Journal of Machine Learning Research 15 (2014) 3743-3773 Submitted 6/13; Published 11/14

What Regularized Auto-Encoders Learn from the

Data-Generating Distribution

Guillaume Alain guillaume.alain@umontreal.ca

Yoshua Bengio yoshua.bengio@umontreal.ca
Department of Computer Science and Operations Research
University of Montreal
Montreal, H3C 3J7, Quebec, Canada

Editors: Aaron Courville, Rob Fergus, and Christopher Manning

Abstract
What do auto-encoders learn about the underlying data-generating distribution? Recent
work suggests that some auto-encoder variants do a good job of capturing the local manifold
structure of data. This paper clarifies some of these previous observations by showing that
minimizing a particular form of regularized reconstruction error yields a reconstruction
function that locally characterizes the shape of the data-generating density. We show that
the auto-encoder captures the score (derivative of the log-density with respect to the input).
It contradicts previous interpretations of reconstruction error as an energy function. Unlike
previous results, the theorems provided here are completely generic and do not depend on
the parameterization of the auto-encoder: they show what the auto-encoder would tend
to if given enough capacity and examples. These results are for a contractive training
criterion we show to be similar to the denoising auto-encoder training criterion with small
corruption noise, but with contraction applied on the whole reconstruction function rather
than just encoder. Similarly to score matching, one can consider the proposed training
criterion as a convenient alternative to maximum likelihood because it does not involve
a partition function. Finally, we show how an approximate Metropolis-Hastings MCMC
can be setup to recover samples from the estimated distribution, and this is confirmed in
sampling experiments.
Keywords: auto-encoders, denoising auto-encoders, score matching, unsupervised repre-
sentation learning, manifold learning, Markov chains, generative models

1. Introduction
Machine learning is about capturing aspects of the unknown distribution from which the
observed data are sampled (the data-generating distribution). For many learning algorithms
and in particular in manifold learning, the focus is on identifying the regions (sets of points)
in the space of examples where this distribution concentrates, i.e., which configurations of
the observed variables are plausible.
Unsupervised representation-learning algorithms try to characterize the data-generating
distribution through the discovery of a set of features or latent variables whose variations
capture most of the structure of the data-generating distribution. In recent years, a number
of unsupervised feature learning algorithms have been proposed that are based on minimiz-
ing some form of reconstruction error, such as auto-encoder and sparse coding variants (Ol-

c 2014 Guillaume Alain and Yoshua Bengio.

 Alain and Bengio

shausen and Field, 1997; Bengio et al., 2007; Ranzato et al., 2007; Jain and Seung, 2008;
Ranzato et al., 2008; Vincent et al., 2008; Kavukcuoglu et al., 2009; Rifai et al., 2011b,a;
Gregor et al., 2011). An auto-encoder reconstructs the input through two stages, an encoder
function f , which outputs a learned representation h = f (x) of an example x, and a decoder
function g, such that g(f (x)) ≈ x for most x sampled from the data-generating distribu-
tion. These feature learning algorithms can be stacked to form deeper and more abstract
representations. Deep learning algorithms learn multiple levels of representation, where the
number of levels is data-dependent. There are theoretical arguments and much empiri-
cal evidence to suggest that when they are well-trained, deep learning algorithms (Hinton
et al., 2006; Bengio, 2009; Lee et al., 2009; Salakhutdinov and Hinton, 2009; Bengio and
Delalleau, 2011; Bengio et al., 2013b) can perform better than their shallow counterparts,
both in terms of learning features for the purpose of classification tasks and for generating
higher-quality samples.
Here we restrict ourselves to the case of continuous inputs x ∈ Rd with the data-
generating distribution being associated with an unknown target density function, denoted
p. Manifold learning algorithms assume that p is concentrated in regions of lower dimen-
sion (Cayton, 2005; Narayanan and Mitter, 2010), i.e., the training examples are by defini-
tion located very close to these high-density manifolds. In that context, the core objective
of manifold learning algorithms is to identify where the density concentrates.
Some important questions remain concerning many of feature learning algorithms based
on reconstruction error. Most importantly, what is their training criterion learning about
the input density? Do these algorithms implicitly learn about the whole density or only some
aspect? If they capture the essence of the target density, then can we formalize that link
and in particular exploit it to sample from the model? The answers may help to establish
that these algorithms actually learn implicit density models, which only define a density
indirectly, e.g., through the estimation of statistics or through a generative procedure.
These are the questions to which this paper contributes.
The paper is divided in two main sections, along with detailed appendices with proofs
of the theorems. Section 2 makes a direct link between denoising auto-encoders (Vincent
et al., 2008) and contractive auto-encoders (Rifai et al., 2011b), justifying the interest
in the contractive training criterion studied in the rest of the paper. Section 3 is the
main contribution and regards the following question: when minimizing that criterion, what
does an auto-encoder learn about the data-generating density? The main answer is that it
estimates the score (first derivative of the log-density), i.e., the direction in which density is
increasing the most, which also corresponds to the local mean, which is the expected value
in a small ball around the current location. It also estimates the Hessian (second derivative
of the log-density).
Finally, Section 4 shows how having access to an estimator of the score can be exploited
to estimate energy differences, and thus perform approximate MCMC sampling. This is
achieved using a Metropolis-Hastings MCMC in which the energy differences between the
proposal and the current state are approximated using the denoising auto-encoder. Ex-
periments on artificial data sets show that a denoising auto-encoder can recover a good
estimator of the data-generating distribution, when we compare the samples generated by
the model with the training samples, projected into various 2-D views for visualization.

3744
What Regularized Auto-Encoders Learn from the Data-Generating Distribution

Figure 1: Regularization forces the auto-encoder to become less sensitive to the input, but
minimizing reconstruction error forces it to remain sensitive to variations along
the manifold of high density. Hence the representation and reconstruction end
up capturing well variations on the manifold while mostly ignoring variations
orthogonal to it.

2. Contractive and Denoising Auto-Encoders

Regularized auto-encoders (see Bengio et al. 2012b for a review and a longer exposition)
capture the structure of the training distribution thanks to the productive opposition be-
tween reconstruction error and a regularizer. An auto-encoder maps inputs x to an internal
representation (or code) f (x) through the encoder function f , and then maps back f (x)
to the input space through a decoding function g. The composition of f and g is called
the reconstruction function r, with r(x) = g(f (x)), and a reconstruction loss function `
penalizes the error made, with r(x) viewed as a prediction of x. When the auto-encoder
is regularized, e.g., via a sparsity regularizer, a contractive regularizer (detailed below), or
a denoising form of regularization (that we find below to be very similar to a contractive
regularizer), the regularizer basically attempts to make r (or f ) as simple as possible, i.e.,
as constant as possible, as unresponsive to x as possible. It means that f has to throw away
some information present in x, or at least represent it with less precision. On the other
hand, to make reconstruction error small on the training set, examples that are neighbors
on a high-density manifold must be represented with sufficiently different values of f (x) or
r(x). Otherwise, it would not be possible to distinguish and hence correctly reconstruct
these examples. It means that the derivatives of f (x) or r(x) in the x-directions along the
manifold must remain large, while the derivatives (of f or r) in the x-directions orthogonal
to the manifold can be made very small. This is illustrated in Figure 1. In the case of prin-
cipal components analysis, one constrains the derivative to be exactly 0 in the directions
orthogonal to the chosen projection directions, and around 1 in the chosen projection di-
rections. In regularized auto-encoders, f is non-linear, meaning that it is allowed to choose
different principal directions (those that are well represented, i.e., ideally the manifold tan-
gent directions) at different x’s, and this allows a regularized auto-encoder with non-linear
encoder to capture non-linear manifolds. Figure 2 illustrates the extreme case when the
regularization is very strong (r wants to be nearly constant where density is high) in the
special case where the distribution is highly concentrated at three points (three training ex-
amples). It shows the compromise between obtaining the identity function at the training
examples and having a flat r near the training examples, yielding a vector field r(x) − x
that points towards the high density points.
Here we show that the denoising auto-encoder (Vincent et al., 2008) with very small
Gaussian corruption and squared error loss is actually a particular kind of contractive auto-
encoder (Rifai et al., 2011b), contracting the whole auto-encoder reconstruction function

3745
 Alain and Bengio

r(x)"

x1" x2" x3" x"

Figure 2: The reconstruction function r(x) (in turquoise) which would be learned by a high-capacity
auto-encoder on a 1-dimensional input, i.e., minimizing reconstruction error at the train-
ing examples xi (with r(xi ) in red) while trying to be as constant as possible otherwise.
The figure is used to exaggerate and illustrate the effect of the regularizer (corresponding
to a large σ 2 in the loss function L later described by Equation 6. The dotted line is
the identity reconstruction (which might be obtained without the regularizer). The blue
arrows shows the vector field of r(x)−x pointing towards high density peaks as estimated
by the model, and estimating the score (log-density derivative), as shown in this paper.

rather than just the encoder, whose contraction penalty coefficient is the magnitude of the
perturbation. This was first suggested in Rifai et al. (2011b).
The contractive auto-encoder, or CAE (Rifai et al., 2011b), is a particular form of reg-
ularized auto-encoder which is trained to minimize the following regularized reconstruction
error:
" #
∂f (x) 2
LCAE = E `(x, r(x)) + λ (1)
∂x F

where r(x) = g(f (x)) and ||A||2F is the sum of the squares of the elements of A. Both the
squared loss `(x, r) = ||x−r||2 and the cross-entropy loss `(x, r) = −x log r−(1−x) log(1−r)
have been used, but here we focus our analysis on the squared loss because of the easier
mathematical treatment it allows. Note that success in minimizing the CAE criterion
strongly depends on the parameterization of f and g and in particular on the tied weights
constraint used, with f (x) = sigmoid(W x + b) and g(h) = sigmoid(W T h + c). The above
regularizing term forces f (as well as g, because of the tied weights) to be contractive, i.e.,
to have singular values less than 1.1 Larger values of λ yield more contraction (smaller
singular values) where it hurts reconstruction error the least, i.e., in the local directions
where there are only little or no variations in the data. These typically are the directions
orthogonal to the manifold of high density concentration, as illustrated in Figure 2.

1. Note that an auto-encoder without any regularization would tend to find many leading singular values
near 1 in order to minimize reconstruction error, i.e., preserve input norm in all the directions of variation
present in the data.

3746
What Regularized Auto-Encoders Learn from the Data-Generating Distribution

The denoising auto-encoder, or DAE (Vincent et al., 2008), is trained to minimize the
following denoising criterion:

LDAE = E [`(x, r(N (x)))] (2)

where N (x) is a stochastic corruption of x and the expectation is over the training dis-
tribution and the corruption noise source. Here we consider mostly the squared loss and
Gaussian noise corruption, again because it is easier to handle them mathematically. In
many cases, the exact same proofs can be applied to any kind of additive noise, but Gaussian
noise serves as a good frame of reference.

Theorem 1 Let p be the probability density function of the data. If we train a DAE using
the expected quadratic loss and corruption noise N (x) = x +  with

 ∼ N 0, σ 2 I ,

then the optimal reconstruction function r∗ (x) will be given by

E [p(x − )(x − )]

r∗ (x) = (3)
E [p(x − )]

for values of x where p(x) 6= 0.

Moreover, if we consider how the optimal reconstruction function rσ∗ (x) behaves asymp-
totically as σ → 0, we get that
∂ log p(x)
rσ∗ (x) = x + σ 2 + o(σ 2 ) as σ → 0. (4)
∂x
The proof of this result is found in the Appendix. We make use of the small o notation
throughout this paper and assume that the reader is familiar with asymptotic notation. In
the context of Theorem 1, it has to be understood that all the other quantities except for
σ are fixed when we study the effect of σ → 0.
Equation 3 reveals that the optimal DAE reconstruction function at every point x is
given by a kind of convolution involving the density function p, or weighted average from
the points in the neighbourhood of x, depending on how we would like to view it. A higher
noise level σ means that a larger neighbourhood of x is taken into account. Note that the
total quantity of “mass” being included in the weighted average of the numerator of (3) is
found again at the denominator.
Gaussian noise is a simple case in the sense that it is additive and symmetrical, so it
avoids the complications that would occur when trying to integrate over the density of pre-
images x0 such that N (x0 ) = x for a given x. The ratio of those quantities that we have
in Equation 3, however, depends strongly on the decision that we made to minimize the
expected square error.

When we look at the asymptotic behavior with Equation 4, the first thing to observe is
that the leading term in the expansion of rσ∗ (x) is x, and then the remainder goes to 0 as
σ → 0. When there is no noise left at all, it should be clear that the best reconstruction
target for any value x would be that x itself.

3747
 Alain and Bengio

We get something even more interesting if we look at the second term of Equation 4
because it gives us an estimator of the score from
∂ log p(x)
= (rσ∗ (x) − x) /σ 2 + o(1) as σ → 0. (5)
∂x
This result is at the core of our paper. It is what allows us to start from a trained DAE,
and then recover properties of the training density p(x) that can be used to sample from
p(x).
Most of the asymptotic properties that we get by considering the limit as the Gaussian
noise level σ goes to 0 could be derived from a family of noise distribution that approaches
a point mass distribution in a relatively “nice” way.

An interesting connection with contractive auto-encoders can also be observed by using a

Taylor expansion of the denoising auto-encoder loss and assuming only that rσ (x) = x+o(1)
as σ → 0. In that case we get the following proposition.

Proposition 1 Let p be the probability density function of the data. Consider a DAE
using
2
the expected quadratic loss and corruption noise N (x) = x + , with  ∼ N 0, σ I . If we
assume that the non-parametric solutions rσ (x) satisfies

rσ (x) = x + o(1) as σ → 0,

then we can rewrite the loss as

" #
2
∂r(x)
LDAE = E kr(x) − xk22 +σ 2
+ o(σ 2 ) as σ → 0.
∂x F

The proof is in Appendix and uses a simple Taylor expansion around x.

Proposition 1 shows that the DAE with small corruption of variance σ 2 is similar to
a contractive auto-encoder with penalty coefficient λ = σ 2 but where the contraction is
imposed explicitly on the whole reconstruction function r(·) = g(f (·)) rather than on f (·)
alone.2
This analysis motivates the definition of the reconstruction contractive auto-encoder
(RCAE), a variation of the CAE where loss function is instead the squared reconstruction
loss plus contractive penalty on the reconstruction:
" #
2
2 2 ∂r(x)
LRCAE = E kr(x) − xk2 + σ . (6)
∂x F

This is an analytic version of the denoising criterion with small noise σ 2 , and also corre-
sponds to a contractive auto-encoder with contraction on both f and g, i.e., on r.
Because of the similarity between DAE and RCAE when taking λ = σ 2 and because
the semantics of σ 2 is clearer (as a squared distance in input space), we will denote σ 2 for
the penalty term coefficient in situations involving RCAE. For example, in the statement of
2. In the CAE there is a also a contractive effect on g(·) as a side effect of the parameterization with weights
tied between f (·) and g(·).

3748
What Regularized Auto-Encoders Learn from the Data-Generating Distribution

Theorem 2, this σ 2 is just a positive constant; there is no notion of additive Gaussian noise,
i.e., σ 2 does not explicitly refer to a variance, but using the notation σ 2 makes it easier to
intuitively see the connection to the DAE setting.
The connection between DAE and RCAE established in Proposition 1 also serves as
the basis for an alternative demonstration to Theorem 1 in which we study the asymptotic
behavior of the RCAE solution. This result is contained in the following theorem.

Theorem 2 Let p be a probability density function that is continuously differentiable once

and with support Rd (i.e., ∀x ∈ Rd we have p(x) 6= 0). Let Lσ2 be the loss function defined
by " #
Z 2
2 2 ∂r(x)
Lσ2 (r) = p(x) kr(x) − xk2 + σ dx (7)
Rd ∂x F

for r : Rd → Rd assumed to be differentiable twice, and 0 ≤ σ 2 ∈ R used as factor to the

penalty term.
Let rσ∗ 2 (x) denote the optimal function that minimizes Lσ2 . Then we have that

∂ log p(x)
rσ∗ 2 (x) = x + σ 2 + o(σ 2 ) as σ 2 → 0. (8)
∂x
Moreover, we also have the following expression for the derivative

∂rσ∗ 2 (x) ∂ 2 log p(x)

= I + σ2 + o(σ 2 ) as σ 2 → 0. (9)
∂x ∂x2
 ∗ Both these asymptotic expansions are to be understood in a context where we consider
rσ2 (x) σ2 ≥0 to be a family of optimal functions minimizing Lσ2 for their corresponding
value of σ 2 . The asymptotic expansions are applicable point-wise in x, that is, with any
fixed x we look at the behavior as σ 2 → 0.

The proof is given in the appendix and uses the Euler-Lagrange equations from the
calculus of variations.

3. Minimizing the Loss to Recover Local Features of p(·)

One of the central ideas of this paper is that in a non-parametric setting (without parametric
constraints on r), we have an asymptotic formula (as the noise level σ → 0) for the optimal
reconstruction function for the DAE and RCAE that allows us to recover the score ∂ log∂xp(x) .
A DAE is trained with a method that knows nothing about p, except through the use of
training samples to minimize a loss function, so it comes as a surprise that we can compute
the score of p at any point x.
In the following subsections we explore the consequences and the practical aspect of
this.

3.1 Empirical Loss

In an experimental setting, the expected loss (7) is replaced by the empirical loss

3749
 Alain and Bengio

N
!
2
1 X (n) (n)
2
2 ∂r(x)
L̂ = r(x )−x +σ
N 2 ∂x x=x(n) F
n=1
N
based on a sample x(n) n=1 drawn from p(x).


Alternatively, the auto-encoder is trained online (by stochastic gradient updates) with a
stream of examples x(n) , which corresponds to performing stochastic gradient descent on the
expected loss (7). In both cases we obtain an auto-encoder that approximately minimizes
the expected loss.
An interesting question is the following: what can we infer from the data-generating
density when given an auto-encoder reconstruction function r(x)?
The premise is that this auto-encoder r(x) was trained to approximately minimize a
loss function that has exactly the form of (7) for some σ 2 > 0. This is assumed to have
been done through minimizing the empirical loss and the distribution p was only available
N
indirectly through the samples x(n) n=1 . We do not have access to p or to the samples.


We have only r(x) and maybe σ 2 .

We will now discuss the usefulness of r(x) based on different conditions such as the
model capacity and the value of σ 2 .

3.2 Perfect World Scenario

As a starting point, we will assume that we are in a perfect situation, i.e., with no con-
straint on r (non-parametric setting), an infinite amount of training data, and a perfect
minimization. We will see what can be done to recover information about p in that ideal
case. Afterwards, we will drop certain assumptions one by one and discuss the possible
paths to getting back some information about p.
We use notation rσ2 (x) when we want to emphasize the fact that the value of r(x) came
from minimizing the loss with a certain fixed σ 2 .
Suppose that rσ2 (x) was trained with an infinite sample drawn from p. Suppose also that
it had infinite (or sufficient) model capacity and that it is capable of achieving the minimum
of the loss function (7) while satisfying the requirement that r(x) be twice differentiable.
Suppose that we know the value of σ 2 and that we are working in a computing environment
of arbitrary precision (i.e., no rounding errors).
As shown by Theorem 1 and Theorem 2, we would be able to get numerically the values
∂ log p(x)
of ∂x at any point x ∈ Rd by simply evaluating

rσ2 (x) − x ∂ log p(x)

→ as σ 2 → 0. (10)
σ2 ∂x
In the setup described, we do not get to pick values of σ 2 so as to take the limit σ 2 → 0.
However, it is assumed that σ 2 is already sufficiently small that the above quantity is close
to ∂ log∂xp(x) for all intents and purposes.

3.3 Simple Numerical Example

To give an example of this in one dimension, we will show what happens when we train
a non-parametric model r̂(x) to minimize numerically the loss relative to p(x). We train

3750
What Regularized Auto-Encoders Learn from the Data-Generating Distribution

both a DAE and an RCAE in this fashion by minimizing a discretized version of their losses
defined by equations (2) and (6). The goal here is to show that, for either a DAE or RCAE,
the approximation of the score that we get through Equation 5 gets arbitrarily close to the
∂
actual score ∂x log p(x) as σ → 0.
The distribution p(x) studied is shown in Figure 3 (left) and it was created to be simple
enough to illustrate the mechanics. We plot p(x) in Figure 3 (left) along with the score of
p(x) (right).

1 ∂ ∂
(a) p(x) = Z
exp(−E(x)) (b) ∂x
log p(x) = − ∂x E(x)

Figure 3: The density p(x) and its score for a simple one-dimensional example.

The model r̂(x) is fitted by dividing the interval [−1.5, 1.5] into M = 1000 partition
points x1 , . . . , xM evenly separated by a distance ∆. The discretized version of the RCAE
loss function is

M M −1  2
X 2 2
X r̂(xi+1 ) − r̂(xi )
p(xi )∆ (r̂(xi ) − xi ) + σ p(xi )∆ . (11)
∆
i=1 i=1

Every value r̂(xi ) for i = 1, . . . , M is treated as a free parameter. Setting to 0 the derivative
with respect to the r̂(xi ) yields a system of linear equations in M unknowns that we can
solve exactly. From that RCAE solution r̂ we get an approximation of the score of p at each
point xi . A similar thing can be done for the DAE by using a discrete version of the exact
solution (3) from Theorem 1. We now have two ways of approximating the score of p.
In Figure 4 we compare the approximations to the actual score of p for decreasingly
small values of σ ∈ {1.00, 0.31, 0.16, 0.06}.

3.4 Vector Field Around a Manifold

We extend the experimentation of Section 3.3 to a 1-dimensional manifold in 2-D space, in
which one can visualize r(x) − x as a vector field, and we go from the non-parametric esti-
mator of the previous section to an actual auto-encoder trained by numerically minimizing
the regularized reconstruction error.

3751
 Alain and Bengio

Figure 4: Comparing the approximation of the score of p given by discrete versions of

optimally trained auto-encoders with infinite capacity. The approximations given
by the RCAE are in orange while the approximations given by the DAE are in
purple. The results are shown for decreasing values of σ ∈ {1.00, 0.31, 0.16, 0.06}
that have been selected for their visual appeal.
As expected, we see in that the RCAE (orange) and DAE (purple) approximations
of the score are close to each other as predicted by Proposition 1. Moreover, they
are also converging to the true score (green) as predicted by Theorem 1 and
Theorem 2.

Two-dimensional data points (x, y) were generated along a spiral according to the fol-
lowing equations:

x = 0.04 sin(t), y = 0.04 cos(t), t ∼ Uniform (3, 12) .

A denoising auto-encoder was trained with Gaussian corruption noise σ = 0.01. The
encoder is f (x) = tanh(b + W x) and the decoder is g(h) = c + V h. The parameters
(b, c, V, W ) are optimized by BFGS to minimize the average squared error, using a fixed
training set of 10 000 samples (i.e., the same corruption noises were sampled once and for
all). We found better results with untied weights, and BFGS gave more accurate models
than stochastic gradient descent. We used 1000 hidden units and ran BFGS for 1000
iterations.

3752
What Regularized Auto-Encoders Learn from the Data-Generating Distribution

The non-convexity of the problem makes it such that the solution found depends on the
initialization parameters. The random corruption noise used can also influence the final
outcome. Moreover, the fact that we are using a finite training sample size with reasonably
small noise may allow for undesirable behavior of r in regions far away from the training
samples. For those reasons, we trained the model multiple times and selected two of the
most visually appealing outcomes. These are found in Figure 5 which features a more global
perspective along with a close-up view.

(a) r(x) − x vector field, acting as sink, zoomed out (b) r(x) − x vector field, close-up

Figure 5: The original 2-D data from the data-generating density p(x) is plotted along
with the vector field defined by the values of r(x) − x for trained auto-encoders
(corresponding to the estimation of the score ∂ log∂xp(x) ).

Figure 5 shows the data along with the learned score function (shown as a vector field).
We see that that the vector field points towards the nearest high-density point on the data
manifold. The vector field is close to zero near the manifold (i.e., the reconstruction error is
close to zero), also corresponding to peaks of the implicitly estimated density. The points
on the manifolds play the role of sinks for the vector field. Other places where reconstruction
error may be low, but where the implicit density is not high, are sources of the vector field.
In Figure 5(b) we can see that we have that kind of behavior halfway between two sections
of the manifold. This shows that reconstruction error plays a very different role as what
was previously hypothesized: whereas in Ranzato et al. (2008) the reconstruction error
was viewed as an energy function, our analysis suggests that in regularized auto-encoders,
it is the norm of an approximate score, i.e., the derivative of the energy w.r.t. input. Note
that the norm of the score should be small near training examples (corresponding to local
maxima of density) but it could also be small at other places corresponding to local minima
of density. This is indeed what happens in the spiral example shown. It may happen
whenever there are high-density regions separated by a low-density region: tracing paths
from one high-density region to another should cross a “median” lower-dimensional region
(a manifold) where the density has a local maximum along the path direction. The reason
such a median region is needed is because at these points the vectors r(x) − x must change

3753
 Alain and Bengio

sign: on one side of the median they point to one of the high-density regions while on the
other side they point to the other, as clearly visible in Figure 5(b) between the arms of the
spiral.
We believe that this analysis is valid not just for contractive and denoising auto-encoders,
but for regularized auto-encoders in general. The intuition behind this statement can be
firmed up by analyzing Figure 2: the score-like behavior of r(x) − x arises simply out of the
opposing forces of (a) trying to make r(x) = x at the training examples and (b) trying to
make r(x) as regularized as possible (as close to a constant as possible).
Note that previous work (Rifai et al., 2012; Bengio et al., 2013b) has already shown that
contractive auto-encoders (especially when they are stacked in a way similar to RBMs in
a deep belief net) learn good models of high-dimensional data (such as images), and that
these models can be used not just to obtain good representations for classification tasks
but that good quality samples can be obtained from the model, by a random walk near the
manifold of high-density. This was achieved by essentially following the vector field and
adding noise along the way.

3.5 Missing σ 2
When we are in the same setting as in Section 3.2 but the value of σ 2 is unknown, we can
modify (10) a bit and avoid dividing by σ 2 . That is, for a trained reconstruction function
r(x) given to us we just take the quantity r(x) − x and it should be approximately the score
up to a multiplicative constant. We get that

∂ log p(x)
r(x) − x ∝ .
∂x

Equivalently, if one estimates the density via an energy function (minus the unnormalized
log density), then x − r(x) estimates the derivative of the energy function.
We still have to assume that σ 2 is small. Otherwise, if the unknown σ 2 is too large we
might get a poor estimation of the score.

3.6 Limited Parameterization

We should also be concerned about the fact that r(x) − x is trying to approximate − ∂E(x)
∂x
as σ 2 → 0 but we have not made any assumptions about the space of functions that r can
represent when we are dealing with a specific implementation.
When using a certain parameterization of r such as the one from Section 3.3, there is
no guarantee that the family of functions in which we select r each represent a conservative
vector field (i.e., the gradient of a potential function). Even if we start from a density
∂
p(x) ∝ exp(−E(x)) and we have that r(x) − x is very close to − ∂x E(x) in terms of some
given norm, there is not guarantee that there exists an associated function E0 (x) for which
∂
r(x) − x ∝ − ∂x E0 (x) and E0 (x) ≈ E(x).
In fact, in many cases we can trivially show the non-existence of such a E0 (x) by
computing the curl of r(x). The curl has to be equal to 0 everywhere if r(x) is indeed the
derivative of a potential function. We can omit the x terms from the computations because
∂
we can easily find its antiderivative by looking at x = ∂x kxk22 .

3754
What Regularized Auto-Encoders Learn from the Data-Generating Distribution

Conceptually, another way to see this is to argue that if such a function E0 (x) existed,
its second-order mixed derivatives should be equal. That is, we should have that
∂ 2 E0 (x) ∂ 2 E0 (x)
= ∀i, j,
∂xi ∂xj ∂xj ∂xi
which is equivalent to
∂ri (x) ∂rj (x)
= ∀i, j.
∂xj ∂xi
Again in the context of Section 3.3, with the parameterization used for that particular
kind of denoising auto-encoder, this would yield the constraint that V T = W . That is,
unless we are using tied weights, we know that no such potential E0 (x) exists, and yet when
running the experiments from Section 3.3 we obtained much better results with untied
weights. To make things worse, it can also be demonstrated that the energy function that
we get from tied weights leads to a distribution that is not normalizable (it has a divergent
integral over Rd ). In that sense, this suggests that we should not worry too much about
the exact parameterization of the denoising auto-encoder as long as it has the required
flexibility to approximate the optimal reconstruction function sufficiently well.

3.7 Relation to Denoising Score Matching

There is a connection between our results and previous research involving score matching for
denoising auto-encoders. We will summarize here the existing results from Vincent (2011)
and show that, whereas they have shown that denoising auto-encoders with a particular form
estimated the score, our results extend this to a very large family of estimators (including
the non-parametric case). This will provide some reassurance given some of the potential
issues mentioned in Section 3.6.
Motivated by the analysis of denoising auto-encoders, the authors of Vincent (2011) are
concerned with the case where we explicitly parameterize an energy function E(x), yielding
an associated score function ψ(x) = − ∂E(x)∂x and we stochastically corrupt the original
samples x ∼ p to obtain noisy samples x̃ ∼ qσ (x̃|x). In particular, the article analyzes the
case where qσ adds Gaussian noise of variance σ 2 to x. The main result is that minimizing
the expected square difference between ψ(x̃) and the score of qσ (x̃|x),
" #
∂ log qσ (x̃|x) 2
Ex,x̃ ψ(x̃) − ,
∂ x̃
is equivalent to performing
R score matching (Hyvärinen, 2005) with estimator ψ(x̃) and target
density qσ (x̃) = qσ (x̃|x)p(x)dx, where p(x) generates the training samples x. Note that
when a finite training set is used, qσ (x̃) is simply a smooth of the empirical distribution
(e.g., the Parzen density with Gaussian kernel of width σ). When the corruption noise
is Gaussian, qσ ∂(x̃|x)
x̃ = x−x̃
σ2
, from which we can deduce that if we define a reconstruction
function
r(x̃) = x̃ + σ 2 ψ(x̃), (12)
then the above expectation is equivalent to
" #
2
r(x̃) − x̃ x − x̃ 1 h
2
i
Ex,x̃ − = Ex,x̃ kr(x̃) − xk
σ2 σ2 σ2

3755
 Alain and Bengio

which is the denoising criterion. This says that when the reconstruction function r is pa-
rameterized so as to correspond to the score ψ of a model density (as per Equation 12, and
where ψ is a derivative of some log-density), the denoising criterion on r with Gaussian cor-
ruption noise is equivalent to score matching with respect to a smooth of the data-generating
density, i.e., a regularized form of score matching. Note that this regularization appears
desirable, because matching the score of the empirical distribution (or an insufficiently
smoothed version of it) could yield undesirable results when the training set is finite. Since
score matching has been shown to be a consistent induction principle (Hyvärinen, 2005), it
means that this denoising score matching (Vincent, 2011; Kingma and LeCun, 2010; Swer-
sky et al., 2011) criterion recovers the underlying density, up to the smoothing induced by
the noise of variance σ 2 . By making σ 2 small, we can make the estimator arbitrarily good
(and we would expect to want to do that as the amount of training data increases). Note
the correspondence of this conclusion with the results presented here, which show (1) the
equivalence between the RCAE’s regularization coefficient and the DAE’s noise variance σ 2 ,
and (2) that minimizing the equivalent analytic criterion (based on a contraction penalty)
estimates the score when σ 2 is small. The difference is that our result holds even when r
is not parameterized as per Equation 12, i.e., is not forced to correspond with the score
function of a density.

3.8 Estimating the Hessian

Since we have r(x)−x
σ2
as an estimator of the score, we readily obtain that the Hessian of
the log-density, can be estimated by the Jacobian of the reconstruction function minus the
identity matrix:
∂ 2 log p(x) ∂r(x)
≈( − I)/σ 2
∂x2 ∂x
as shown by Equation 9 of Theorem 2.
In spite of its simplicity, this result is interesting because it relates the derivative of
the reconstruction function, i.e., a Jacobian matrix, with the second derivative of the log-
density (or of the energy). This provides insights into the geometric interpretation of the
reconstruction function when the density is concentrated near a manifold. In that case, near
the manifold the score is nearly 0 because we are near a ridge of density, and the density’s
second derivative matrix tells us in which directions the first density remains close to zero
or increases. The ridge directions correspond to staying on the manifold and along these
directions we expect the second derivative to be close to 0. In the orthogonal directions,
the log-density should decrease sharply while its first and second derivatives would be large
in magnitude and negative in directions away from the manifold.
Returning to the above equation, keep in mind that in these derivations σ 2 is near
0 and r(x) is near x, so that ∂r(x) ∂x is close to the identity. In particular, in the ridge
(manifold) directions, we should expect ∂r(x) ∂x to be closer to the identity, which means
that the reconstruction remains faithful (r(x) = x) when we move on the manifold, and this
corresponds to the eigenvalues of ∂r(x)
∂x that are near 1, making the corresponding eigenvalues
∂ 2 log p(x) ∂r(x)
of ∂x2
near 0. On the other hand, in the directions orthogonal to the manifold, ∂x
∂2 log p(x)
should be smaller than 1, making the corresponding eigenvalues of ∂x2
negative.

3756
What Regularized Auto-Encoders Learn from the Data-Generating Distribution

Besides first and second derivatives of the density, other local properties of the density
are its local mean and local covariance, discussed in the Appendix, Section D.

4. Sampling with Metropolis-Hastings

In this section we show how a technique to generate samples from a given denoising auto-
encoder by using Metropolis-Hastings.
We start by explaining how to compute differences in energies Section 4.1, and then we
use this in Section 4.2 to generate samples. We provide an experimental example and we
discuss the potential problems that this method has.
This section serves as a demonstration that the main result of this paper, i.e., the
connection between denoising auto-encoders and the score ∂ log∂xp(x) , is more than just an
observation : it can have practical uses.

4.1 Estimating Energy Differences

One of the immediate consequences of Theorem 2 and Equation 10 is that, while we cannot
easily recover the energy E(x) itself, it is possible to approximate the energy difference
E(x∗ ) − E(x) between two states x and x∗ . This can be done by using a first-order Taylor
approximation
∂E(x) T ∗
E(x∗ ) − E(x) = (x − x) + o(kx∗ − xk).
∂x
To get a more accurate approximation, we can also use a path integral from x to x∗ that we
can discretize in sufficiently many steps. With a smooth path γ(t) : [0, 1] → Rd , assuming
that γ stays in a region where our DAE/RCAE can be used to approximate ∂E ∂x well enough,
we have that Z 1 T
∗ ∂E
E(x ) − E(x) = (γ(t)) γ 0 (t)dt. (13)
0 ∂x
The simplest way to discretize this path integral is to pick points {xi }ni=1 spread at even
distances on a straight line from x1 = x to xn = x∗ . We approximate (13) by
n  T
∗ 1 X ∂E
E(x ) − E(x) ≈ (xi ) (x∗ − x) (14)
n ∂x
i=1

4.2 Sampling
With Equation 13 from Section 4.1 we can perform approximate sampling from the es-
timated distribution, using the score estimator to approximate energy differences which
are needed in the Metropolis-Hastings accept/reject decision. Using a symmetric proposal
q(x∗ |x), the acceptance ratio is

p(x∗ )
α= = exp(−E(x∗ ) + E(x))
p(x)

which can be computed with (13) or approximated with (14) as long as we trust that
our DAE/RCAE was trained properly and has enough capacity to be a sufficiently good

3757
 Alain and Bengio

estimator of ∂E
∂x . An example of this process is shown in Figure 6 in which we sample from a
density concentrated around a 1-d manifold embedded in a space of dimension 10. For this
particular task, we have trained only DAEs and we are leaving RCAEs out of this exercise.
Given that the data is roughly contained in the range [−1.5, 1.5] along all dimensions, we
selected a training noise level σtrain = 0.1 so that the noise would have an appreciable
effect while still being relatively small. As required by Theorem 1, we have used isotropic
2
Gaussian noise of variance σtrain .
The Metropolis-Hastings proposal q(x∗ |x) = N (0, σMH 2 I) has a noise parameter σ
MH
that needs to be set. In the situation shown in Figure 6, we used σMH = 0.1. After some
hyperparameter tweaking and exploring various scales for σtrain , σMH , we found that setting
both to be 0.1 worked well.
When σtrain is too large, the DAE trained learns a “blurry” version of the density that
fails to represent the details that we are interested in. The samples shown in Figure 6 are
very convincing in terms of being drawn from a distribution that models well the original
density. We have to keep in mind that Theorem 2 describes the behavior as σtrain → 0 so
we would expect that the estimator becomes worse when σtrain is taking on larger values.
In this particular case with σtrain = 0.1, it seems that we are instead modeling something
2
like the original density to which isotropic Gaussian noise of variance σtrain has been added.
In the other extreme, when σtrain is too small, the DAE is not exposed to any training
example farther away from the density manifold. This can lead to various kinds of strange
behaviors when the sampling algorithm falls into those regions and then has no idea what
to do there and how to get back to the high-density regions. We come back to that topic
in Section 4.3.
It would certainly be possible to pick both a very small value for σtrain = σMH = 0.01
to avoid the spurious maxima problem illustrated in Section 4.3. However, this leads to the
same kind of mixing problems that any kind of MCMC algorithm has. Smaller values of
σMH lead to higher acceptance ratios but worse mixing properties.

4.3 Spurious Maxima

There are two very real concerns with the sampling method discussed in Section 4.2. The
first problem is with the mixing properties of MCMC and it is discussed in that section.
The second issue is with spurious probability maxima resulting from inadequate training of
the DAE. It happens when an auto-encoder lacks the capacity to model the density with
enough precision, or when the training procedure ends up in a bad local minimum (in terms
of the DAE parameters).
This is illustrated in Figure 7 where we show an example of a vector field r(x) − x
for a DAE that failed to properly learn the desired behavior in regions away from the
spiral-shaped density.

5. Conclusion
Whereas auto-encoders have long been suspected of capturing information about the data-
generating density, this work has clarified what some of them are actually doing, showing
that they can actually implicitly recover the data-generating density altogether. We have
shown that regularized auto-encoders such as the denoising auto-encoder and a form of con-

3758
 What Regularized Auto-Encoders Learn from the Data-Generating Distribution

original sampled original sampled

Figure 6: Samples drawn from the estimate of ∂E ∂x given by a DAE by the Metropolis-
Hastings method presented in Section 4. By design, the data density distribu-
tion is concentrated along a 1-d manifold embedded in a space of dimension 10.
This data can be visualized in the plots above by plotting pairs of dimensions
(x0 , x1 ), . . . , (x8 , x9 ), (x9 , x0 ), going in reading order from left to right and then
line by line. For each pair of dimensions, we show side by side the original data
(left) with the samples drawn (right).

tractive auto-encoder are closely related to each other and estimate local properties of the
data-generating density: the first derivative (score) and second derivative of the log-density,
as well as the local mean. This contradicts the previous interpretation of reconstruction
error as being an energy function (Ranzato et al., 2008) but is consistent with our exper-
imental findings. Our results do not require the reconstruction function to correspond to
the derivative of an energy function as in Vincent (2011), but hold simply by virtue of
minimizing the regularized reconstruction error training criterion. This suggests that min-
imizing a regularized reconstruction error may be an alternative to maximum likelihood
for unsupervised learning, avoiding the need for MCMC in the inner loop of training, as
in RBMs and deep Boltzmann machines, analogously to score matching (Hyvärinen, 2005;

3759
 Alain and Bengio

(a) DAE misbehaving when away from manifold (b) sampling getting trapped into bad attractor

Figure 7: (a) On the left we show a r(x) − x vector field similar to that of the earlier
Figure 5. The density is concentrated along a spiral manifold and we should have
the reconstruction function r bringing us back towards the density. In this case,
it works well in the region close to the spiral (the magnitude of the vectors is so
small that the arrows are shown as dots). However, things are out of control in
the regions outside. This is because the level of noise used during training was so
small that not enough of the training examples were found in those regions.
(b) On the right we sketch what may happen when we follow a sampling procedure
as described in Section 4.2. We start in a region of high density (in purple) and we
illustrate in red the trajectory that our samples may take. In that situation, the
DAE/RCAE was not trained properly. The resulting vector field does not reflect
the density accurately because it should not have this attractor (i.e., stable fixed
point) outside of the manifold on which the density is concentrated. Conceptually,
the sampling procedure visits that spurious attractor because it assumes that it
corresponds to a region of high probability. In some cases, this effect is regrettable
but not catastrophic, but in other situations we may end up with completely
unusable samples. In the experiments, training with enough of the examples
involving sufficiently large corruption noise typically eliminates that problem.

Vincent, 2011). Toy experiments have confirmed that a good estimator of the density can be
obtained when this criterion is non-parametrically minimized. The experiments have also
confirmed that an MCMC could be setup that approximately samples from the estimated
model, by estimating energy differences to first order (which only requires the score) to
perform approximate Metropolis-Hastings MCMC.
Many questions remain open and deserve further study. A big question is how to gener-
alize these ideas to discrete data, since we have heavily relied on the notions of scores, i.e.,
of derivatives with respect to x. A natural extension of the notion of score that could be
applied to discrete data is the notion of relative energy, or energy difference between a point
x and a perturbation x̃ of x. This notion has already been successfully applied to obtain the
equivalent of score matching for discrete models, namely ratio matching (Hyvärinen, 2007).

3760
What Regularized Auto-Encoders Learn from the Data-Generating Distribution

More generally, we would like to generalize to any form of reconstruction error (for exam-
ple many implementations of auto-encoders use a Bernoulli cross-entropy as reconstruction
loss function) and any (reasonable) form of corruption noise (many implementations use
masking or salt-and-pepper noise, not just Gaussian noise). More fundamentally, the need
to rely on σ → 0 is troubling, and getting rid of this limitation would also be very useful. A
possible solution to this limitation, as well as adding the ability to handle both discrete and
continuous variables, has recently been proposed while this article was under review (Bengio
et al., 2013a).
It would also be interesting to generalize the results presented here to other regularized
auto-encoders besides the denoising and contractive types. In particular, the commonly
used sparse auto-encoders seem to fit the qualitative pattern illustrated in 2 where a score-
like vector field arises out of the opposing forces of minimizing reconstruction error and
regularizing the auto-encoder.
We have mostly considered the harder case where the auto-encoder parameterization
does not guarantee the existence of an analytic formulation of an energy function. It would
be interesting to compare experimentally and study mathematically these two formulations
to assess how much is lost (because the score function may be somehow inconsistent) or
gained (because of the less constrained parameterization).

Acknowledgments

The authors thank Salah Rifai Max Welling, Yutian Chen and Pascal Vincent for fruitful
discussions, and acknowledge the funding support from NSERC, Canada Research Chairs
and CIFAR.

Appendix A. Exact Solution for DAE

There is a way to get an exact solution to the DAE loss (2) without assuming that σ → 0
or that the noise is Gaussian (but still using the quadratic loss).
We let p be the density function of the data such that ∀x ∈ Rd , p(x) > 0, and we use
additive isotropic Gaussian noise of variance σ 2 . We are in the non-parametric setting in
which we are minimizing
Z h i
LDAE = E∼N (0,σ2 I) p(x) kr(x + ) − xk22 dx (15)
Rd

with respect to the function r : Rd → Rd .

By using an auxiliary variable x̃ = x + , we can rewrite this loss in a way that puts
the quantity r(x̃) in focus and allows us to perform the minimization with respect to each
choice of r(x̃) independently. That is, we have that
Z h i
LDAE = E∼N (0,σ2 I) p(x̃ − ) kr(x̃) − x̃ + k22 dx̃ (16)
Rd

which can be differentiated with respect to the quantity r(x̃) and set to be equal to 0.
Denoting the optimum by r∗ (x̃), we get

3761
 Alain and Bengio

0 = E∼N (0,σ2 I) [p(x̃ − )r∗ (x̃) − x̃ + ] (17)

E∼N (0,σ2 I) [p(x̃ − )r∗ (x̃)] = E∼N (0,σ2 I) [p(x̃ − )(x̃ − )] (18)

E∼N (0,σ2 I) [p(x̃ − )(x̃ − )]

r∗ (x̃) = (19)
E∼N (0,σ2 I) [p(x̃ − )]

Conceptually, this means that the optimal DAE reconstruction function at every point
x̃ ∈ Rd is given by a kind of convolution involving the density function p, or weighted average
from the points in the neighbourhood of x̃, depending on how we would like to view it. A
higher noise level σ means that a larger neighbourhood of x̃ is taken into account. Note
that the total quantity of “mass” being included in the weighted average of the numerator
of (19) is found again at the denominator.

Appendix B. Relationship between Contractive Penalty and Denoising

Criterion
Theorem 1 When using corruption noise N (x) = x +  with

 ∼ N 0, σ 2 I ,

the objective function LDAE is

" #!
2
∂r(x)
E kx − r(x)k2 + σ 2 E + o(σ 2 )
 
LDAE =
∂x F

as σ → 0.

Proof With a Taylor expansion around x we have that

∂r(x)
r(x + ) = r(x) +  + o(σ 2 ).
∂x
Substituting this into LDAE we have that
   2
1 ∂r(x) 2
LDAE = E 2 x − r(x) + ∂x  + o(σ )
  
 2
 T ∂r(x) T
= E kx − r(x)k − 2E[] E ∂x (x − r(x))
  
T ∂r(x)
+T r E T E ∂r(x) + o(σ 2 )
 
∂x ∂x
  
2
1 2 2 ∂r(x)
+ o(σ 2 )
 
= 2 E kx − r(x)k + σ E ∂x (20)
F

where in the second line we used the independence of the noise from x and properties
of the trace, while in the last line we used E T = σ 2 I and E[] = 0 by definition of .
 

3762
 What Regularized Auto-Encoders Learn from the Data-Generating Distribution

Appendix C. Calculus of Variations

Theorem 2 Let p be a probability density function that is continuously differentiable once
and with support Rd (i.e., ∀x ∈ Rd we have p(x) 6= 0). Let Lσ2 be the loss function defined
by " #
Z 2
2 2 ∂r(x)
Lσ2 (r) = p(x) kr(x) − xk2 + σ dx
Rd ∂x F
for r : Rd → Rd assumed to be differentiable twice, and 0 ≤ σ 2 ∈ R used as factor to the
penalty term.
Let rσ∗ 2 (x) denote the optimal function that minimizes Lσ2 . Then we have that
∂ log p(x)
rσ∗ 2 (x) = x + σ 2 + o(σ 2 ) as σ 2 → 0.
∂x
Moreover, we also have the following expression for the derivative
∂rσ∗ 2 (x) ∂ 2 log p(x)
= I + σ2 + o(σ 2 ) as σ 2 → 0.
∂x ∂x2
Both
 ∗ these asymptotic expansions are to be understood in a context where we consider
rσ2 (x) σ2 ≥0 to be a family of optimal functions minimizing Lσ2 for their corresponding
value of σ 2 . The asymptotic expansions are applicable point-wise in x, that is, with any
fixed x we look at the behavior as σ 2 → 0.
Proof
This proof is done in two parts. In the first part, the objective is to get to Equation 23
that has to be satisfied for the optimum solution.
We will leave out the σ 2 indices from the expressions involving r(x) to make the notation
lighter. We have a more important need for indices k in rk (x) that denote the d components
of r(x) ∈ Rd .
We treat σ 2 as given and constant for the first part of this proof.
In the second part we work out the asymptotic expansion in terms of σ 2 . We again work
with the implicit dependence of r(x) on σ 2 .
(part 1 of the proof )
We make use of the Euler-Lagrange equation from the Calculus of Variations. We would
refer the reader to either (Dacorogna, 2004) or Wikipedia for more on the topic. Let
" #
2
∂r(x)
f (x1 , . . . , xn , r, rx1 , . . . , rxn ) = p(x) kr(x) − xk22 + σ 2
∂x F
∂f
where x = (x1 , . . . , xd ) , r(x) = (r1 (x), . . . , rd (x)) and rxi = ∂xi .
We can rewrite the loss L(r) more explicitly as
 
Z d d X d 2
X X ∂ri (x) 
L(r) = p(x)  (ri (x) − xi )2 + σ 2 dx
Rd ∂xj
i=1 i=1 j=1
 
d Z d 2
X X ∂ri (x) 
= p(x) (ri (x) − xi )2 + σ 2 dx (21)
R d ∂xj
i=1 j=1

3763
 Alain and Bengio

to observe that the components r1 (x), . . . , rd (x) can each be optimized separately.

The Euler-Lagrange equation to be satisfied at the optimal r : Rd → Rd is

d
∂f X ∂ ∂f
= .
∂r ∂xi ∂rxi
i=1

In our situation, the expressions from that equation are given by

∂f
= 2(r(x) − x)p(x)
∂r

∂f h
∂r1 ∂r2 ∂rd
iT
= 2σ 2 p(x) ∂xi ∂xi ··· ∂xi
∂rxi

 
∂ ∂f ∂p(x) h ∂r1 ∂r2 ∂rd
iT
= 2σ 2 ∂xi ∂xi ··· ∂xi
∂xi ∂rxi ∂xi
h iT
∂ 2 r1 ∂ 2 r2 ∂ 2 rd
+2σ 2 p(x) ∂x2i ∂x2i
··· ∂x2i

and the equality to be satisfied at the optimum becomes

∂p(x) ∂r1 2
+ p(x) ∂∂xr21
 
d ∂xi ∂xi i

(r(x) − x)p(x) = σ 2
X  .. 
. (22)
.

 
i=1 ∂p(x) ∂rd 2
∂xi ∂xi + p(x) ∂∂xr2d
i

As Equation 21 hinted, the expression (22) can be decomposed into the different com-
ponents rk (x) : Rd → R that make r. For k = 1, . . . , d we get
d 
∂ 2 rk (x)

2
X ∂p(x) ∂rk (x)
(rk (x) − xk )p(x) = σ + p(x) .
i=1
∂xi ∂xi ∂x2i
∂p(x)
As p(x) 6= 0 by hypothesis, we can divide all the terms by p(x) and note that ∂xi /p(x) =
∂ log p(x)
∂xi .

We get
d 
∂ 2 rk (x)

2
X ∂ log p(x) ∂rk (x)
rk (x) − xk = σ + . (23)
i=1
∂xi ∂xi ∂x2i

This first thing to observe is that when σ 2 = 0 the solution is just rk (x) = xk , which
translates into r(x) = x. This is not a surprise because it represents the perfect reconstruc-
tion value that we get when we the penalty term vanishes in the loss function.

(part 2 of the proof )

3764
What Regularized Auto-Encoders Learn from the Data-Generating Distribution

This linear partial differential Equation 23 can be used as a recursive relation for rk (x)
to obtain a Taylor series in σ 2 . The goal is to obtain an expression of the form

r(x) = x + σ 2 h(x) + o(σ 2 ) as σ2 → 0 (24)

where we can solve for h(x) and for which we also have that
∂r(x) ∂h(x)
= I + σ2 + o(σ 2 ) as σ 2 → 0.
∂x ∂x
We can substitute in the right-hand side of Equation 24 the value for rk (x) that we get
from Equation 24 itself. This substitution would be pointless in any other situation where
we are not trying to get a power series in terms of σ 2 around 0.

d 
∂ 2 rk (x)

2
X ∂ log p(x) ∂rk (x)
rk (x) = xk +σ +
i=1
∂xi
∂xi ∂x2i
  !
d d
X
 ∂ log p(x) ∂ xk + σ 2
X ∂ log p(x) ∂rk (x) ∂ 2 rk (x)
= xk +σ 2 + 
i=1
∂xi ∂xi
j=1
∂xj ∂xj ∂x2j
d
X ∂ 2 rk (x)
+σ 2
i=1
∂x2i
d d
X ∂ 2 rk (x)
X ∂ log p(x)
= xk +σ 2 I (i = k) + σ 2
i=1
∂xi
i=1
∂x2i
d X
d
!!
22
X ∂ log p(x) ∂ ∂ log p(x) ∂rk (x) ∂ 2 rk (x)
+σ +
i=1 j=1
∂xi ∂xi ∂xj ∂xj ∂x2j
d
X ∂ 2 rk (x)
2∂ log p(x) 2
rk (x) = xk +σ + σ2 2 + σ 2 ρ(σ 2 , x)
∂xk ∂xi
i=1
Pd ∂ 2 rk (x)
Now we would like to get rid of that σ 2 i=1 ∂x2i
term by showing that it is a
22
term that involves only powers of σ or higher. We get this by showing what we get by
differentiating the expression for rk (x) in line (25) twice with respect to some l.
d
!
∂rk (x) ∂ 2 log p(x) ∂ ∂ 2 r (x)
k
X
= I (i = l) + σ 2 + σ2 2 + σ 2 ρ(σ 2 , x)
∂xl ∂xl ∂xk ∂xl ∂x i
i=1
d
!
∂ 2 rk (x) 3
2 ∂ log p(x) 2 ∂
X ∂ 2 rk (x) 2 2
=σ +σ + σ ρ(σ , x)
∂x2l ∂x2l ∂xk ∂x2l i=1 ∂x2i
∂ 2 rk (x)
Since σ 2 is a common factor in all the terms of the expression of ∂x2l
we get what we
needed. That is,
∂ log p(x) 2
rk (x) = xk + σ 2 + σ 2 η(σ 2 , x).
∂xk

3765
 Alain and Bengio

This shows that

∂ log p(x)
r(x) = x + σ 2 + o(σ 2 ) as σ2 → 0
∂x
and
∂r(x) ∂ 2 log p(x)
= I + σ2 + o(σ 2 ) as σ2 → 0
∂x ∂x2
which completes the proof.

Appendix D. Local Mean

In preliminary work (Bengio et al., 2012a), we studied how the optimal reconstruction
could possibly estimate so-called local moments. We revisit this question here, with more
appealing and precise results.
What previous work on denoising and contractive auto-encoders suggest is that regu-
larized auto-encoders can capture the local structure of the density through the value of the
encoding (or reconstruction) function and its derivative. In particular, Rifai et al. (2012);
Bengio et al. (2012a) argue that the first and second derivatives tell us in which directions
it makes sense to randomly move while preserving or increasing the density, which may
be used to justify sampling procedures. This motivates us here to study so-called local
moments as captured by the auto-encoder, and in particular the local mean, following the
definitions introduced in Bengio et al. (2012a).

D.1 Definitions for Local Distributions

Let p be a continuous probability density function with support Rd . That is, ∀x ∈ Rd
we have that p(x) 6= 0. We define below the notion of a local ball Bδ (x0 ), along with an
associated local density, which is the normalized product of p with the indicator for the ball:
Bδ (x0 ) = {x s.t. kx − x0 k2 < δ}
Z
Zδ (x0 ) = p(x)dx
Bδ (x0 )
1
pδ (x|x0 ) = p(x)I (x ∈ Bδ (x0 ))
Zδ (x0 )
where Zδ (x0 ) is the normalizing constant required to make pδ (x|x0 ) a valid pdf for a
distribution centered on x0 . The support of pδ (x|x0 ) is the ball of radius δ around x0
denoted by Bδ (x0 ). We stick to the 2-norm in terms of defining the balls Bδ (x0 ) used, but
everything could be rewritten in terms of another p-norm to have slightly different formulas.
We use the following notation for what will be referred to as the first two local moments
(i.e., local mean and local covariance) of the random variable described by pδ (x|x0 ).
Z
def
mδ (x0 ) = xpδ (x|x0 )dx
Rd
Z
def
Cδ (x0 ) = (x − mδ (x0 ))(x − mδ (x0 ))T pδ (x|x0 )dx
Rd
Based on these definitions, one can prove the following theorem.

3766
What Regularized Auto-Encoders Learn from the Data-Generating Distribution

Theorem 3 Let p be of class C 3 and represent a probability density function. Let x0 ∈ Rd

with p(x0 ) 6= 0. Then we have that

1 ∂ log p(x)
mδ (x0 ) = x0 + δ 2 + o δ3 .


d+2 ∂x x0

This links the local mean of a density with the score associated with that density.
Combining this theorem with Theorem 2, we obtain that the optimal reconstruction function
r∗ (·) also estimates the local mean:

δ2
mδ (x) − x = (r∗ (x) − x) + A(δ) + δ 2 B(σ 2 ) (25)
σ 2 (d + 2)

for error terms A(δ), B(σ 2 ) such that

A(δ) ∈ o(δ 3 ) as δ → 0,
2
B(σ ) ∈ o(1) as σ 2 → 0.

This means that we can loosely estimate the direction to the local mean by the direction
of the reconstruction:
mδ (x) − x ∝ r∗ (x) − x. (26)

Appendix E. Asymptotic formulas for localized moments

Proposition 4 Let p be of class C 2 and let x0 ∈ Rd . Then we have that

π d/2
 
d 2 Tr(H(x0 )) 3
Zδ (x0 ) = δ p(x0 ) + δ + o(δ )
Γ (1 + d/2) 2(d + 2)
∂ 2 p(x)
where H(x0 ) = ∂x2 x=x
. Moreover, we have that
0
 
1 −d Γ (1 + d/2) 1 2 1 Tr(H(x0 )) 3
=δ −δ + o(δ ) .
Zδ (x0 ) π d/2 p(x0 ) p(x0 )2 2(d + 2)

Proof

Z "
∂p(x) 1
Zδ (x0 ) = p(x0 ) + (x − x0 ) + (x − x0 )T H(x0 )(x − x0 )
Bδ (x0 ) ∂x x0 2!

1 (3) 3
+ D p(x0 )(x − x0 ) + o(δ ) dx
3!
Z Z
1
= p(x0 ) dx + 0 + (x − x0 )T H(x0 )(x − x0 )dx + 0 + o(δ d+3 )
Bδ (x0 ) 2 Bδ (x0 )
π d/2 π d/2
= p(x0 )δ d + δ d+2 Tr (H(x0 )) + o(δ d+3 )
Γ (1 + d/2) 4Γ (2 + d/2)
π d/2
 
d 2 Tr(H(x0 )) 3
= δ p(x0 ) + δ + o(δ )
Γ (1 + d/2) 2(d + 2)

3767
 Alain and Bengio

We use Proposition 10 to get that trace come up from the integral involving H(x0 ). The
expression for 1/Zδ (x0 ) comes from the fact that, for any a, b > 0 we have that

a−1
 
1 1 b 2 3 4
= = 1 − ( δ + o(δ )) + o(δ )
a + bδ 2 + o(δ 3 ) 1 + ab δ 2 + o(δ 3 ) a a
1 b
= − δ 2 + o(δ 3 ) as δ → 0.
a a2
1
by using the classic result from geometric series where 1+r = 1 − r + r2 − . . . for |r| < 1.
Now we just apply this to

1 Γ (1 + d/2) 1
= δ −d d/2
h i
Zδ (x0 ) π p(x0 ) + δ 2 Tr(H(x 0 ))
+ o(δ 3)
2(d+2)

and get the expected result.

Theorem 5 Let p be of class C 3 and represent a probability density function. Let x0 ∈ Rd

with p(x0 ) 6= 0. Then we have that

1 ∂ log p(x)
mδ (x0 ) = x0 + δ 2 + o δ3 .


d+2 ∂x x0

Proof
The leading term in the expression for mδ (x0 ) is obtained by transforming the x in the
integral into a x − x0 to make the integral easier to integrate.
Z Z
1 1
mδ (x0 ) = xp(x)dx = x0 + (x − x0 )p(x)dx.
Zδ (x0 ) Bδ (x0 ) z δ 0)
(x Bδ (x0 )

Now using the Taylor expansion around x0

Z "
1 ∂p(x)
mδ (x0 ) = x0 + (x − x0 ) p(x0 ) + (x − x0 )
Zδ (x0 ) Bδ (x0 ) ∂x x0
#
1 ∂ 2 p(x)
+ (x − x0 )T (x − x0 ) + o(kx − x0 k2 ) dx.
2 ∂x2 x0

R
Remember that Bδ (x0 ) f (x)dx = 0 whenever we have a function f is anti-symmetrical
(or “odd”) relative to the point x0 (i.e., f (x − x0 ) = f (−x − x0 )). This applies to the terms
2 p(x)
(x − x0 )p(x0 ) and (x − x0 )(x − x0 ) ∂ ∂x 2 (x − x0 )T . Hence we use Proposition 9 to get
x=x0

3768
What Regularized Auto-Encoders Learn from the Data-Generating Distribution

Z " #
1 ∂p(x)
mδ (x0 ) = x0 + (x − x0 )T (x − x0 ) + o(kx − x0 k3 ) dx
Zδ (x0 ) Bδ (x0 ) ∂x x0
d
!
1 π 2 ∂p(x)
= x0 + δ d+2 d

+ o(δ 3 ).
Zδ (x0 ) 2Γ 2 + 2 ∂x x0

∂p(x)
Now, looking at the coefficient in front of ∂x x in the first term, we can use Propo-
0
sition 4 to rewrite it as

d
!
1 d+2 π2
δ
2Γ 2 + d2


Zδ (x0 )
  d
Γ (1 + d/2) 1 1 Tr(H(x0 )) π2
= δ −d − δ 2
+ o(δ 3
) δ d+2
π d/2 p(x0 )2 2(d + 2) 2Γ 2 + d2


p(x0 )
d 


2 Γ 1+ 2 1 2 1 Tr(H(x0 )) 1 1
= δ d

−δ 2
+ o(δ ) = δ 2
3
+ o(δ 3 ).
2Γ 2 + 2 p(x0 ) p(x0 ) 2(d + 2) p(x0 ) d + 2

Γ(1+ d )
There is no reason the keep the −δ 4 2Γ 2+2d p(x10 )2 Tr(H(x 0 ))
2(d+2) in the above expression be-
( 2)
cause the asymptotic error from the remainder term in the main expression is o(δ 3 ). That
would swallow our exact expression for δ 4 and make it useless.
We end up with

1 ∂ log p(x)
mδ (x0 ) = x0 + δ 2 + o(δ 3 ).
d+2 ∂x x0

Appendix F. Integration on balls and spheres

This result comes from Multi-dimensional Integration : Scary Calculus Problems from Tim
Reluga (who got the results from How to integrate a polynomial over a sphere by Gerald
B. Folland).
n Pd o
Theorem 6 Let B = x ∈ Rd x2 ≤1
j=1 j be the ball of radius 1 around the origin.
Then
Q  aj +1 
d
Z Y Γ 2
|xj |aj dx =
Γ 1 + 2 + 12 aj
d P

B j=1

for any real numbers aj ≥ 0.

3769
 Alain and Bengio

Corollary 7 Let B be the ball of radius 1 around the origin. Then

 Q  aj +1 
d
Z Y Γ 2
if all the aj are even integers

a
xj j dx = Γ(1+ d2 + 12 aj )
P

B j=1
0 otherwise


for any non-negative integers aj ≥ 0. Note the absence of the absolute values put on the
a
xj j terms.

Corollary 8 Let Bδ (0) ⊂ Rd be the ball of radius δ around the origin. Then
 Q  aj +1 
Z d  d+P aj Γ 2
Yaj δ Γ(1+ d2 + 12 aj )
if all the aj are even integers
xj dx =
P

Bδ (0) j=1
0 otherwise


a
for any non-negative integers aj ≥ 0. Note the absence of the absolute values on the xj j
terms.

Proof
We take the theorem as given and concentrate here on justifying the two corollaries.
Note how in Corollary 7 we dropped the absolute values that were in the original Q Theo-
a
rem 6. In situations where at least one aj is odd, we have that the function f (x) = dj=1 xj j
becomes odd in the sense that f (−x) = −f (x). Because of the symmetrical nature of the
integration on the unit ball, we get that the integral is 0 as a result of cancellations.
For Corollary 8, we can rewrite the integral by changing the domain with yj = xj /δ so
that
Z d Z d Z d
a
P Y Y Y
− aj aj
δ xj j dx = (xj /δ) dx = y aj δ d dy.
Bδ (0) j=1 Bδ (0) j=1 B1 (0) j=1

We pull out the δ d that we got from the determinant of the Jacobian when changing
from dx to dy and Corollary 8 follows.

Proposition 9 Let v ∈ Rd and let Bδ (0) ⊂ Rd be the ball of radius δ around the origin.
Then
d
Z !
d+2 π2
y < v, y > dy = δ v
2Γ 2 + d2


Bδ (0)

where < v, y > is the usual dot product.

Proof
We have that

3770
What Regularized Auto-Encoders Learn from the Data-Generating Distribution

v1 y12
 
 .. 
y < v, y > =  . 
vd yd2
which is decomposable into d component-wise applications

1 of 1Corollary

1 √ 8. This yields
3
the expected result with the constant obtained from Γ 2 = 2 Γ 2 = 2 π.

Proposition 10 Let H ∈ Rd×d and let Bδ (x0 ) ⊂ Rd be the ball of radius δ around x0 ∈ Rd .
Then

π d/2
Z
(x − x0 )T H(x − x0 )dx = δ d+2 trace (H) .
Bδ (x0 ) 2Γ (2 + d/2)

Proof
First, by substituting y = (x − x0 ) /δ we have that this is equivalent to showing that

π d/2
Z
y T Hydy = trace (H) .
B1 (0) 2Γ (2 + d/2)
This integral yields a real number which can be written as
Z Z X XZ
T
y Hydy = yi Hi,j yj dy = yi yj Hi,j dy.
B1 (0) B1 (0) i,j i,j B1 (0)

Now we know from Corollary 8 that this integral is zero when i 6= j. This gives

π d/2
X Z X Z
Hi,j yi yj dy = Hi,i yi2 dy = trace (H) .
B1 (0) B1 (0) 2Γ (2 + d/2)
i,j i

References
Y. Bengio. Learning deep architectures for AI. Foundations & Trends in Mach. Learn., 2
(1):1–127, 2009.
Y. Bengio, P. Lamblin, D. Popovici, and H. Larochelle. Greedy layer-wise training of deep
networks. In NIPS’2006, 2007.
Yoshua Bengio and Olivier Delalleau. On the expressive power of deep architectures. In
ALT’2011, 2011.
Yoshua Bengio, Guillaume Alain, and Salah Rifai. Implicit density estimation by local
moment matching to sample from auto-encoders. Technical report, arXiv:1207.0057,
2012a.

3771
 Alain and Bengio

Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review
and new perspectives. Technical report, arXiv:1206.5538, 2012b.

Yoshua Bengio, Yao Li, Guillaume Alain, and Pascal Vincent. Generalized denoising auto-
encoders as generative models. Technical Report arXiv:1305.6663, Universite de Montreal,
2013a.

Yoshua Bengio, Grégoire Mesnil, Yann Dauphin, and Salah Rifai. Better mixing via deep
representations. In ICML’13, 2013b.

Lawrence Cayton. Algorithms for manifold learning. Technical Report CS2008-0923, UCSD,
2005.

B. Dacorogna. Introduction to the Calculus of Variations. World Scientific Publishing

Company, 2004.

Karol Gregor, Arthur Szlam, and Yann LeCun. Structured sparse coding via lateral inhi-
bition. In NIPS’2011, 2011.

G. E. Hinton, S. Osindero, and Y.-W. Teh. A fast learning algorithm for deep belief nets.
Neural Computation, 18:1527–1554, 2006.

Aapo Hyvärinen. Estimation of non-normalized statistical models using score matching. J.

Machine Learning Res., 6, 2005.

Aapo Hyvärinen. Some extensions of score matching. Computational Statistics and Data
Analysis, 51:2499–2512, 2007.

Viren Jain and Sebastian H. Seung. Natural image denoising with convolutional networks.
In NIPS’2008, 2008.

K. Kavukcuoglu, M-A. Ranzato, R. Fergus, and Y. LeCun. Learning invariant features

through topographic filter maps. In CVPR’2009, 2009.

Diederik Kingma and Yann LeCun. Regularized estimation of image statistics by score
matching. In NIPS’2010, 2010.

Honglak Lee, Roger Grosse, Rajesh Ranganath, and Andrew Y. Ng. Convolutional deep
belief networks for scalable unsupervised learning of hierarchical representations. In
ICML’2009. 2009.

Hariharan Narayanan and Sanjoy Mitter. Sample complexity of testing the manifold hy-
pothesis. In NIPS’2010. 2010.

B. A. Olshausen and D. J. Field. Sparse coding with an overcomplete basis set: a strategy
employed by V1? Vision Research, 37:3311–3325, 1997.

M. Ranzato, C. Poultney, S. Chopra, and Y. LeCun. Efficient learning of sparse represen-

tations with an energy-based model. In NIPS’2006, 2007.

3772
What Regularized Auto-Encoders Learn from the Data-Generating Distribution

M. Ranzato, Y. Boureau, and Y. LeCun. Sparse feature learning for deep belief networks.
In NIPS’2007, 2008.

Salah Rifai, Yann Dauphin, Pascal Vincent, Yoshua Bengio, and Xavier Muller. The man-
ifold tangent classifier. In NIPS’2011, 2011a.

Salah Rifai, Pascal Vincent, Xavier Muller, Xavier Glorot, and Yoshua Bengio. Contractive
auto-encoders: Explicit invariance during feature extraction. In ICML’2011, 2011b.

Salah Rifai, Yoshua Bengio, Yann Dauphin, and Pascal Vincent. A generative process for
sampling contractive auto-encoders. In ICML’2012, 2012.

R. Salakhutdinov and G.E. Hinton. Deep Boltzmann machines. In AISTATS’2009, 2009.

Kevin Swersky, Marc’Aurelio Ranzato, David Buchman, Benjamin Marlin, and Nando
de Freitas. On autoencoders and score matching for energy based models. In ICML’2011.
2011.

Pascal Vincent. A connection between score matching and denoising autoencoders. Neural
Computation, 23(7), 2011.

Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting
and composing robust features with denoising autoencoders. In ICML 2008, 2008.

3773