Rd corresponding to the observation samples x(i) , then there is variant of the latent
response matrix, termed the conditioned response matrix, to quantify how well the learned variables
match with the true ones, which is closely related to the disentanglement of the representation.
The key is to carefully select our interventions such that they only affect one true factor at a time, and
then evaluate to what extent these interventions for each of the latent variables are still detectable.
Intuitively, if a latent variable Zj only captures information pertaining to factor Yc , then if we select
interventions that only change factor Yc0 where c0 6= c, then interventions on Zj do not produce a
response, so Mjj ≈ 0.
To condition the set of interventions on a specific factor Yc , we choose a subset of observations
which are all semantically identical except for a single factor Yc , x ∼ p(X | Yc , Y−c )p(Yc )p(Y−c )
where p(X | Y ) refers to the true generative process given label Y and Y−c refers to all
true causal variables except Yc . Then the latent response matrix is computed with interven-
tions
R exclusively sampled from the resulting aggregate posterior of this subset q(Z | Y−c ; φ) =
q(Z | X; φ)p(X | Yc , Y−c )p(Yc )dXdYc .
∗2 1 h i
Mcj = Ez∼p(Z);z̃j ∼q(Zj | Y−c ;φ)p(Y−c ) |hφθ
j (∆ (zj ←z̃j )
(z)) − hφθ
j (z)|2
(9)
2
In the context of controllable generation, the conditioned response matrix quantifies how much an
intervention on each latent variable can affect each of the true factors of variation. Ideally, each
true causal factor would only be manipulable by disjoint subsets of the latent variables, which is
commonly referred to as "disentangling" the factors of variation. From the conditioned response
matrix, we can identify not only which latent variables contain information pertaining to each of the
true factors of variation, but also which factors are affected by an intervention in each of the latent
variables.

Causal Disentanglement Score To more easily compare representations, we can aggregate the
information in the conditioned response matrix into a single value to measure the degree to which the
representation is able to disentangle the true factors of variation. The causal disentanglement score
(CDS) in equation 10 which allows each latent variable to causally affect a single factor, but penalizes
any additional responses. As written, the score is between d1∗ and 1, but we re-scale it to [0, 1].
∗
P
j maxc Mcj
CDS = P ∗ (10)
cj Mcj

(To our knowledge) none of the existing disentanglement metrics take the learned generative process
into account at all. If the task is only to infer the true labels from the observations (as is common),
then the decoder is admittedly superfluous, which is why disentanglement methods generally focus on
the encoder. However, in tasks such as controllable generation, where disentanglement is obviously
valuable, the behavior of the decoder is critical. Here, the main risk in evaluating disentanglement
from the encoder alone is that some latent variables may be correlated with true factors of variation
without them having any causal effect on those factors when generating new samples. These spurious

5
correlations also potentially decrease the resulting disentanglement score, which may falsely penalize
larger representations.
The conditioned response matrix and associated CDS mirrors the responsibility matrix and disentan-
glement score introduced by [42]. However, crucially, the responsibility matrix identifies how well
each latent variable correlates with each true factor of variation.

Response Maps The last type of analysis we propose focuses on gaining a qualitative understanding
of how the data manifold is embedded in the latent space. Specifically, we exploit the ability to probe
the latent manifold using the response function to map out the manifolds extent, including estimating
its extrinsic curvature. From the definition of s = f φ (x) where s ∈ Z, x ∈ X and f φ is a smooth
deterministic function, we may interpret s as a projection of the data manifold X into the latent space.
Usually, our analysis of Z is limited to the observation samples of x we have access, from which man-
ifold learning methods often estimate the intrinsic curvature of the data manifold X . However, using
latent responses, we can use any latent sample z ∼ p(Z) to probe Z as long as ŝ ≈ s. Rearranging
the terms gives us u(z) in equation 11, whose magnitude can be interpretted as the unsigned distance
function to the latent manifold, which is evocative of the neural implicit functions [63, 63, 64]
suggesting a variety of further tools we leave for future work.
u(z) = s − z ≈ hφθ (z) − z (11)
Treating |u(z)| as an approximate distance function to the manifold, we compute the mean curvature
H of the latent manifold (see appendix for further discussion). In practice, we estimate the necessary
gradients by finite differencing across a 2D grid in the latent space, we call the resulting map the
response map, visualized similarly as in [65].
Qualitatively, with our sign convention, high curvature corresponds to regions where u(z) is small
and locally convergent, and consequently where we find the data manifold. Empirically, we also find
the divergence of u(z), which is closely related to the mean curvature, and is useful for identifying
the regions of the latent space where u diverges, which may be interpreted as possible "holes" in the
latent space.
When interpolating between latent samples z, ideally we would follow the geodesic of the data
manifold. This is can done by estimating the Riemannian metric from the data samples and integrating
an expensive ODE to find the geodesic connecting samples [25]. The Riemmanian metric relates to
the intrinsic curvature of the data manifold, which is independent of the embedding, consequently
guaranteeing we find the geodesic independent of the representation. However, in our case, we use
the latent responses to estimate the extrinsic curvature which does depend on the embedding, so the
resulting path may not be optimal with respect to the underlying data manifold. Nevertheless, we
optimize the path along the response map to stay in high curvature regions, thereby effectively are
finding a path in the latent space which stays near the data manifold.

4 Toy Example: The Double Helix

To illustrate how the latent response framework can be used to study the representation learned by
a VAE, we show the process when learning a 2D representation for samples from a double helix
embedded in R3 . Disregarding the additive noise, the data manifold has two degrees of freedom (data
manifold formally defined in the appendix).This analysis is largely independent of the precise neural
network architecture, provided the model has sufficient capacity to learn a satisfactory representation
(hyperparameters in the appendix).
Figure 2 provides an example for how the response maps can be used to trace the latent representation
through the mean curvature and divergence. Note that the magnitude of the response |u(z)| is not
sufficient to identify the latent manifold since both the regions with maximal and minimal curvature
have a minimal response. The divergence shows most of the latent space has a slightly negative
divergence, implying most of the latent space converges, rather than diverges, which is consistent
with expectations. The mean curvature shows what regions of the latent space the map converges to
in yellow, from which we can recognize the structure of the learned latent manifold. Finally, the right
most plot shows the aggregate posterior q(S | X; φ) of the 1024 training samples.
This toy example also motivates the value of meaningful interpolations as seen in figure 3. The
path in red shows the shortest euclidean path between the two samples in orange, but note that the

6
 Magnitude of Response Divergence Mean Curvature Posterior
2.0 0.20
0.8
1.5 1.4 0.15 0.6
1.0 1.2 0.10 0.4
0.5 1.0 0.05 0.2
0.0 0.8 0.00 0.0
0.5 0.6 -0.01 0.2
1.0 0.4 -0.02 0.4
1.5 0.2 -0.03 0.6

2.0 -0.05 0.8

2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2

Figure 2: Depicted are three quantities derived using the latent response function compared to the
aggregate posterior shown on the far right for the representation learned by a VAE for the double
helix toy dataset. Note that almost all the density of the posterior is in regions that have positive
curvature, corresponding to the data manifold.
resulting path in the observation space jumps from one strand to another twice. Meanwhile, the path
the maximizes the estimated mean curvature is shown in green and produces a much more reasonable
path.
2.0
1.5
1.0
1.0
0.5
0.5
0.0
0.0 0.5
0.5 1.0
1.0 1.0
0.5
0.00.5
1.5
1.0 1.0 0.5 0.0 0.5 1.0
2.0
2 1 0 1 2

Figure 3: The left plot shows the 2D latent space including the aggregate posterior density in
black, and two possible interpolations between the two pink points. Meanwhile, the plot on the right
shows the ambient space with the black points being the observed data samples, with the blue points
showing the reconstructed samples, and the paths in the ambient space corresponding to the ones in
the latent space. Note how the path in green follows the learned manifold and consequently much
more consistent in the ambient space compared to the shortest euclidean path in red.

5 Experiments & Results

Experimental Setup We apply our new tools on a small selection of common benchmark datasets,
including 3D-Shapes [66], MNIST [67], and Fashion-MNIST [68] 3 . Our methods are directly
applicable to any VAE-based model, and can readily be extended to any autoencoders. Nevertheless,
we focus our empirical evaluation on vanilla VAEs and some β-VAEs (denoted by substituting the
β, so 4-VAE refers to a β-VAE where β = 4). Specifically, here we mostly analyze a 4-VAE model
with a d = 24 latent space trained on 3D-Shapes (except for table 1) referred to as Model A, and
include results on a range of other models in the appendix. All our models use four convolution and
two fully-connected layers in the encoder and decoder. The models are trained using Adam [69] with
a learning rate of 0.001 for 100k steps (see appendix for details).

Qualitative Understanding The most direct way to get a better understanding the manifold struc-
ture is the visualization of the response maps, in particular with the mean curvature (see figure 4).
Unfortunately, since the curvature is estimated numerically using a grid of samples, the maps do
not scale well to the whole latent space. Here the latent response matrices help identify pairs of
related latent dimensions which can then be analyzed more closely with a response map. Furthermore,
currently, all these response maps are aligned to the axes of the latent space. Although VAEs do
align information along the axes somewhat [70], any off-axis structure is missed since the off-axes
responses are completely ignored. Since, an implicit requirement of disentangled representations
is that the information is axis-aligned [19], the response maps present the most striking results for
disentangled representations (see appendix for more examples).
3
All are provided with an MIT, Apache or Creative Commons License

7
 1.5 0.34
1.5
0.25
1.0 0.75
0.17 1.0

0.5 0.50
0.08 0.5
0.25
0.0 0.00
0.0 0.00
-0.01
0.5
0.5 0.25
-0.02
1.0 0.50
-0.02 1.0
0.75
1.5 -0.03
1.5 1.0 0.5 0.0 0.5 1.0 1.5 1.5
1.5 1.0 0.5 0.0 0.5 1.0 1.5
(c) Reconstructions over
(a) Divergence Map and Aggregate Pos-
(b) Mean Curvature Map the same space
terior
Figure 4: A projection along dimensions 16 (horizontally) and 22 for Model A (4-VAE) shows
the computed divergence of the response field in blue and red while the green points are samples
from the aggregate posterior. 4b shows the mean curvature, which identifies 10 points where the
curvature spikes and the boundaries between the regions corresponding to different clusters in the
posterior. Finally, from the corresponding reconstructions in 4c (with all other latent variables fixed)
it becomes clear that each of the clusters in the posterior corresponds to a different floor hue. Note
that, although the aggregate posterior is highly concentrated at a few points, the negative divergence
almost everywhere suggests the extent of U | X the decoder can handle extends well beyond the
posterior (as confirmed by the reconstructions).
1.5
1.5 Shortest
Best
0.8 1.0
1.0
0.6
0.5
Dimension 14

0.5 0.4
0.0
0.2
0.0
0.0 0.5 (c) Shortest Path Reconstructions
0.5
0.2
1.0
1.0 0.4
1.5
0.6 1.5 1.0 0.5 0.0 0.5 1.0 1.5
Dimension 13
1.5
1.5 1.0 0.5 0.0 0.5 1.0 1.5
(d) Best Path Reconstructions
(b) Interpolation between latent
(a) Mean Curvature Map.
samples on the manifold.
Figure 5: Using Model A (4-VAE) in a similar setting to figure 3, we now search for an interpolation
between two latent samples from the posterior along the latent manifold, visualized in 5a. Figure 5b
compares this best path (in green) to the shortest path in euclidean distance (in red), Finally, latent
samples along each of the paths at even intervals are decoded showing that the shortest path results in
blurry, unrealistic shadows in the middle of 5c compared to 5d.
From the mean curvature response map in figure 5a, we see that the manifold is particularly nonlinear
along these two latent dimensions (13 and 14). Consequently, there can be a dramatic difference
between the geodesic (here approximated using the mean curvature), compared to the shortest path in
euclidean space as seen in figure 5.

Causal Disentanglement Table 1, unsurprisingly, shows the disentanglement increasing with

increasing β. Our proposed CDS score correlates strongly with the other disentanglement metrics.
Perhaps noteworthy is that even though the CDS and DCI-D scores are computed in similar ways
(the vital difference being whether responsibility rests on a causal link or a statistical correlation), the
DCI-D scores are consistently lower than the CDS scores. This may be explained by the DCI-D metric

Name CDS DCI-D IRS MIG Table 1: Comparing disentanglement met-

1-VAE 0.44 0.3 0.44 0.07 rics for β-VAEs trained on 3D-shapes with
2-VAE 0.49 0.36 0.46 0.09 varying β. For the models in the first four
4-VAE 0.58 0.46 0.51 0.16 rows d = 12, while d = 24 for the remain-
8-VAE 0.71 0.66 0.63 0.21 ing four. While the CDS generally correlates
1-VAE 0.52 0.49 0.48 0.09 well with other metrics, notably, the DCI-D
2-VAE 0.61 0.58 0.52 0.15 score is consistently slightly lower which
4-VAE 0.72 0.67 0.57 0.17 may be due to spurious correlations between
8-VAE 0.78 0.73 0.64 0.2 latent variables and the true factors.

8
 2 4 9 13 14 15 16 20 21 22 23
2 4 9 13 14 15 16 20 21 22 23 floor_hue 0.00 0.00 0.00 0.00 0.00 0.00 0.44 0.00 0.00 0.56 0.00

2 0.96 0.04 0.13 0.02 0.03 0.03 0.03 0.02 0.06 0.02 0.03 wall_hue 0.00 0.00 0.00 0.00 0.00 0.54 0.00 0.00 0.00 0.00 0.46
object_hue 0.45 0.00 0.54 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
4 0.04 0.90 0.03 0.04 0.04 0.02 0.03 0.03 0.52 0.02 0.02
scale 0.00 0.13 0.00 0.00 0.00 0.00 0.00 0.86 0.00 0.00 0.00
9 0.13 0.07 0.98 0.02 0.03 0.02 0.02 0.03 0.06 0.02 0.02 13 9
shape 0.00 0.50 0.00 0.00 0.00 0.00 0.00 0.00 0.50 0.00 0.00 14
13 0.02 0.04 0.02 1.01 0.13 0.02 0.02 0.02 0.04 0.02 0.02 orientation 0.00 0.00 0.00 0.43 0.57 0.00 0.00 0.00 0.00 0.00 0.00
4
15
14 0.02 0.05 0.02 0.18 0.99 0.02 0.02 0.02 0.04 0.01 0.01 2
Intervention

16
15 0.04 0.04 0.03 0.05 0.05 1.05 0.04 0.03 0.05 0.03 0.16 (b) DCI Responsibility Matrix 23
20
16 0.04 0.06 0.04 0.02 0.02 0.03 0.98 0.05 0.07 0.18 0.02 21 22
2 4 9 13 14 15 16 20 21 22 23
20 0.04 0.17 0.03 0.03 0.04 0.03 0.03 1.00 0.08 0.02 0.02 floor_hue 0.00 0.00 0.00 0.01 0.01 0.00 1.04 0.02 0.00 1.06 0.00
wall_hue 0.00 0.00 0.00 0.01 0.01 1.10 0.00 0.02 0.00 0.00 1.06
(d) Graph correspond-
21 0.02 0.06 0.01 0.01 0.01 0.02 0.01 0.03 0.75 0.01 0.01
object_hue 0.98 0.00 0.99 0.00 0.01 0.00 0.00 0.02 0.00 0.00 0.00
ing to the Response
22 0.03 0.07 0.03 0.02 0.02 0.02 0.16 0.06 0.10 0.96 0.03
scale 0.00 0.02 0.00 0.00 0.02 0.00 0.00 0.99 0.00 0.00 0.00
Matrix
23 0.04 0.05 0.04 0.04 0.05 0.18 0.03 0.03 0.04 0.03 1.02
shape 0.00 1.13 0.00 0.01 0.02 0.00 0.00 0.21 0.84 0.00 0.00
Response
orientation 0.00 0.00 0.00 0.97 0.97 0.00 0.00 0.04 0.00 0.00 0.00

(a) Latent Response Matrix

(c) Conditioned Response Matrix
Figure 6: These are the (6a) latent response matrix, (6b) DCI Responsibility matrix [42], (6c)
conditioned response matrix, and the (6d) derived from the latent response matrix for Model A
(4-VAE) (d = 24) trained on 3D-Shapes. The responsibility matrix shows the predictability of
each latent dimension (column) for each factor of variation (row), while the conditioned response
matrix measures the effect an intervention in each latent dimension (column) has provided that the
intervention can changes a specific factor of variation (row). Note that from the conditional response
matrix we see only dimensions 16 and 22 are causally linked to the floor hue, for which the structure
is further visualized in figure 4.

taking additional spurious correlations between latent variables into account (as seen in figure 6),
while the CDS metric focuses on the causal links, so the DCI-D has an undeservedly low score.

Causal Structure Closer comparison between the conditioned response matrices and the responsi-
bility matrices reveal more how the DCI-D metric and CDS differ. Figure 6 shows the responsibility
matrix matches the conditioned response matrix for the most part. The only exception being latent
variables 4 and 20. Since the DCI framework identifies which latent variables are most predictive for
the true factors, it cannot distinguish between a correlation and a causal link. In this case, the DCI
metric recognized a correlation between dimension 4 and the "scale" factor. However, from the latent
response matrix (and the graph 6d) we see interventions on dim 20 have a significant effect on dim 4,
but not vice versa. Consequently, we identify dim 20 is a parent of dim 4 in the learned causal graph.
Since dim 20 is closely related with the "scale" factor the causal link to dim 4 results in dim 4 being
correlated with "scale". The conditioned response matrix correctly identifies that it is dimension 20
which primarily affects the scale, but indirectly also affects the shape through dimension 4. This
is a prime example of how causal reasoning can avoid misattributing responsibility due to spurious
correlations.

6 Conclusion
In this work, we have introduced and motivated the latent response framework including a variety of
tools to better visualize and understand the representations learned by variational autoencoders. Given
an intervention on a sample in the latent space, the latent response quantifies the degree to which that
intervention affects the semantic information in the sample. Therefore, we can think of this analysis
as leveraging the interventional consistency of a representation to study the geometric and causal
structure therein. Notably, the current analysis relies on a certain degree of axis-aligned structure in the
latent space, which makes these tools especially useful for understanding the structure of disentangled
representation. Another limitation is that computing the latent response maps to, for example,
improve interpolations, does not scale well for the large representations of high fidelity generative
models [8]. Consequently, our experiments thus far have focused on synthetic datasets designed
for evaluating disentanglement methods. However, latent responses do not require any ground truth
label information, which is particularly promising for better understanding representations of real
datasets and consequently speed up development of not just better performing representation learning
techniques, but also more interpretable and trustworthy [71] models.

9
Acknowledgements
This work was supported by the German Federal Ministry of Education and Research (BMBF):
Tübingen AI Center, FKZ: 01IS18039B, and by the Machine Learning Cluster of Excellence, EXC
number 2064/1 – Project number 390727645. The authors thank the International Max Planck
Research School for Intelligent Systems (IMPRS-IS) for supporting Felix Leeb.

References
[1] David E Rumelhart, Geoffrey E Hinton, and Ronald J Williams. Learning internal represen-
tations by error propagation. Technical report, California Univ San Diego La Jolla Inst for
Cognitive Science, 1985.
[2] Dana H Ballard. Modular learning in neural networks. In AAAI, pages 279–284, 1987.
[3] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint
arXiv:1312.6114, 2013.
[4] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images
with vq-vae-2. arXiv preprint arXiv:1906.00446, 2019.
[5] James Townsend, Tom Bird, and David Barber. Practical lossless compression with latent
variables using bits back coding. arXiv preprint arXiv:1901.04866, 2019.
[6] Jinwon An and Sungzoon Cho. Variational autoencoder based anomaly detection using recon-
struction probability. Special Lecture on IE, 2(1):1–18, 2015.
[7] Jiwei Li, Minh-Thang Luong, and Dan Jurafsky. A hierarchical neural autoencoder for para-
graphs and documents. arXiv preprint arXiv:1506.01057, 2015.
[8] Arash Vahdat and Jan Kautz. Nvae: A deep hierarchical variational autoencoder. Advances in
Neural Information Processing Systems, 33:19667–19679, 2020.
[9] Rewon Child. Very deep vaes generalize autoregressive models and can outperform them on
images. arXiv preprint arXiv:2011.10650, 2020.
[10] Ligong Han, Sri Harsha Musunuri, Martin Renqiang Min, Ruijiang Gao, Yu Tian, and Dimitris
Metaxas. Ae-stylegan: Improved training of style-based auto-encoders. In Proceedings of the
IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3134–3143, 2022.
[11] Ðord̄e Miladinović, Aleksandar Stanić, Stefan Bauer, Jürgen Schmidhuber, and Joachim M
Buhmann. Spatial dependency networks: Neural layers for improved generative image modeling.
International Conference on Learning Representations (ICLR), 2021.
[12] Diederik P Kingma and Max Welling. An introduction to variational autoencoders. arXiv
preprint arXiv:1906.02691, 2019.
[13] Dor Bank, Noam Koenigstein, and Raja Giryes. Autoencoders. arXiv preprint arXiv:2003.05991,
2020.
[14] Laurent Girin, Simon Leglaive, Xiaoyu Bie, Julien Diard, Thomas Hueber, and Xavier Alameda-
Pineda. Dynamical variational autoencoders: A comprehensive review. arXiv preprint
arXiv:2008.12595, 2020.
[15] Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and
new perspectives, 2012. URL arXiv:1206.5538.
[16] Andrea Dittadi, Frederik Träuble, Francesco Locatello, Manuel Wüthrich, Vaibhav Agrawal, Ole
Winther, Stefan Bauer, and Bernhard Schölkopf. On the transfer of disentangled representations
in realistic settings. arXiv preprint arXiv:2010.14407, 2020.
[17] Aravind Srinivas, Michael Laskin, and Pieter Abbeel. Curl: Contrastive unsupervised represen-
tations for reinforcement learning. arXiv preprint arXiv:2004.04136, 2020.
[18] Lukas Schott, Julius von Kügelgen, Frederik Träuble, Peter Gehler, Chris Russell, Matthias
Bethge, Bernhard Schölkopf, Francesco Locatello, and Wieland Brendel. Visual representation
learning does not generalize strongly within the same domain. arXiv preprint arXiv:2107.08221,
2021.

10
[19] Francesco Locatello, Stefan Bauer, Mario Lucic, Gunnar Rätsch, Sylvain Gelly, Bernhard
Schölkopf, and Olivier Bachem. Challenging common assumptions in the unsupervised learning
of disentangled representations. arXiv preprint arXiv:1811.12359, 2018.
[20] Ricky TQ Chen, Xuechen Li, Roger Grosse, and David Duvenaud. Isolating sources of
disentanglement in variational autoencoders. arXiv preprint arXiv:1802.04942, 2018.
[21] Emile Mathieu, Tom Rainforth, Nana Siddharth, and Yee Whye Teh. Disentangling disentan-
glement in variational autoencoders. In International Conference on Machine Learning, pages
4402–4412. PMLR, 2019.
[22] Wenqian Liu, Runze Li, Meng Zheng, Srikrishna Karanam, Ziyan Wu, Bir Bhanu, Richard J
Radke, and Octavia Camps. Towards visually explaining variational autoencoders. In Pro-
ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages
8642–8651, 2020.
[23] Bernhard Schölkopf, Francesco Locatello, Stefan Bauer, Nan Rosemary Ke, Nal Kalchbrenner,
Anirudh Goyal, and Yoshua Bengio. Toward causal representation learning. Proceedings of the
IEEE, 109(5):612–634, 2021.
[24] Mengyue Yang, Furui Liu, Zhitang Chen, Xinwei Shen, Jianye Hao, and Jun Wang. Causalvae:
Disentangled representation learning via neural structural causal models. In Proceedings of the
IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9593–9602, 2021.
[25] Georgios Arvanitidis, Lars Kai Hansen, and Søren Hauberg. Latent space oddity: on the
curvature of deep generative models. arXiv preprint arXiv:1710.11379, 2017.
[26] Tao Yang, Georgios Arvanitidis, Dongmei Fu, Xiaogang Li, and Søren Hauberg. Geodesic
clustering in deep generative models. arXiv preprint arXiv:1809.04747, 2018.
[27] Marissa Connor, Gregory Canal, and Christopher Rozell. Variational autoencoder with learned
latent structure. In International Conference on Artificial Intelligence and Statistics, pages
2359–2367. PMLR, 2021.
[28] Clément Chadebec, Clément Mantoux, and Stéphanie Allassonnière. Geometry-aware hamilto-
nian variational auto-encoder. arXiv preprint arXiv:2010.11518, 2020.
[29] Nutan Chen, Alexej Klushyn, Francesco Ferroni, Justin Bayer, and Patrick Van Der Smagt.
Learning flat latent manifolds with vaes. arXiv preprint arXiv:2002.04881, 2020.
[30] Dimitris Kalatzis, Johan Ziruo Ye, Jesper Wohlert, and Søren Hauberg. Multi-chart flows. arXiv
preprint arXiv:2106.03500, 2021.
[31] Mike Yan Michelis and Quentin Becker. On linear interpolation in the latent space of deep
generative models. arXiv preprint arXiv:2105.03663, 2021.
[32] Luis A Pérez Rey, Vlado Menkovski, and Jacobus W Portegies. Diffusion variational autoen-
coders. arXiv preprint arXiv:1901.08991, 2019.
[33] Nutan Chen, Alexej Klushyn, Richard Kurle, Xueyan Jiang, Justin Bayer, and Patrick Smagt.
Metrics for deep generative models. In International Conference on Artificial Intelligence and
Statistics, pages 1540–1550. PMLR, 2018.
[34] Michael M Bronstein, Joan Bruna, Yann LeCun, Arthur Szlam, and Pierre Vandergheynst.
Geometric deep learning: going beyond euclidean data. IEEE Signal Processing Magazine, 34
(4):18–42, 2017.
[35] Federico Monti, Davide Boscaini, Jonathan Masci, Emanuele Rodola, Jan Svoboda, and
Michael M Bronstein. Geometric deep learning on graphs and manifolds using mixture model
cnns. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages
5115–5124, 2017.
[36] Christos Louizos, Kevin Swersky, Yujia Li, Max Welling, and Richard Zemel. The variational
fair autoencoder. arXiv preprint arXiv:1511.00830, 2015.
[37] Francesco Locatello, Gabriele Abbati, Thomas Rainforth, Stefan Bauer, Bernhard Schölkopf,
and Olivier Bachem. On the fairness of disentangled representations. In Advances in Neural
Information Processing Systems, pages 14584–14597, 2019.
[38] Justin Ker, Lipo Wang, Jai Rao, and Tchoyoson Lim. Deep learning applications in medical
image analysis. Ieee Access, 6:9375–9389, 2017.

11
[39] Xiaoran Chen, Nick Pawlowski, Martin Rajchl, Ben Glocker, and Ender Konukoglu. Deep
generative models in the real-world: An open challenge from medical imaging. arXiv preprint
arXiv:1806.05452, 2018.
[40] Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation
and approximate inference in deep generative models. arXiv preprint arXiv:1401.4082, 2014.
[41] Sjoerd van Steenkiste, Francesco Locatello, Jürgen Schmidhuber, and Olivier Bachem. Are
disentangled representations helpful for abstract visual reasoning? In Advances in Neural
Information Processing Systems, 2019.
[42] Cian Eastwood and Christopher KI Williams. A framework for the quantitative evaluation of
disentangled representations. 2018.
[43] Rui Shu, Yining Chen, Abhishek Kumar, Stefano Ermon, and Ben Poole. Weakly supervised
disentanglement with guarantees. arXiv preprint arXiv:1910.09772, 2019.
[44] William F Whitney, Min Jae Song, David Brandfonbrener, Jaan Altosaar, and Kyunghyun Cho.
Evaluating representations by the complexity of learning low-loss predictors. arXiv preprint
arXiv:2009.07368, 2020.
[45] Jakub Tomczak and Max Welling. Vae with a vampprior. In International Conference on
Artificial Intelligence and Statistics, pages 1214–1223. PMLR, 2018.
[46] Abhishek Sinha, Jiaming Song, Chenlin Meng, and Stefano Ermon. D2c: Diffusion-decoding
models for few-shot conditional generation. Advances in Neural Information Processing
Systems, 34, 2021.
[47] Tom White. Sampling generative networks. arXiv preprint arXiv:1609.04468, 2016.
[48] Shujian Yu and Jose C Principe. Understanding autoencoders with information theoretic
concepts. Neural Networks, 117:104–123, 2019.
[49] Shengjia Zhao, Jiaming Song, and Stefano Ermon. InfoVAE: Information maximizing varia-
tional autoencoders. arXiv preprint arXiv:1706.02262, 2017.
[50] James Lucas, George Tucker, Roger Grosse, and Mohammad Norouzi. Understanding posterior
collapse in generative latent variable models. 2019.
[51] Danilo Jimenez Rezende and Fabio Viola. Taming VAEs. arXiv preprint arXiv:1810.00597,
2018.
[52] Taylan Cemgil, Sumedh Ghaisas, Krishnamurthy Dj Dvijotham, and Pushmeet Kohli. Adversar-
ially robust representations with smooth encoders. In International Conference on Learning
Representations, 2019.
[53] Felix Leeb, Guilia Lanzillotta, Yashas Annadani, Michel Besserve, Stefan Bauer, and Bernhard
Schölkopf. Structure by architecture: Disentangled representations without regularization.
arXiv preprint arXiv:2006.07796, 2020.
[54] A Taylan Cemgil, Sumedh Ghaisas, Krishnamurthy Dvijotham, Sven Gowal, and Pushmeet
Kohli. Autoencoding variational autoencoder. arXiv preprint arXiv:2012.03715, 2020.
[55] Steve Dias Da Cruz, Bertram Taetz, Thomas Stifter, and Didier Stricker. Autoencoder attractors
for uncertainty estimation. arXiv preprint arXiv:2204.00382, 2022.
[56] Zijun Zhang, Ruixiang Zhang, Zongpeng Li, Yoshua Bengio, and Liam Paull. Perceptual
generative autoencoders. In International Conference on Machine Learning, pages 11298–
11306. PMLR, 2020.
[57] Giulia Lanzillotta, Felix Leeb, Stefan Bauer, and Bernhard Schölkopf. On the interventional
consistency of autoencoders. 2021.
[58] Salah Rifai, Pascal Vincent, Xavier Muller, Xavier Glorot, and Yoshua Bengio. Contractive
auto-encoders: Explicit invariance during feature extraction. In Icml, 2011.
[59] Adityanarayanan Radhakrishnan, Karren Yang, Mikhail Belkin, and Caroline Uhler. Memoriza-
tion in overparameterized autoencoders. arXiv preprint arXiv:1810.10333, 2018.
[60] Bin Dai and David Wipf. Diagnosing and enhancing VAE models. arXiv preprint
arXiv:1903.05789, 2019.

12
[61] Jan Stühmer, Richard Turner, and Sebastian Nowozin. Independent subspace analysis for
unsupervised learning of disentangled representations. In International Conference on Artificial
Intelligence and Statistics, pages 1200–1210. PMLR, 2020.
[62] Matthew D Hoffman and Matthew J Johnson. Elbo surgery: yet another way to carve up the
variational evidence lower bound. In Workshop in Advances in Approximate Bayesian Inference,
NIPS, volume 1, 2016.
[63] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Im-
plicit neural representations with periodic activation functions. Advances in Neural Information
Processing Systems, 33:7462–7473, 2020.
[64] Kyle Genova, Forrester Cole, Daniel Vlasic, Aaron Sarna, William T Freeman, and Thomas
Funkhouser. Learning shape templates with structured implicit functions. In Proceedings of the
IEEE/CVF International Conference on Computer Vision, pages 7154–7164, 2019.
[65] Holger Theisel. Vector field curvature and applications. PhD thesis, 1995.
[66] Chris Burgess and Hyunjik Kim. 3d shapes dataset, 2018.
[67] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning
applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.
[68] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for
benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747, 2017.
[69] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint
arXiv:1412.6980, 2014.
[70] Christopher P Burgess, Irina Higgins, Arka Pal, Loic Matthey, Nick Watters, Guillaume Des-
jardins, and Alexander Lerchner. Understanding disentangling in β-VAE. arXiv preprint
arXiv:1804.03599, 2018.
[71] Bo Li, Peng Qi, Bo Liu, Shuai Di, Jingen Liu, Jiquan Pei, Jinfeng Yi, and Bowen Zhou.
Trustworthy ai: From principles to practices. ArXiv, abs/2110.01167, 2021.
[72] Cian Eastwood, Andrei Liviu Nicolicioiu, Julius Von Kügelgen, Armin Kekic, Frederik Träuble,
Andrea Dittadi, and Bernhard Schölkopf. On the dci framework for evaluating disentangled
representations: Extensions and connections to identifiability. In UAI 2022 Workshop on Causal
Representation Learning.
[73] Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and
new perspectives. IEEE transactions on pattern analysis and machine intelligence, 35(8):
1798–1828, 2013.
[74] Georgios Arvanitidis, Soren Hauberg, Philipp Hennig, and Michael Schober. Fast and robust
shortest paths on manifolds learned from data. In The 22nd International Conference on
Artificial Intelligence and Statistics, pages 1506–1515. PMLR, 2019.
[75] Dimitris Kalatzis, David Eklund, Georgios Arvanitidis, and Søren Hauberg. Variational autoen-
coders with riemannian brownian motion priors. arXiv preprint arXiv:2002.05227, 2020.
[76] Tim R Davidson, Luca Falorsi, Nicola De Cao, Thomas Kipf, and Jakub M Tomczak. Hyper-
spherical variational auto-encoders. arXiv preprint arXiv:1804.00891, 2018.
[77] Luca Falorsi, Pim de Haan, Tim R Davidson, Nicola De Cao, Maurice Weiler, Patrick Forré,
and Taco S Cohen. Explorations in homeomorphic variational auto-encoding. arXiv preprint
arXiv:1807.04689, 2018.
[78] Tong Lin and Hongbin Zha. Riemannian manifold learning. IEEE Transactions on Pattern
Analysis and Machine Intelligence, 30(5):796–809, 2008.
[79] Alessandra Tosi, Søren Hauberg, Alfredo Vellido, and Neil D Lawrence. Metrics for probabilistic
geometries. arXiv preprint arXiv:1411.7432, 2014.

13
A Appendix

A.1 Latent Responses

In a similar setup as [54], we can extend the joint distribution p(X, Z) to include the reconstruction
and response as p(X, Z, X̂, Ẑ), where crucial question is how the posterior q(Z | X; φ) relates to the
response q(Ẑ | X̂; φ) where Z and Ẑ are related by (also shown in equation 7):
Z
r(Ẑ | Z; θ, φ) = q(Ẑ | X̂; φ)p(X̂ | Z; θ)dX̂ (12)

Note, that r(Ẑ | Z; θ, φ) is equivalent to the transition kernel QAVAE in [54]. However, crucially,
we do not make two assumptions used to derive the AVAE objective. Firstly, we do not assume
that the decoder is a one-to-one mapping between latent samples and a corresponding generated
sample. The contractive behavior observed in the latent space of autoencoders [59], suggests a
many-to-one mapping is more realistic, which may be interpretted as the decoder filtering out useless
exogenous information from the latent code. Consequently, we also do not treat p(Ẑ; θ, φ) =
Ep(Z) [r(Ẑ | Z; θ, φ)] as a normal distribution, which would imply the encoder perfectly inverts the
decoder.
Consider the reconstructions X̂ of the maximally overfit encoder q(z = Z | xi = X; φ̃) = δ(si − z)
(recall si = f φ (xi )) and decoder p(X̂ | Z; θ̃). Since the autoencoder is trained on the empirical
generative process π(X) rather
R than the true generative process p(X), the overfit decoder generates
samples from p(X̂; θ̃) = p(X̂ | Z; θ̃)p(Z)dZ = π(X), which does not have continuous support.
For such a decoder, all exogenous noise is completely removed and the decoder mapping is obviously
many-to-one, and it follows that r(Ẑ = ẑ | Z = z; θ̃, φ̃) = δ(ẑ − s) (recall z = s + u).
Now consider the more desirable (and perhaps slightly more realistic) setting where the autoencoder
extrapolates somewhat beyond π(X) to resemble p(X), in which case decoding the latent sample z ∼
q(Z | X = x; φ) to generate x̂ ∼ p(x̂ = X̂ | z = Z; θ) will not necessarily match the observation x,
which, by our definition of endogenous information, implies a change in the endogenous information
contained in z. When re-encoding to get q(Ẑ | X̂ = x̂; φ), the changes in the endogenous information
result in some width to the distribution over Ẑ.

A.2 Derivation of Equation 6

Starting from our definition of ŝ = f φ (x̂) where x̂ = g θ (z), z = f φ (x), z = s + u, and  = x̂ − x.

The high-level goal is expand f φ around x and then g θ around s to first order.

ŝ = f φ (x̂)

ŝ = f φ (x + )

ŝ = f φ (x) + Jf φ (x) + O 2
 

ŝ = f φ (x) + Jf φ (x) g θ (z) − x + O 2

   

ŝ = f φ (x) + Jf φ (x) g θ (s + u) − x + O 2
   

ŝ = f φ (x) + Jf φ (x) g θ (s) + Jgθ (s)u − x + O 2 + O u2

     

ŝ = f φ (x) + Jf φ (x)(g θ (s) − x) + Jf φ (x)Jgθ (s)u + O 2 + O u2

   

14
A.3 Comparing the Conditioned Response Matrix and the DCI Responsibility Matrix

In [42], the responsibility matrix is used to evaluate the disentanglement of a learned representation.
In the matrix, element Rij corresponds to the relative importance of latent variable j in predicting
the true factor of variation i for a simple classifier trained with full supervision to recover the true
factors from the latent vector. Although the scalar scores (DCI-d and CDS) are computed identically
from the respective matrices, there are important practical and theoretical distinctions in the DCI and
latent response frameworks.
First and most importantly, since the DCI framework only uses the encoder, the learned generative
process is not taken into account at all. Consequently, the DCI framework (and other existing
disentanglement metrics) fail to evaluate how disentangled the causal drivers of the learned generative
process are, and instead evaluate which latent variables are correlated with true factors. Furthermore,
practically speaking, the DCI framework is sensitive to a variety of hyperparameters such as the exact
design and training of the model [72], while the conditioned response matrix has far fewer (and more
intuitive) hyperparameters relating to the Monte carlo integration.
Interestingly, the DCI responsibility matrices do often resemble the conditioned response matrices,
suggesting that relying on correlations instead of a full causal analysis, can yield similar results.
Obviously as the data becomes more challenging and realistic, and the true generative process involves
a more complicated causal structure, then we may expect the DCI responsibility matrix to become
less reliable for analyzing the generative model structure. In fact, then the learned causal structure
estimated using the latent response matrix may be used in tandem to develop a structure-aware
disentanglement metric.

A.4 Mean Curvature for Manifold Learning

The geometry of learned representations with a focus on the generalization ability of neural networks
has been discussed in [73]. One key problem is that the standard Gaussian prior used in variational
autoencoders relies on the usual Lebesgue measure which in turn, assumes a Euclidean structure over
the latent space. This has been demonstrated to lead to difficulties in particular when interpolating
in the latent space [25, 74, 75] due to a manifold mismatch [76, 77]. Given the complexity of
the underlying data manifold, a viable alternative is based on riemanian geometry [78] which has
previously been investigated for alternative probabilistic models like Gaussian Process regression
[79].
These methods focus on the intrinsic curvature of the data manifold, which does not depend on the
specific embedding of the manifold in the latent space. However, our focus is precisely on how the
data manifold is embedded in the latent space, to (among other things) quantify the relationships
between latent variables and how well the representation disentangles the true factors of variation.
Consequently, we focus on the extrinsic curvature, and more specifically the mean curvature which
can readily be estimated using the response maps.
As discussed in the main paper, |u(z)| = |z − s| is interpreted as a distance where |u(z)| = 0 implies
z is on the latent manifold and there is no exogenous noise. The gradient of this function ∇z |u(z)|,
effectively projects any point in the latent space onto the endogenous manifold. Similarly, the mean
curvature (equation 13) can be computed, which can be interpreted as identifying the regions in the
latent space where the |u(z)| converges and diverges. These gradients are estimated numerically by
finite differencing.
 
1 ∇z |u(z)| 1 u(z)
H = − ∇z · = − ∇z · (13)
2 |∇z |u(z)|| 2 |u(z)|

A.5 Double Helix Example Details

To illustrate how the latent response framework can be used to study the representation learned by
a VAE, we show the process when learning a 2D representation for samples from a double helix
embedded in R3 , defined as:
T
xi = [A1 cos(π(ωti + ni )), A2 sin(π(ωti + ni )), A3 ti ] + i (14)

15
where ti ∼ Uniform(−1, 1), ni ∼ Bernoulli(0.5), i ∼ N (0, σI). For this experiment, we set
A1 = A2 = A3 = ω = 1 and σ = 0.1.
Disregarding the additive noise i , the data manifold has two degrees of freedom, which are the strand
location ti and the strand number ni .
To provide the model sufficient capacity, we use four hidden layers with 32 units each for the encoder
and decoder. We train until convergence (at most 5k steps) with β = 0.05 using an Adam optimizer on
a total of N = 1024 training samples (see the supplementary code for the full training and evaluation
details).

2.0

1.5

1.0

0.5

0.0

0.5

1.0

1.5

2.0
2.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5

Figure 7: The response map of the representation trained on the double helix. Starting from the latent
samples (blue dots), applying the decoder followed by the encoder (i.e. response function) results in
the orange dots connected by the black arrows. Note that applying the response function effectively
contracts points all over the latent space into a relatively small non-linear region, corresponding to
endogenous information.

A.6 Architecture and Training Details

All our models are based on the same convolutional neural network architecture detailed in table 10 so
that in total models have approximately 500k trainable parameters. For the smaller datasets MNIST
and Fashion-MNIST, samples are upsampled to 32x32 pixels from their original 28x28 and the one
convolutional block is removed from both the encoder and decoder.
The datasets are split into a 70-10-20 train-val-test split, and are optimized using Adam [69] with a
learning rate of 0.0001, weight decay 0, and β1 , β2 of 0.9 and 0.999 respectively. The models are
trained for 100k iterations with a batch size of 64 (128 for MNIST and Fashion-MNIST).

16
 Input 64x64x3 image
Input d latent vector
Conv Layer (64 filters, k=5x5, s=1x1, p=2x2)
Fully-connected Layer (128 units)
Max pooling (filter 2x2, s=2x2)
ELU activation
Group Normalization (8 groups, affine)
Fully-connected Layer (256 units)
ELU activation
ELU activation
Conv Layer (64 filters, k=3x3, s=1x1, p=1x1)
Fully-connected Layer (256 units)
Max pooling (filter 2x2, s=2x2)
ELU activation
Group Normalization (8 groups, affine)
Bilinear upsampling (scale 2x2)
ELU activation
Conv Layer (64 filters, k=3x3, s=1x1, p=1x1)
Conv Layer (64 filters, k=3x3, s=1x1, p=1x1)
Group Normalization (8 groups, affine)
Max pooling (filter 2x2, s=2x2)
ELU activation
Group Normalization (8 groups, affine)
Bilinear upsampling (scale 2x2)
ELU activation
Conv Layer (64 filters, k=3x3, s=1x1, p=1x1)
Conv Layer (64 filters, k=3x3, s=1x1, p=1x1)
Group Normalization (8 groups, affine)
Max pooling (filter 2x2, s=2x2)
ELU activation
Group Normalization (8 groups, affine)
Bilinear upsampling (scale 2x2)
ELU activation
Conv Layer (64 filters, k=3x3, s=1x1, p=1x1)
Conv Layer (64 filters, k=3x3, s=1x1, p=1x1)
Group Normalization (8 groups, affine)
Max pooling (filter 2x2, s=2x2)
ELU activation
Group Normalization (8 groups, affine)
Bilinear upsampling (scale 2x2)
ELU activation
Conv Layer (64 filters, k=3x3, s=1x1, p=1x1)
Fully-connected Layer (256 units)
Group Normalization (8 groups, affine)
ELU activation
ELU activation
Fully-connected Layer (128 units)
Conv Layer (3 filters, k=3x3, s=1x1, p=1x1)
ELU activation
Sigmoid activation
Fully-connected Layer (2d units)
Output 64x64x3 image
Output posterior µ and log σ
Figure 9: Decoder Architecture
Figure 8: Encoder Architecture
Figure 10: Model architectures where "k" is the kernel size, "s" is the stride, and "p" is the zero-
padding

17
 B Additional Results

B.1 3D-Shapes

1-VAE 2-VAE 4-VAE 8-VAE

0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10
0 0.6 0.1 0.2 0.1 0.2 0.1 0.0 0.1 0.1 0.1 0.2 0.1 0.7 0.1 0.2 0.1 0.2 0.1 0.0 0.1 0.0 0.1 0.2 0.0 1.0 0.0 0.0 0.1 0.1 0.1 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.1 0.1 0.0 0.0 0.1 0.0 0.0 0.0 0.0
0.2 0.9 0.2 0.1 0.2 0.1 0.1 0.2 0.1 0.1 0.2 0.1 0.2 0.9 0.1 0.1 0.2 0.1 0.1 0.2 0.1 0.1 0.2 0.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.7 0.0 0.0 0.1 0.1 0.0 0.0 0.0 0.0 0.0 0.0
2 0.1 0.1 0.7 0.1 0.2 0.1 0.0 0.1 0.0 0.0 0.1 0.1 0.2 0.0 0.7 0.1 0.1 0.1 0.0 0.0 0.0 0.0 0.2 0.1 0.0 0.0 0.8 0.0 0.0 0.1 0.0 0.1 0.0 0.0 0.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.2 0.1 0.2 0.8 0.2 0.1 0.1 0.1 0.1 0.1 0.2 0.1 0.1 0.1 0.2 0.9 0.2 0.1 0.0 0.1 0.1 0.0 0.1 0.1 0.2 0.0 0.0 1.0 0.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.1 0.0 0.0 0.9 0.1 0.0 0.0 0.1 0.0 0.0 0.0 0.0
4 0.1 0.0 0.1 0.0 0.5 0.0 0.0 0.0 0.0 0.0 0.1 0.0 0.0 0.0 0.0 0.0 0.5 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.6 0.0 0.0 0.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.7 0.0 0.0 0.1 0.0 0.0 0.0 0.0
Intervention

0.2 0.1 0.3 0.2 0.2 0.9 0.1 0.2 0.1 0.1 0.2 0.1 0.2 0.1 0.2 0.2 0.3 0.8 0.1 0.1 0.1 0.1 0.2 0.1 0.1 0.0 0.2 0.0 0.2 0.8 0.0 0.2 0.0 0.0 0.1 0.0 0.0 0.2 0.0 0.0 0.1 0.9 0.0 0.1 0.0 0.0 0.0 0.0
6 0.2 0.1 0.2 0.1 0.2 0.1 0.9 0.1 0.1 0.1 0.1 0.1 0.2 0.1 0.1 0.1 0.2 0.1 0.9 0.1 0.1 0.1 0.1 0.1 0.0 0.0 0.0 0.0 0.1 0.0 1.0 0.1 0.2 0.0 0.1 0.0 0.0 0.0 0.0 0.0 0.1 0.0 0.9 0.1 0.1 0.0 0.1 0.0
0.2 0.2 0.2 0.1 0.3 0.1 0.1 0.8 0.1 0.1 0.2 0.1 0.2 0.2 0.2 0.1 0.3 0.1 0.1 0.9 0.1 0.1 0.2 0.1 0.0 0.0 0.2 0.0 0.4 0.2 0.0 0.9 0.0 0.0 0.1 0.0 0.0 0.1 0.0 0.0 0.4 0.0 0.0 0.9 0.0 0.0 0.2 0.0
8 0.2 0.1 0.2 0.2 0.3 0.2 0.2 0.2 0.9 0.1 0.2 0.1 0.2 0.1 0.1 0.1 0.3 0.1 0.2 0.1 0.9 0.1 0.1 0.1 0.0 0.0 0.0 0.0 0.1 0.1 0.2 0.1 0.9 0.0 0.1 0.0 0.0 0.0 0.0 0.0 0.1 0.0 0.2 0.1 1.0 0.0 0.1 0.0
0.3 0.1 0.3 0.2 0.3 0.2 0.2 0.1 0.2 1.2 0.3 0.1 0.3 0.1 0.3 0.1 0.2 0.1 0.2 0.1 0.2 1.0 0.3 0.1 0.0 0.0 0.1 0.0 0.1 0.1 0.0 0.1 0.0 1.0 0.0 0.2 0.0 0.1 0.0 0.0 0.1 0.0 0.0 0.1 0.0 1.1 0.0 0.2
10 0.1 0.0 0.1 0.1 0.1 0.1 0.0 0.0 0.0 0.0 0.6 0.1 0.1 0.0 0.2 0.0 0.1 0.1 0.0 0.0 0.0 0.0 0.6 0.0 0.1 0.0 0.2 0.1 0.2 0.2 0.0 0.1 0.0 0.1 0.9 0.0 0.0 0.0 0.0 0.1 0.3 0.0 0.0 0.2 0.0 0.0 0.9 0.0
0.2 0.1 0.3 0.2 0.3 0.2 0.1 0.1 0.2 0.1 0.3 1.1 0.3 0.1 0.4 0.1 0.3 0.1 0.1 0.1 0.1 0.1 0.3 1.0 0.0 0.0 0.1 0.0 0.1 0.1 0.0 0.1 0.0 0.2 0.1 1.0 0.0 0.1 0.0 0.0 0.1 0.0 0.0 0.0 0.0 0.1 0.0 1.0
Response Response Response Response

(a) Latent Response Matrices

1-VAE 2-VAE 4-VAE 8-VAE
0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10
floor_hue 0.1 0.1 0.0 0.2 0.0 0.2 1.0 0.2 0.6 0.4 0.0 0.2 0.0 0.1 0.0 0.1 0.0 0.1 0.9 0.1 0.8 0.3 0.0 0.1 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 1.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 1.0 0.0 0.0 0.0
wall_hue 0.3 0.3 0.5 0.2 0.1 0.5 0.2 0.2 0.3 0.9 0.5 0.8 0.3 0.2 0.6 0.1 0.0 0.3 0.2 0.2 0.3 1.1 0.4 0.9 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.1 0.0 1.0
object_hue 0.2 0.4 0.2 0.6 0.1 0.3 0.2 0.5 0.3 0.0 0.0 0.1 0.1 0.5 0.1 0.9 0.0 0.3 0.0 0.5 0.1 0.0 0.0 0.1 1.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
scale 0.1 0.5 0.1 0.1 0.1 0.3 0.1 0.2 0.1 0.1 0.1 0.4 0.1 0.5 0.0 0.1 0.0 0.4 0.0 0.2 0.1 0.1 0.1 0.4 0.0 0.0 0.0 0.0 0.0 0.3 0.0 0.5 0.0 0.0 0.8 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.3 0.0 0.0 0.9 0.0
shape 0.3 0.4 0.3 0.4 0.3 0.3 0.1 0.5 0.2 0.1 0.3 0.1 0.3 0.4 0.1 0.2 0.4 0.4 0.0 0.6 0.1 0.1 0.3 0.1 0.0 0.0 0.1 0.0 0.9 0.5 0.0 0.9 0.0 0.0 0.2 0.0 0.0 0.0 0.0 0.0 0.7 0.0 0.0 0.9 0.0 0.0 0.2 0.0
orientation 0.4 0.1 0.5 0.2 0.3 0.2 0.1 0.1 0.1 0.1 0.5 0.1 0.4 0.1 0.7 0.1 0.2 0.2 0.0 0.1 0.1 0.2 0.4 0.2 0.1 0.0 0.8 0.0 0.0 0.7 0.0 0.2 0.0 0.0 0.4 0.0 0.0 0.8 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0
Latent Dimension Latent Dimension Latent Dimension Latent Dimension

(b) Conditioned Response Matrices

1-VAE 2-VAE 4-VAE 8-VAE
0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10
floor_hue 0.0 0.0 0.0 0.0 0.0 0.0 0.5 0.0 0.2 0.2 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.5 0.0 0.3 0.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.4 0.0 0.6 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.5 0.0 0.5 0.0 0.0 0.0
wall_hue 0.0 0.0 0.0 0.0 0.0 0.1 0.0 0.0 0.1 0.4 0.0 0.3 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.4 0.0 0.3 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.5 0.0 0.5 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.3 0.0 0.7
object_hue 0.0 0.2 0.0 0.2 0.0 0.1 0.0 0.2 0.1 0.0 0.0 0.0 0.0 0.2 0.0 0.4 0.0 0.1 0.0 0.2 0.0 0.0 0.0 0.0 0.4 0.0 0.0 0.6 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.4 0.0 0.0 0.6 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
scale 0.0 0.3 0.0 0.0 0.0 0.2 0.0 0.0 0.0 0.0 0.0 0.3 0.0 0.3 0.0 0.0 0.0 0.2 0.0 0.1 0.0 0.0 0.0 0.2 0.0 0.0 0.0 0.0 0.0 0.2 0.0 0.3 0.0 0.0 0.5 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.3 0.0 0.0 0.7 0.0
shape 0.1 0.1 0.1 0.1 0.3 0.1 0.0 0.2 0.0 0.0 0.1 0.0 0.2 0.0 0.0 0.0 0.3 0.1 0.0 0.2 0.0 0.0 0.1 0.0 0.0 0.0 0.0 0.0 0.6 0.1 0.0 0.2 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.5 0.0 0.0 0.4 0.0 0.0 0.1 0.0
orientation 0.2 0.0 0.1 0.1 0.1 0.1 0.0 0.0 0.0 0.1 0.2 0.1 0.2 0.0 0.2 0.0 0.2 0.0 0.0 0.0 0.0 0.1 0.2 0.1 0.0 0.0 0.6 0.0 0.0 0.2 0.0 0.0 0.0 0.0 0.1 0.0 0.0 0.4 0.0 0.0 0.0 0.6 0.0 0.0 0.0 0.0 0.0 0.0
Latent Dimension Latent Dimension Latent Dimension Latent Dimension

(c) DCI Responsibility Matrices

Figure 11: Response and Responsibility matrices for several VAEs (d = 12).

18
 0 5 10 15 20
0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.1 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.9 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.5 0.0 0.0
5 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.1 0.0 0.1 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.1 0.0 0.0
10 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Intervention

0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.2 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
15 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.1 1.0 0.0 0.0 0.0 0.0 0.0 0.1 0.0 0.2
0.0 0.0 0.0 0.0 0.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.1 0.1 0.2 0.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
20 0.0 0.0 0.0 0.0 0.2 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.1 0.0 0.0
0.0 0.0 0.0 0.0 0.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.7 0.0 0.0
0.0 0.0 0.0 0.0 0.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.2 0.0 0.0 0.0 0.1 0.1 1.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.1 0.2 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0
Response

(a) Latent Response Matrix

0 5 10 15 20
floor_hue 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 1.1 0.0
wall_hue 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.1
object_hue 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
scale 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0
shape 0.0 0.0 0.0 0.0 1.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.2 0.8 0.0 0.0
orientation 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Latent Dimension

(b) Conditioned Response Matrix

0 5 10 15 20
floor_hue 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.4 0.0 0.0 0.0 0.0 0.0 0.6 0.0
wall_hue 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.5 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.5
object_hue 0.0 0.0 0.5 0.0 0.0 0.0 0.0 0.0 0.0 0.5 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
scale 0.0 0.0 0.0 0.0 0.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.9 0.0 0.0 0.0
shape 0.0 0.0 0.0 0.0 0.5 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.5 0.0 0.0
orientation 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.4 0.6 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Latent Dimension

(c) DCI Responsibility Matrix

Figure 12: Full response and responsibility matrices of the 4-VAE (d = 24) also shown in figure 6.
Note how the Latent Response matrices (12a) shows a categorical difference between the latent
dimensions where the diagonal element is close to zero (non-causal), compared to the dimensions
with diagonal elements close to 1 (causal).

19
 2.0 0.53
2.0
1.5 0.40
1.5 0.75
1.0 0.26
1.0 0.50
0.5 0.13
0.5
Dimension 16

0.25

Dimension 16
0.0 0.00
0.0 0.00
0.5 -0.02
0.5
0.25
1.0 -0.04
1.0
0.50
1.5 -0.06
1.5
2.0 -0.08
0.75
2 1 0 1 2 2.0
Dimension 22 2 1 0 1 2
Dimension 22
(c) Corresponding Reconstruc-
(a) Divergence Map and Aggregate Pos- tions
terior (b) Mean Curvature Map

Figure 13: Visualization of the representation learned by a 4-VAE trained on 3D-Shapes (same
model as in figure 12).

2.0 0.58
2.0
1.5 0.43
1.5 0.75
1.0 0.29
1.0 0.50
0.5 0.14
0.5
Dimension 23

0.25
Dimension 23

0.0 0.00
0.0
0.00
0.5 -0.02
0.5
0.25
1.0 -0.05
1.0
1.5 -0.07 0.50
1.5
2.0 -0.10 0.75
2 1 0 1 2 2.0
Dimension 15 2 1 0 1 2
Dimension 15
(c) Corresponding Reconstruc-
(a) Divergence Map and Aggregate Pos- tions
terior (b) Mean Curvature Map

Figure 14: Visualization of the representation learned by a 4-VAE trained on 3D-Shapes (same
model as in figure 12).

2.0 0.49
2.0
1.5 0.37
1.5 0.75
1.0 0.25
1.0 0.50
0.5 0.12
0.5
Dimension 9

0.25
Dimension 9

0.0 0.00
0.0
0.00
0.5 -0.02
0.5
1.0 -0.04 0.25
1.0
1.5 -0.06 0.50
1.5
2.0 -0.08 0.75
2 1 0 1 2 2.0
Dimension 2 2 1 0 1 2
Dimension 2
(c) Corresponding Reconstruc-
(a) Divergence Map and Aggregate Pos- tions
terior (b) Mean Curvature Map

Figure 15: Visualization of the representation learned by a 4-VAE trained on 3D-Shapes (same
model as in figure 12).

20
 2.0 0.42
2.0
1.5 0.32 0.8
1.5
1.0 0.21 0.6
1.0
0.5 0.11 0.4
0.5
Dimension 4

Dimension 4
0.0 0.00 0.2
0.0
0.5 -0.02 0.0
0.5
1.0 -0.05 0.2
1.0
1.5 -0.07 0.4
1.5
2.0 -0.09 0.6
2 1 0 1 2 2.0
Dimension 21 2 1 0 1 2
Dimension 21
(c) Corresponding Reconstruc-
(a) Divergence Map and Aggregate Pos- tions
terior (b) Mean Curvature Map

Figure 16: Visualization of the representation learned by a 4-VAE trained on 3D-Shapes (same
model as in figure 12). This projection is particularly interesting as the information encoding shape
is not exactly axis-aligned, leading to a slight mismatch between the aggregate posterior and the
divergence maps. As our visualizations are presently confined to two dimensions, the structure can
become significantly more obscured to us if the information is not disentangled and axis-aligned.

2.0 0.24
2.0
1.5 0.18 0.8
1.5
1.0 0.12 0.6
1.0
0.5 0.06 0.4
0.5
Dimension 14

Dimension 14

0.0 0.00 0.2

0.0
0.0
0.5 -0.02
0.5
0.2
1.0 -0.04
1.0
0.4
1.5 -0.07
1.5 0.6
2.0 -0.09
2 1 0 1 2 2.0
Dimension 13 2 1 0 1 2
Dimension 13
(c) Corresponding Reconstruc-
(a) Divergence Map and Aggregate Pos- tions
terior (b) Mean Curvature Map

Figure 17: Visualization of the representation learned by a 4-VAE trained on 3D-Shapes (same
model as in figure 12).

21
B.2 MPI3D

1-VAE 2-VAE 4-VAE 8-VAE

0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10
0 1.1 0.0 0.0 0.0 0.0 0.1 0.1 0.0 0.0 0.0 0.0 0.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
2 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.1 0.0 0.0 0.0 0.0 0.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
4 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Intervention

Intervention

Intervention

Intervention

0.0 0.0 0.0 0.0 0.0 1.0 0.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
6 0.1 0.0 0.0 0.0 0.0 0.2 1.1 0.0 0.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.2 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
8 0.0 0.0 0.0 0.0 0.0 0.1 0.1 0.0 1.0 0.0 0.0 0.0 0.1 0.1 0.0 0.1 0.0 0.0 0.1 0.1 1.2 0.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.2 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
10 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.1 0.1 0.0 0.0 0.0 0.0 1.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Response Response Response Response

(a) Latent Response Matrices

1-VAE 2-VAE 4-VAE 8-VAE
0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10
object_color 0.3 0.0 0.0 0.0 0.0 0.9 0.7 0.0 0.3 0.0 0.0 0.0 0.0 0.0 0.0 0.1 0.0 0.0 0.0 0.0 0.3 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.6 0.0 0.3 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.2 0.0 0.0 0.0
object_shape 0.1 0.0 0.0 0.0 0.0 0.2 0.4 0.0 0.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
object_size 0.1 0.0 0.0 0.0 0.0 0.1 0.3 0.0 0.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
camera_height 0.2 0.0 0.0 0.0 0.0 0.1 0.3 0.0 0.2 0.0 0.0 1.3 0.0 0.0 0.0 0.2 0.0 0.0 0.0 0.0 1.6 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.2 0.0 1.2 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.2 0.0 0.0 0.0
background_color 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.4 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.3 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.3 0.0 0.3 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.5 0.0 0.0 0.0
horizonal_axis 0.9 0.0 0.0 0.0 0.0 0.4 0.6 0.0 0.9 0.0 0.0 0.0 0.0 0.0 0.0 1.1 0.0 0.0 0.0 0.0 0.2 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.5 0.0 0.5 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.1 0.0 0.0 0.0
vertical_axis 0.9 0.0 0.0 0.0 0.0 0.2 0.5 0.0 0.4 0.0 0.0 0.0 0.0 0.0 0.0 0.9 0.0 0.0 0.0 0.0 0.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.2 0.0 0.3 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.1 0.0 0.0 0.0
Latent Dimension Latent Dimension Latent Dimension Latent Dimension

(b) Conditioned Response Matrices

1-VAE 2-VAE 4-VAE 8-VAE
0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10
object_color 0.1 0.0 0.0 0.0 0.0 0.5 0.2 0.0 0.1 0.1 0.0 0.0 0.1 0.1 0.0 0.1 0.0 0.1 0.0 0.1 0.3 0.1 0.0 0.1 0.0 0.0 0.1 0.0 0.0 0.0 0.3 0.0 0.2 0.1 0.0 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.0 0.1 0.2 0.1 0.1 0.1
object_shape 0.1 0.1 0.1 0.1 0.0 0.1 0.2 0.0 0.1 0.1 0.1 0.0 0.1 0.1 0.0 0.1 0.1 0.0 0.1 0.1 0.2 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1
object_size 0.1 0.1 0.1 0.2 0.1 0.1 0.2 0.1 0.0 0.1 0.1 0.0 0.0 0.2 0.0 0.0 0.0 0.0 0.2 0.1 0.2 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.0 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.0 0.0 0.1
camera_height 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.3 0.0 0.6 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.2 0.0 0.0 0.7 0.0 0.0 0.0
background_color 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.7 0.0 0.0 0.0 0.0 0.1 0.0 0.0 0.0 0.0 0.2 0.1 0.4 0.0 0.1 0.0 0.0 0.0 0.0 0.1 0.1 0.4 0.0 0.0 0.3 0.0 0.0 0.1
horizonal_axis 0.1 0.1 0.1 0.0 0.1 0.1 0.1 0.0 0.3 0.1 0.0 0.0 0.1 0.0 0.1 0.2 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.2 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1
vertical_axis 0.2 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.0 0.1 0.1 0.1 0.2 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.2 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1
Latent Dimension Latent Dimension Latent Dimension Latent Dimension

(c) DCI Responsibility Matrices

Figure 18: Response and Responsibility matrices for several VAEs (d = 12) trained on the MPI3D
Toy dataset.

22
 1-VAE 2-VAE 4-VAE 8-VAE
0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10
0 1.3 0.2 0.0 0.0 0.1 0.0 0.2 0.0 0.0 0.1 0.1 0.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.2 1.1 0.0 0.0 0.2 0.0 0.2 0.0 0.0 0.1 0.1 0.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
2 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Intervention 4 0.0 0.1 0.0 0.0 0.8 0.0 0.0 0.0 0.0 0.0 0.0 0.2 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

Intervention

Intervention

Intervention
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
6 0.2 0.3 0.0 0.0 0.1 0.0 1.3 0.0 0.0 0.2 0.1 0.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
8 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.6 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.9 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.2 0.3 0.0 0.0 0.1 0.0 0.2 0.0 0.0 1.1 0.1 0.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.1 1.0 0.0 0.0 0.1 0.0 0.0 0.0 0.1 0.0 0.0 0.0 0.1 1.2 0.0 0.0 0.1 0.1 0.0 0.0 0.1 0.0 0.1 0.1 0.1 0.9 0.0 0.4
10 0.2 0.3 0.0 0.0 0.1 0.0 0.2 0.0 0.0 0.2 1.3 0.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.1 0.1 1.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.1 0.1 0.0 0.0 0.4 0.0 0.1 0.0 0.0 0.0 0.0 1.4 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.1 0.1 0.0 1.2 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.1 0.0 0.1 0.0 0.0 0.0 0.1 0.2 0.0 1.1
Response Response Response Response

(a) Latent Response Matrices

1-VAE 2-VAE 4-VAE 8-VAE
0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10
object_color 0.7 0.9 0.0 0.0 0.1 0.0 0.5 0.0 0.0 0.4 0.4 0.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.6 0.1 0.4 0.5 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.5 0.4 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.3 0.0 0.6
object_shape 0.2 0.3 0.0 0.0 0.1 0.0 0.5 0.0 0.0 0.3 0.3 0.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.3 0.0 0.3 0.2 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.2 0.2 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.2 0.0 0.3
object_size 0.3 0.3 0.0 0.0 0.1 0.0 0.3 0.0 0.0 0.4 0.4 0.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.4 0.0 0.3 0.3 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.1 0.2 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.2 0.0 0.4
camera_height 0.4 0.6 0.0 0.0 0.9 0.0 0.6 0.0 0.0 0.5 0.4 1.7 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.4 1.0 0.2 0.3 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.2 1.2 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.7
background_color 0.1 0.1 0.0 0.0 0.3 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.1 0.5 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.1 0.0 0.2
horizonal_axis 0.9 0.5 0.0 0.0 0.2 0.0 0.8 0.0 0.0 1.1 1.3 0.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.4 0.1 0.8 0.9 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.8 0.4 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.3 0.0 0.7
vertical_axis 0.8 0.5 0.0 0.0 0.1 0.0 0.9 0.0 0.0 1.0 0.9 0.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.3 0.1 0.6 0.5 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.6 0.2 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.2 0.0 0.6
Latent Dimension Latent Dimension Latent Dimension Latent Dimension

(b) Conditioned Response Matrices

1-VAE 2-VAE 4-VAE 8-VAE
0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10
object_color 0.3 0.3 0.0 0.0 0.0 0.0 0.1 0.0 0.0 0.1 0.2 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.1 0.0 0.6 0.0 0.0 0.0 0.1 0.0 0.0 0.1 0.1 0.1 0.0 0.0 0.1 0.1 0.1 0.2 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1
object_shape 0.1 0.1 0.0 0.0 0.1 0.1 0.2 0.1 0.0 0.1 0.2 0.0 0.1 0.1 0.1 0.0 0.1 0.0 0.1 0.1 0.2 0.1 0.2 0.1 0.1 0.1 0.0 0.1 0.2 0.1 0.1 0.1 0.1 0.0 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1
object_size 0.1 0.2 0.0 0.0 0.0 0.0 0.2 0.0 0.0 0.1 0.2 0.0 0.0 0.1 0.1 0.0 0.0 0.0 0.0 0.0 0.4 0.0 0.2 0.1 0.0 0.1 0.0 0.0 0.3 0.1 0.1 0.0 0.1 0.1 0.0 0.0 0.1 0.1 0.2 0.1 0.1 0.1 0.1 0.0 0.1 0.1 0.1 0.0
camera_height 0.0 0.0 0.0 0.0 0.6 0.0 0.0 0.0 0.0 0.0 0.0 0.4 0.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.8 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.5 0.2 0.1 0.0 0.0 0.1 0.0 0.0 0.0 0.0 0.0 0.1 0.6 0.0 0.0
background_color 0.0 0.0 0.1 0.0 0.3 0.0 0.0 0.0 0.0 0.0 0.0 0.6 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.9 0.0 0.0 0.1 0.1 0.1 0.0 0.1 0.1 0.1 0.1 0.0 0.1 0.1 0.1 0.1 0.1 0.1 0.2 0.1 0.1 0.1 0.1 0.0 0.2 0.1 0.1
horizonal_axis 0.1 0.1 0.0 0.0 0.1 0.0 0.1 0.0 0.1 0.2 0.2 0.0 0.0 0.1 0.0 0.1 0.0 0.0 0.0 0.1 0.1 0.1 0.2 0.3 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1
vertical_axis 0.1 0.1 0.1 0.0 0.1 0.1 0.2 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.2 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1
Latent Dimension Latent Dimension Latent Dimension Latent Dimension

(c) DCI Responsibility Matrices

Figure 19: Response and Responsibility matrices for several VAEs (d = 12) trained on the MPI3D
Real dataset.

Name CDS DCI-D IRS MIG

1-VAE 0.69 0.33 0.58 0.32
2-VAE 0.86 0.17 0.59 0.14
4-VAE 0.66 0.11 0.61 0.05
8-VAE 1 0.13 0.79 0.1
1-VAE 0.61 0.24 0.51 0.07
2-VAE 0.69 0.26 0.72 0.24
4-VAE 0.4 0.09 0.75 0.04
8-VAE 0.7 0.08 0.71 0.04
Table 2: disentanglement scores for the MPI3D Toy (first four rows) and Real (last four rows).

23
 2.0 0.17
2.0
0.8
1.5 0.13
1.5
0.6
1.0 0.08
1.0 0.4
0.5 0.04
0.5 0.2
Dimension 11

Dimension 11
0.0 0.00
0.0 0.0
0.5 -0.01
0.5 0.2
1.0 -0.03 0.4
1.0
1.5 -0.04 0.6
1.5
2.0 -0.05 0.8
2 1 0 1 2 2.0
Dimension 0 2 1 0 1 2
Dimension 0
(c) Corresponding Reconstruc-
(a) Divergence Map and Aggregate Pos- tions
terior (b) Mean Curvature Map

Figure 20: Visualization of the representation learned by the 1-VAE trained on MPI3D Toy.

2.0 0.61
2.0
1.5 0.46 1.00
1.5
1.0 0.30 0.75
1.0
0.5 0.15 0.50
0.5
Dimension 5

Dimension 5

0.0 0.00 0.25

0.0
0.5 -0.08 0.00
0.5
1.0 -0.16 0.25
1.0
1.5 -0.24 0.50
1.5
2.0 -0.32 0.75
2 1 0 1 2 2.0
Dimension 6 2 1 0 1 2
Dimension 6
(c) Corresponding Reconstruc-
(a) Divergence Map and Aggregate Pos- tions
terior (b) Mean Curvature Map

Figure 21: Visualization of the representation learned by the 1-VAE trained on MPI3D Toy.

2.0 0.31
2.0
1.5 0.24
1.5
1.0 0.16 0.5
1.0
0.5 0.08
0.5
Dimension 6

0.0
Dimension 6

0.0 0.00
0.0
0.5 -0.26
0.5 0.5
1.0 -0.53
1.0
1.5 -0.79
1.5 1.0
2.0 -1.05
2 1 0 1 2 2.0
Dimension 8 2 1 0 1 2
Dimension 8
(c) Corresponding Reconstruc-
(a) Divergence Map and Aggregate Pos- tions
terior (b) Mean Curvature Map

Figure 22: Visualization of the representation learned by the 4-VAE trained on MPI3D Toy. Note that
due to posterior collapse, the full latent manifold is contained in this projection (see the corresponding
response matrix in figure 18).

24
 2.0 0.63
2.0
1.5 0.47
1.5 0.75
1.0 0.31
1.0 0.50
0.5 0.16
0.5 0.25
Dimension 4

Dimension 4
0.0 0.00
0.0 0.00
0.5 -0.05
0.5
0.25
1.0 -0.10
1.0
0.50
1.5 -0.16
1.5
0.75
2.0 -0.21
2 1 0 1 2 2.0
Dimension 11 2 1 0 1 2
Dimension 11
(c) Corresponding Reconstruc-
(a) Divergence Map and Aggregate Pos- tions
terior (b) Mean Curvature Map

Figure 23: Visualization of the representation learned by the 1-VAE trained on MPI3D Real.

2.0 0.28
2.0
1.5 0.21 0.8
1.5
1.0 0.14 0.6
1.0
0.4
0.5 0.07
0.5
Dimension 0

0.2
Dimension 0

0.0 0.00
0.0
0.0
0.5 -0.04
0.5 0.2
1.0 -0.08
1.0 0.4
1.5 -0.11
1.5 0.6
2.0 -0.15
2 1 0 1 2 2.0 0.8
Dimension 9 2 1 0 1 2
Dimension 9
(c) Corresponding Reconstruc-
(a) Divergence Map and Aggregate Pos- tions
terior (b) Mean Curvature Map

Figure 24: Visualization of the representation learned by the 1-VAE trained on MPI3D Real.

2.0 2.19
2.0
1.5 1.64 0.75
1.5
1.0 1.09 0.50
1.0
0.5 0.55
0.5 0.25
Dimension 9

Dimension 9

0.0 0.00
0.0 0.00
0.5 -0.34
0.5 0.25
1.0 -0.68
1.0 0.50
1.5 -1.01
1.5
0.75
2.0 -1.35
2 1 0 1 2 2.0
Dimension 11 2 1 0 1 2
Dimension 11
(c) Corresponding Reconstruc-
(a) Divergence Map and Aggregate Pos- tions
terior (b) Mean Curvature Map

Figure 25: Visualization of the representation learned by the 8-VAE trained on MPI3D Real. Note that
due to posterior collapse, the full latent manifold is contained in this projection (see the corresponding
response matrix in figure 19).

25
B.3 MNIST

Due to the computational cost of evaluating the response function over a dense grid, we focus our
visualizations to 2D projections of the latent space. However, for MNIST and Fashion-MNIST, we
train several VAE models to embed the whole representation into two dimensions d = 2, so that
we can visualize the full representation. While the resulting divergence and curvature maps do not
demonstrate as intuitive structure as in the disentangled representations for 3D-Shapes or MPI-3D,
we can nevertheless appreciate the learned manifold beyond qualitatively observing reconstructions.

26
27
B.3.1 MNIST
2.0 0.48 2.0

0.75
1.5 0.36 1.5

1.0 0.24 1.0 0.50

0.5 0.12 0.5 0.25

0.0 0.00 0.0

0.00

0.5 -0.10 0.5

0.25

1.0 -0.19 1.0

0.50

1.5 -0.29 1.5

0.75

2.0 -0.38 2.0

2.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5 2.0 2.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5 2.0

(a) Divergence Map and Aggregate Posterior (b) Mean Curvature Map

(c) Corresponding Reconstructions

Figure 26: The full latent space for a VAE (d = 2) model trained on MNIST. 26a shows the computed
divergence of the response field in blue and red while the green points are samples from the aggregate
posterior. 26b shows the resulting mean curvature, which identifies 10 points where the curvature
spikes and the boundaries between the regions corresponding to different clusters in the posterior.
Finally 26c shows the reconstructions over the same region. Note how the high divergence (red)
regions correspond to boundaries between significantly different samples (such as changing digit
value or stroke thickness).
28
 4 0.59 4

3 0.44 3 0.75

2 0.29 2 0.50

1 0.15 1
0.25

0 0.00 0
0.00

1 -0.13 1
0.25

2 -0.27 2
0.50

3 -0.40 3

0.75

4 -0.53 4
4 3 2 1 0 1 2 3 4 4 3 2 1 0 1 2 3 4

(a) Divergence Map and Aggregate Posterior (b) Mean Curvature Map

(c) Corresponding Reconstructions

Figure 27: Same plot and model as figure 26, except over a larger range of the latent space [−4, 4].
Note that even though the posterior (green dots) is concentrated near the prior (standard normal),
reconstructions far away (along the edges of the figure) still look recognizable, demonstrating the
exceptional robustness of VAEs to project unexpected latent vectors back onto the learned manifold.

29
 4

2
1

0
0

2
2

3 4

3 2 1 0 1 2 3 4 2 0 2 4

(a) Response field [−2, 2] (b) Response field [−4, 4]

Figure 28: Response fields for the same model analyzed in figures 26 and 27. The blue dots show
the initial latent samples, and the orange dots connected by the black arrows show the corresponding
responses (the latent sample after decoding and reencoding).

30
31
B.3.2 Fashion-MNIST
2.0 0.25 2.0

1.5 0.19 1.5 0.75

1.0 0.13 1.0 0.50

0.5 0.06 0.5 0.25

0.0 0.00 0.0 0.00

0.5 -0.06 0.5 0.25

1.0 -0.13 1.0

0.50

1.5 -0.19 1.5

0.75

2.0 -0.25 2.0

2.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5 2.0 2.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5 2.0

(a) Divergence Map and Aggregate Posterior (b) Mean Curvature Map

(c) Corresponding Reconstructions

Figure 29: The full latent space for a 8-VAE (d = 2) model trained on Fashion-MNIST. 29a shows
the computed divergence of the response field in blue and red while the green points are samples from
the aggregate posterior. 29b shows the resulting mean curvature, which identifies 10 points where the
curvature spikes and the boundaries between the regions corresponding to different clusters in the
posterior. Finally 29c shows the reconstructions over the same region. Note how the high divergence
(red) regions correspond to boundaries between significantly different samples (such as changing
digit value or stroke thickness).
32
 4 0.24 4

0.75
3 0.18 3

2 0.12 2 0.50

1 0.06 1 0.25

0 0.00 0 0.00

1 -0.10 1 0.25

2 -0.20 2 0.50

3 -0.31 3
0.75

4 -0.41 4
4 3 2 1 0 1 2 3 4 4 3 2 1 0 1 2 3 4

(a) Divergence Map and Aggregate Posterior (b) Mean Curvature Map

(c) Corresponding Reconstructions

Figure 30: Same plot and model as figure 29, except over a larger range of the latent space [−4, 4].
Note that even though the posterior (green dots) is concentrated near the prior (standard normal),
reconstructions far away (along the edges of the figure) still look recognizable, demonstrating the
exceptional robustness of VAEs to project unexpected latent vectors back onto the learned manifold.

33
 2.0 4

1.5 3

1.0 2

0.5 1

0.0 0

0.5 1

1.0 2

1.5 3

2.0
4
2.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5 2.0 4 3 2 1 0 1 2 3 4

(a) Response field [−2, 2] (b) Response field [−4, 4]

Figure 31: Response fields for the same model analyzed in figures 29 and 30. The blue dots show
the initial latent samples, and the orange dots connected by the black arrows show the corresponding
responses (the latent sample after decoding and reencoding).

34