Deep Compressed Sensing

Yan Wu 1 Mihaela Rosca 1 Timothy Lillicrap 1

Abstract training at all. CS has been successfully applied in scenarios

Compressed sensing (CS) provides an elegant where measurements are noisy and expensive to take, such
framework for recovering sparse signals from as in MRI (Lustig et al., 2007). Its sample efficiency enables
compressed measurements. For example, CS can the development of, for example, the “single pixel camera”,
exploit the structure of natural images and recover which reconstructs a full resolution image from a single
an image from only a few random measurements. light sensor (Duarte et al., 2008).
CS is flexible and data efficient, but its applica- However, the wide application of CS, especially in process-
tion has been restricted by the strong assumption ing large scale data where modern deep learning approaches
of sparsity and costly reconstruction process. A thrive, is hindered by its assumption of sparse signals and
recent approach that combines CS with neural the slow optimisation process for reconstruction. Recently,
network generators has removed the constraint of Bora et al. (2017) combined CS with separately trained neu-
sparsity, but reconstruction remains slow. Here ral network generators. Although these pre-trained neural
we propose a novel framework that significantly networks were not optimized for CS, they demonstrated re-
improves both the performance and speed of sig- construction performance superior to existing methods such
nal recovery by jointly training a generator and as the Lasso (Tibshirani, 1996). Here we propose the deep
the optimisation process for reconstruction via compressed sensing (DCS) framework in which neural net-
meta-learning. We explore training the measure- works can be trained from-scratch for both measuring and
ments with different objectives, and derive a fam- online reconstruction. We show that this framework leads
ily of models based on minimising measurement naturally to a family of models, including GANs (Good-
errors. We show that Generative Adversarial Nets fellow et al., 2014), which can be derived by training the
(GANs) can be viewed as a special case in this measurement functions with different objectives. In sum-
family of models. Borrowing insights from the mary, this work contributes the following:
CS perspective, we develop a novel way of im-
proving GANs using gradient information from • We demonstrate how to train deep neural networks
the discriminator. within the CS framework.

• We show that a meta-learned reconstruction process

1. Introduction leads to a more accurate and orders of magnitudes
faster method compared with previous models.
Encoding and decoding are central problems in communica-
tion (MacKay, 2003). Compressed sensing (CS) provides • We develop a new GAN training algorithm based on la-
a framework that separates encoding and decoding into in- tent optimisation, which improves GAN performance.
dependent measurement and reconstruction processes (Can- The non-saturated generator loss − ln (D(G(z)))
des et al., 2006; Donoho, 2006). Unlike commonly used emerges as a measurement error.
auto-encoding models (Bourlard & Kamp, 1988; Kingma
& Welling, 2013; Rezende et al., 2014), which feature end- • We extend our framework to training semi-supervised
to-end trained encoder and decoder pairs, CS reconstructs GANs, and show that latent optimisation results in
signals from low-dimensional measurements via online op- semantically meaningful latent spaces.
timisation. This model architecture is highly flexible and
sample efficient: high dimensional signals can be recon-
structed from a few random measurements with little or no Notations
1
DeepMind, London, UK. Correspondence to: Yan Wu We use bold letters for vectors and matrices and normal
<[email protected]>. letters for scalars. Ep(x) [f (x)] indicates taking the expec-
Proceedings of the 36 th International Conference on Machine tation of f (x) over the distribution p(x). We use subscrip-
Learning, Long Beach, California, PMLR 97, 2019. Copyright tions of Greek letters to indicate function parameters. For
2019 by the author(s). example, Gθ is a function parametrised by θ.
 Deep Compressed Sensing

Candes et al., 2006). This constrained optimisation problem

is computationally intensive — a price for the measuring
process that only requires sparse random projections of the
signals (Baraniuk, 2007).
Figure 1. Illustration of Deep Compressed Sensing. F is a mea-
surement process that produces a measurement of the signal, m 2.2. Compressed Sensing using Generative Models
and G is a generator that reconstructs the signal from a latent rep-
resentation ẑ. The latent representation is optimised to minimise a The requirement of sparsity poses a strong restriction on
measurement error Eθ (m, m̂). CS. Sparse bases, such as the Fourier basis or wavelet, only
partially relieve this constraints, since they are restricted to
domains known to be sparse in these bases and cannot adapt
2. Background to data distributions. Recently, Bora et al. (2017) proposed
compressed sensing using generative models (CSGM) to
2.1. Compressed Sensing relax this requirement. This model uses a pre-trained deep
Compressed sensing aims to recover signal x from a linear neural network Gθ (from a VAE or GAN) as the structural
measurement m: constraint in the place of sparsity. This generator maps a
latent representation z to the signal space:
m = Fx + η (1)
x = Gθ (z) (4)
where F is the C × D measurement matrix, and η is the
measurement noise which is usually assumed to be Gaus- Instead of requiring sparse signals, Gθ implicitly constrains
sian distributed. F is typically a “wide” matrix, such that output x in a low-dimensional manifold via its architecture
C  D. As a result, the measurement m has much lower and the weights adapted from data. This constraint is suf-
dimensionality compared with the original signal; solving x ficient to provide a generalised Set-Restricted Eigenvalue
is generally impossible for such under-determined problems. Condition (S-REC) with random matrices, under which low
The elegant CS theory shows that one can nearly perfectly reconstruction error can be achieved with high probability.
recover x with high probability given a random matrix F A minimisation process similar to that in CS is used for
and sparse x (Donoho, 2006; Candes et al., 2006). In prac- reconstruction:
tice, the requirement that x be sparse can be replaced by
sparsity in a set of basis Φ, such as the Fourier basis or ẑ = arg min Eθ (m, z) (5)
z
wavelet, so that Φ x can be non-sparse signals such as natu- 2
Eθ = km − F Gθ (z)k2 (6)
ral images. Here we omit the basis Φ for brevity; the linear
transform from Φ does not affect our following discussion. such that x̂ = Gθ (z) is the reconstructed signal. In con-
At the centre of CS theory is the Restricted Isometry Prop- trast to directly optimising the signal x in CS (eq.3), here
erty (RIP) 1 , which is defined for F and the difference be- optimisation is in the space of latent representation z.
tween signals x1 − x2 as The arg min operator in eq. 5 is intractable since Eθ is
2 2 highly non-convex. It is therefore approximated using
(1 − δ) kx1 − x2 k2 ≤ kF (x1 − x2 )k2 gradient descent starting from a randomly sampled point
2
(2)
≤ (1 + δ) kx1 − x2 k2 ẑ ∼ pz (z):

where δ ∈ (0, 1) is a small constant. The RIP states that ∂Eθ (m, z)
ẑ ← ẑ − α (7)
the projection from F preserves the distance between two ∂z z=ẑ
signals bounded by factors of 1 − δ and 1 + δ. This property
holds with high probability for various random matrices F where α is a learning rate. One can take a specified T steps
and sparse signals x. It guarantees minimising the measure- of gradient descent. Typically, hundreds or thousands of
ment error gradient descent steps and several re-starts from the initial
step are needed to obtain a sufficiently good ẑ (Bora et al.,
2
x̂ = arg min km − F xk2 (3) 2017; Bojanowski et al., 2018). This process is illustrated
x in Figure 1.
under the constraint that x is sparse, leads to accurate re- This work established the connection between compressed
construction x̂ ≈ x with high probability (Donoho, 2006; sensing and deep neural networks, and demonstrated perfor-
1 mance superior to the Lasso (Tibshirani, 1996), especially
The theory can also be proved from the closely related and
more general Restricted Eigenvalue condition (Bora et al., 2017). when the number of measurements is small. The theoreti-
We focus on RIP in this form for its more straightforward connec- cal properties of CSGM have been more closely examined
tion with the training loss (see section 3.1). by Hand & Voroninski (2017), who also proved stronger
 Deep Compressed Sensing

convergence guarantees. More recently, Dhar et al. (2018) modes, unless extra care is taken in designing and training
proposed additional constraints to allow sparse deviation the model (Radford et al., 2015; Salimans et al., 2016).
from the generative model’s support set, thus improving
A widely adapted trick is using − ln (D(G(z))) as the ob-
generalisation. However, CSGM still suffers from two re-
jective for the generator (Goodfellow et al., 2014). Com-
strictions:
pared with eq. 10, this alternative objective avoids saturat-
ing the discriminator in the early stage of training when
1. The optimisation for reconstruction is still slow, as it the generator is too weak. However, this objective voids
requires thousands of gradient descent steps. most theoretical analyses (Hu et al., 2018), since the new
adversarial objective is no longer a zero-sum game (eq. 10).
2. It relies on random measurement matrices, which are
known to be sub-optimal for highly structured signals In most GAN models, discriminators become useless after
such natural images. Learned measurements can per- training. Recently, Tao et al. (2018) and Azadi et al. (2019)
form significantly better (Weiss et al., 2007). proposed methods using the discriminator for importance
sampling. Our work provides an alternative: our model
2.3. Model-Agnostic Meta Learning moves latent representations to areas more likely to generate
realistic images as deemed by the discriminator.
Meta-learning, or learning to learn, allows a model adapt-
ing to new tasks by self-improving (Schmidhuber, 1987).
Model-Agnostic Meta learning (MAML) provides a general 3. Deep Compressed Sensing
method to adapt parameters for a number of tasks (Finn We start by showing the benefit of combining meta-learning
et al., 2017). Given a differentiable loss function L(Ti ; θ) with the model in Bora et al. (2017). We then generalise
for task Ti sampled from the task distribution ptask (T ), the measurement matrices to parametrised measurement func-
task-specific parameters are adapted by gradient descent tions, including deep neural networks. While previous work
from the initial parameters θ: relies on random projections as measurement functions, our
approach learns measurement functions by imposing the RIP
θi ← θ − α∇θ L(Ti ; θ) (8) as a training objective. We then derive two novel models by
imposing properties other than the RIP on the measurements,
The initial parameters θ are trained to minimise the loss
including a GAN model with discriminator-guided latent
across all tasks
optimisation, which leads to more stable training dynamics
min ETi ∼ptask (T ) [L(Ti ; θi )] (9) and better results.
θ

Multiple steps and more sophisticated optimisation algo- 3.1. Compressed Sensing with Meta-Learning
rithms can be used in the place of eq. 8. Despite L usually We hypothesise that the run-time efficiency and performance
being a highly non-convex function, by back-propagating in CSGM (Bora et al. 2017, section 2.2), can be improved
through the gradient-descent process, only a few gradient by training the latent optimisation procedure using meta-
steps are sufficient to adapt to new tasks. learning, by back-propagating through the gradient descent
steps (Finn et al., 2017). The latent optimisation procedure
2.4. Generative Adversarial Networks for CS models can take hundreds or thousands of steps.
By employing meta-learning to optimise this optimisation
A Generative Adversarial Network (GAN) trains a
procedure we aim to achieve similar results with far fewer
parametrised generator Gθ to fool a discriminator Dφ that
updates.
tries to distinguish real data from fake data sampled from
the generator (Goodfellow et al., 2014). The generator Gθ To this end, the model parameters, as well as the latent
is a deterministic function that transforms samples z from optimisation procedure, are trained to minimise the expected
a source pz (z) to the same space as the data x, which has measurement error:
the distribution pdata (x). This adversarial game can be sum-
marised by the following min-max problem with the value min LG , for LG = Exi ∼pdata (x) [Eθ (mi , ẑi )] (11)
θ
function V (Gθ , Dφ ):
where ẑi is obtained from gradient descent (eq. 7). The
min maxV (Gθ , Dφ ) = Ex∼pdata (x) [ln Dφ (x)] gradient descent in eq. 7 and the loss function in eq. 11
Gθ Dφ
(10)
mirror their counterparts in MAML (eq. 8 and 9), except
+ Ez∼pz (z) [ln(1 − Dφ (Gθ (z)))]
that:
GANs are usually difficult to train due to this adversarial
game (Balduzzi et al., 2018). Training may either diverge 1. Instead of the stochastic gradient computed in the out-
or converge to bad equilibrium with, for example, collapsed side loop, here each measurement error Eθ only de-
 Deep Compressed Sensing

pends on a single sample z, so eq. 7 computes the exact Algorithm 1 Compressed Sensing with Meta Learning
gradient of Eθ . Input: minibatchs of data {xi }N i=1 , random matrix F,
2. The online optimisation is over latent variables rather generator Gθ , learning rate α, number of latent optimisa-
than parameters. There are usually much fewer latent tion steps T
variables than parameters, so the update is quicker. repeat
Initialize generator parameters θ
Like in MAML, we implicitly perform second order optimi- for i = 1 to N do
sation, by back-propagating through the latent optimisation Measure the signal mi ← F xi
steps which compute ẑi when optimising eq. 11. We empiri- Sample ẑi ∼ pz (z)
cally observed that this dramatically improves the efficiency for t = 1 to T do
∂
of latent optimisation, with only 3-5 gradient descent steps Optimise ẑi ← ẑi − ∂z Eθ (mi , ẑi )
being sufficient to improve upon baseline methods. end for
end for P
Unlike Bora et al. (2017), we also train the generator Gθ . N
LG = N1 i=1 Eθ (mi , ẑi )
Merely minimising eq. 9 would fail — the generator can Compute LF using eq. 12
exploit F by mapping all Gθ (z) into the null space of F. Update θ ← θ − ∂θ ∂
(LG + LF )
This trivial solution always gives zero measurement error, until reaches the maximum training steps
but may contain no useful information. Our solution is
to enforce the RIP (eq. 2) via training, by minimising the
measurement loss:
h i The distance preserving property as a counterpart of the RIP
2 can be enforced by minimising a loss similar to eq. 12:
LF = Ex1 ,x2 (kF (x1 − x2 )k2 − kx1 − x2 k2 ) (12)
h
2 i
x1 and x2 can be sampled in various ways. While the choice LF = Ex1 ,x2 kFφ (x1 − x2 )k2 − kx1 − x2 k2
is not unique, it is important to sample from both the data (14)
distribution pdata (x) and generated samples Gθ (z), so that
the trained RIP holds for both real and generated data. In Minimising LF provides a relaxation of the constraint spec-
our experiments, we randomly sampled one image from the ified by the RIP (eq. 2). When LF is small, the projection
data and two generated images at the beginning and end of from F better preserves the distance between x1 and x2 .
latent optimisation, then computed the average between the This relaxation enables us to transform the RIP into a train-
3 pairs of losses between these 3 points as a form of “triplet ing objective for the measurements, which can then be inte-
loss”. grated into training other model components. Empirically,
we found this relaxation leads to high quality reconstruction.
Our algorithm is summarised in Algorithm 1. Since Al-
gorithm 1 still uses a random measurement matrix F, it The rest of the algorithm is identical to Algorithm 1, except
can be used as any other CS algorithm when ground truth that we also update the measurement function’s parameters
reconstructions are available for training the generator. φ. Consequently, different schemes can be employed to
coordinate updating θ and φ, which will be discussed more
3.2. Deep Compressed Sensing with Learned in section 3.3. This extended algorithm is summarised in
Measurement Function Algorithm 2. We call it Deep Compressed Sensing (DCS)
to emphasise that both the measurement and reconstruction
In Algorithm 1, we use the RIP property to train the gen- can be deep neural networks. Next, we turn to generalising
erator. We can use the same approach and enforce the RIP the measurements to properties other than the RIP.
property to learn the measurement function F itself, rather
than using a random projection. 3.2.2. G ENERALISED CS 1: CS-GAN

3.2.1. L EARNING M EASUREMENT F UNCTION Here we consider an extreme case: a one-dimensional mea-
surement that only encodes how likely an input is a real data
We start by generalising the measurement matrix F (eq. 1), point or fake one sampled from the generator. One way to
and define a parametrised measurement function m ← formulate this is to train the measurement function Fφ using
Fφ (x). The model introduced in the previous section cor- the following loss instead of eq. 14:
responds to a linear function Fφ (x) = F x; now both Fφ
and Gθ can be deep neural networks. Similar to CS, the
(
2
kFφ (x) − 1k2 x ∼ pdata (x)
central problem in this generalised setting is inverting the LF = 2 (15)
measurement function to recover the signal x ← Fφ−1 (m) kFφ (x̂)k2 x̂ ∼ Gθ (ẑ), ∀ẑ
via minimising the measurement error similar to eq. 6:
Algorithm 2 then becomes the Least Squares Generative
2
Eθ (m, z) = km − Fφ (Gθ (z))k2 (13) Adversarial Nets (LSGAN, Mao et al., 2017) with latent
 Deep Compressed Sensing

Algorithm 2 Deep Compressed Sensing ment function is achieved by the perfect classifier:
Input: minibatchs of data {xi }N i=1 , measurement func-
tion Fφ , generator Gθ , learning rate α, number of latent (
1 x ∼ pdata (x)
optimisation steps T Dφ (x) = (19)
repeat 0 x ∼ Gθ (z), ∀z
Initialize generator parameters θ
for i = 1 to N do
Measure the signal mi ← Fφ (xi ) We can therefore simplify eq. 18 by replacing m with its
Sample ẑi ∼ pz (z) target value 1 as in teacher-forcing (Williams & Zipser,
for t = 1 to T do 1989):
∂
Optimise ẑi ← ẑi − ∂z Eθ (mi , ẑi )
end for E(m, z) = ln [Dφ (Gθ (z))] (20)
end for P
N
LG = N1 i=1 Eθ (mi , ẑi )
Compute LF using eq. 12 This objective recovers the vanilla GAN formulation with
∂
Option 1 : joint update θ ← θ − ∂θ (LG + LF ) the commonly used alternative loss (Goodfellow et al.,
Option 2 : alternating update 2014), which we derived as a measurement error. When
∂ ∂
θ ← θ − ∂θ LG φ ← φ − ∂φ LF latent optimisation is disabled (T = 0), Algorithm 2 is
until reaches the maximum training steps identical to a vanilla GAN.
In our experiments (section 4.2), we observed that the addi-
tional latent optimisation steps introduced from the CS per-
optimisation — they are exactly equivalent when the la- spective significantly improved GAN training. We reckon
tent optimisation is disabled (T = 0, zero step). LSGAN this is because latent optimisation moves the representa-
is an alternative to the original GAN (Goodfellow et al., tion to areas more likely to generate realistic images as
2014) that can be motivated from Pearson χ2 Divergence. deemed by the discriminator. Since the gradient descent
To demonstrate a closer connection with original GANs process remains local, the latent representations are still
(Goodfellow et al., 2014), we instead focus on another for- spread broadly in latent space, which avoids mode collapse.
mulation whose measurement function is a binary classifier Although a sufficiently powerful generator Gθ can trans-
(the discriminator). form the source pz (z) into arbitrarily complex distribution,
This is realised by using a binary classifier Dφ as the mea- a more informative source, as implicitly manifested from
surement function, where we can interpret Dφ (x) as the the optimised z, may significantly reduce the complexity
probability that x comes from the dataset. In this case, the required for Gθ , thus striking a better trade-off in terms of
measurement function is equivalent to the discriminator in the overall computation.
GANs. Consequently, we change the the squared-loss in
eq. 13 to the cross-entropy loss as the matching measure- 3.2.3. G ENERALISED CS 2: S EMI - SUPERVISED GAN S
ment loss function (Bishop, 2006) (ignoring the expectation So far, we have shown two extreme cases of Deep Com-
over x for brevity): pressed Sensing: in one case, the distance preserving mea-
surements (section 3.2.1) essentially encode all information
LF = t(x) ln [Dφ (x)] + (1 − t(x)) ln [1 − Dφ (x)] (16)
for recovering the original signals; on the other hand, the
CS-GAN (section 3.2.2) has one-dimensional measurements
where the binary scalar t is an indicator function identifies
that only indicates whether signals are real or fake. We now
whether x is a real data point.
seek a middle ground, by using measurements that preserve
( class information for labelled data.
1 x ∼ pdata (x)
t(x) = (17) We generalise CS-GAN by replacing the binary classifier
0 x ∼ Gθ (z), ∀z
(discriminator) Dφ with a multi-class classifier Cφ . For data
Similarly, a cross-entropy measurement error is employed with K classes, this classifier outputs K + 1 classes with
to quantify the discrepancy between Dφ (Gθ (z)) and the the (K + 1)’th class reserved for “fake” data that comes
scalar measurement m = Dφ (x): from the generator. This specification is the same as the
classifier used in semi-supervised GANs (SGANs, Salimans
Eθ (m, z) = m ln [Dφ (Gθ (z))] et al. (2016)). Consequently, we extend the binary indicator
(18) function in eq. 17 to multi-class indicator, so that its k’the
+ (1 − m) ln [1 − Dφ (Gθ (z))] element tk (x) = 1 when x in class k. The k’th output of
the classifier Cφk (x) indicates the predicted probability that
At the minimum of LF = 0 (eq. 16), the optimal measure- x is in the k’th class, and multi-class cross-entropy loss is
 Deep Compressed Sensing

as an optimisation cost and add it to LG :

Table 1. A family of DCS models differentiated by the properties
of measurements in comparison with CS. The CS measurement 2
matrix does not need training, so it does not have a training loss. LO = β · kẑ − z0 k2 (23)

M ODEL P ROPERTY L OSS where β is a scalar controlling the strength of this regu-
CS RIP N/A
lariser. This regulariser encourages small moves of z in
DCS TRAINED RIP EQ .
14 optimisation, and can be interpreted as approximating an
CS-GAN VALIDITY PRESERVING EQ .
16 optimal transport cost (Villani, 2008). We found a range
CS-SGAN CLASS PRESERVING EQ .
21 of β from 1.0 to 10.0 made little difference in training, and
used β = 3.0 in our experiments with CS-GAN.

used for the measurement loss and measurement error: 4. Experiments

K+1
X 4.1. Deep Compressed Sensing for Reconstruction
tk (x) ln Cφk (x)
 
LF = (21)
k=1
We first evaluate the DCS model using the MNIST (Yann
et al., 1998) and CelebA (Liu et al., 2015) datasets. To
E(m, z) = t (x) ln Cφk (Gθ (z))
k
 
(22) compare with the approach in Bora et al. (2017), we used
the same generators as in their model. For the measure-
When latent optimisation is disabled (T = 0), the model ments functions, we considered both linear projection and
is similar to other semi-supervised GANs (Salimans et al., neural networks. We considered both random projections
2016; Odena et al., 2017). However, when T > 0 the online and trained measurement functions, while the generator
optimisation moves latent representations towards regions was always trained jointly with the latent optimisation pro-
representing particular classes. This provides a novel way cess. Unless otherwise specified, we use 3 gradient de-
of training conditional GANs. scent steps for latent optimisation. More details, includ-
Compared with conditional GANs which concatenate labels ing hyperparameter values, are reported in the Appendix.
to latent variables (Mirza & Osindero, 2014), optimising la- Our code will be available at https://github.com/
tent variables is more adaptive and uses information from the deepmind/deep-compressed-sensing.
entire model. Compared with Batch-Norm based methods Tables 2 and 3 summarise the results from our models as
(Miyato & Koyama, 2018), the information for conditioning well as the baseline model from Bora et al. (2017). The
is presented in the target measurements, and does not need reconstruction loss for the baseline model is estimated from
to be trained as Batch-Norm statistics (Ioffe & Szegedy, Figure 1 in Bora et al. (2017). DCS performs significantly
2015). Since both of these methods use separate sources better than the baseline. In addition, while the baseline
(label inputs or batch statistics) to provide the condition, model used hundreds or thousands of gradient-descent steps
their latent variables tend to retain no information about with several re-starts, we only used 3 steps without any re-
the condition. Our model, on the other hand, distils the starting, achieving orders of magnitudes higher efficiency.
condition information into the latent representation, which Interestingly, for fixed F , random linear projections outper-
results in semantically meaningful latent space (Figure 5). formed neural networks as the measurement functions in
both datasets across different neural network structures (row
3.3. Optimising Models 2 and 3 of Table 2 and 3). This empirical result is consistent
with the optimality of random projections described in the
The three models we derived as examples in the DCS frame-
compressed sensing literature and the more general Johnson-
work are summarised in Table 1 along side CS. The main
Lindenstrauss lemma (Donoho, 2006; Candes et al., 2006;
difference between them lies is the training objective used
Johnson & Lindenstrauss, 1984).
for the measurement functions LF . Once LF is specified,
the generator objective LG , in the form of measurement The advantage of neural networks manifested when Fφ was
error, can be derived follow suit. When LF and LG are optimised; this variant reached the best performance in all
adversarial, such as in the CS-GAN, Fφ and Gθ need to be scenarios. As argued in (Weiss et al., 2007), we observed
optimised separately as in GANs. This is implemented as that random projections are sub-optimal for highly struc-
the alternating update option in Algorithm 2. In optimis- tured signals such as images, as seen in the improved per-
ing the latent variables (eq. 7), we normalise ẑ after each formance when optimising the measurement matrices (row
gradient descent step, as in (Bojanowski et al., 2018). We 4 of Table 2 and 3). The reconstruction performance was
treat the step size α in latent optimisation as a parameter further improve when the linear measurement projections
and back-propagate through it in optimising the model loss were replaced by neural networks (row 5 of Table 2 and 3).
function. An additional technique we found useful in stabil- Examples of reconstructed MNIST images from different
ising CS-GAN training is to penalise the distance z moves models are shown in Figure 2.
 Deep Compressed Sensing

Table 2. Reconstruction loss on MNIST test data using different

measurement functions. All rows except the first are from our
models. “±” shows the standard deviation across test samples. (L)
indicates learned measurement functions. Lower is better.

M ODEL 10 25 MEASUREMENTS STEPS

BASELINE 54.8 17.2 10 × 1000

L INEAR 10.8 ± 3.8 6.9 ± 2.7 3 Figure 3. Samples from CS-GANs using 0 (left), 3 (central) and 5
NN 12.5 ± 2.2 10.2 ± 1.7 3 (right) gradient descent steps in latent optimisation. The CS-GAN
L INEAR (L) 6.5 ± 2.1 4. ± 1.4 3 using 0 step was equivalent to a vanilla GAN.
NN(L) 5.3 ± 1.9 3.4 ± 1.2 3

Table 3. Reconstruction loss on CelebA test data using different

vantage of latent optimisation. For quantitative evaluation,
measurement functions. All rows except the first are from our we trained larger and more standard models on CIFAR10
models. “±” shows the standard deviation across test samples. (L) (Krizhevsky & Hinton, 2009), and evaluate them using the
indicates learned measurement functions. Lower is better. Inception Score (IS) (Salimans et al., 2016) and Frchet In-
ception Distance (FID) (Heusel et al., 2017). To our knowl-
edge, latent optimisation has not been previously used to
M ODEL 20 50 MEASUREMENTS STEPS improving GANs, so our approach is orthogonal to existing
BASELINE 156.8 82.3 2 × 500
methods such as Arjovsky et al. (2017); Miyato et al. (2018).
L INEAR 34.7 ± 7.9 27.1 ± 6.1 3 We first compare our model with vanilla GANs, which is a
NN 46.1 ± 8.9 41.2 ± 8.3 3 special case of the CS-GAN (see section 3.2.2).
L INEAR (L) 26.2 ± 5.9 20.5 ± 4.3 3
NN(L) 23.4 ± 5.8 18.5 ± 4.3 3 We use the same MNIST architectures as in section 4.1, but
changed the the measurement function to a GAN discrimina-
tor (section 3.2.2). We use the alternating update option in
Unlike autoencoder-based methods, our models were not Algorithm 2 in this setting. All other hyper-parameters are
trained with any pixel reconstruction loss, which we only the same as in previous experiments. We use this relatively
use for testing. Despite this, our results are comparable with weak model to reveal failure modes as well as advantages
the recently proposed “Uncertainty Autoencoders” (Grover of the CS-GAN. Figure 3 shows samples from models with
& Ermon, 2018). We have worse MNIST reconstructions: the same setting but different latent optimisation iterations.
with 10 and 25 measurements, ours best model achieved The three panels show samples from models using 0, 3 and
5.3 and 3.4 per-image reconstruction errors compared with 5 gradient descent steps respectively. The model using 0
theirs 3.8 and 2.5 (estimated from figure 4). However, we iteration was equivalent to a vanilla GAN. Optimising latent
achieved better CelebA results: with 20 and 50 measure- variables exhibits no mode collapse, one of the common
ments, we have errors of 23.4 and 18.5 compared with their failure modes of GAN training.
27 and 22 (estimated from Figure 6). To confirm this advantage, we more systematically evaluate
our method across a range of 144 hyper-parameters (similar
4.2. CS-GANs to Kurach et al. (2018)). We use the CIFAR dataset which
contains various categories of natural images, whose fea-
To evaluate our proposed CS-GANs, we first trained a tures from an Inception Network (Ioffe & Szegedy, 2015)
small model on MNIST to demonstrate intuitively the ad- are meaningful for evaluating the IS and FID. Other than
the number of gradient descent steps (0 vs. 3) the model
architectures and training procedures were identical. The
change of IS and FID during training are plot in figure 4.
CS-GANs achieved better performance in both IS and FID,
and had less variance across the range of hyper-parameters.
The blue horizontal lines at the bottom of Fig. 4 (left) and
the top of Fig. 4 (right) shows failed vanilla GANs, but none
of the CS-GANs diverged in training.
We also applied our latent optimisation method on Spectral-
Figure 2. Reconstructions using 10 measurements from random Normalised GANs (SN-GANs) (Miyato et al., 2018), which
linear projection (top), trained linear projection (middle), and use Batch Normalisation (Ioffe & Szegedy, 2015) for the
trained neural network (bottom). generator and Spectral Normalisation for the discriminator.
 Deep Compressed Sensing

Table 4. Comparison with Spectral Normalised GANs. 0

1
2
3
4

SN-GAN SN-GAN ( OURS ) CS+SN-GAN 5

6
7

IS 7.42 ± 0.08 7.34 ± 0.07 7.80 ± 0.05 8

9

FID 29.3 29.53 ± 0.36 23.13 ± 0.50

Figure 5. Left: samples from the generative classifier, and t-SNE

illustration of the generator’s latent space.

our approach significantly improves upon the performance

and speed of reconstruction obtain in this work. In addi-
Figure 4. Inception Score (higher is better) and FID (lower is bet- tion, we derived a family of models, including a novel GAN
ter) during CIFAR training. model, by expanding the set of properties we consider for
the measurement function (Table 1).
Our method differs from existing algorithms that aim to
We compared our model with SN-GAN in Table 4: the combine compressed sensing with deep networks in that
SN-GAN column reproduces the numbers from (Miyato our approach preserves the online minimisation of measure-
et al., 2018), and the next column are numbers from our ment errors in generic neural networks. Previous attempts
replication of the same baseline. Our results demonstrate that combine CS and deep learning generally fall into two
that deeper architectures, Batch Normalisation and Spectral categories. One category of methods interprets taking com-
Normalisation can further improve CS-GAN and that CS- pressed measurements and reconstructing from these mea-
GAN can improve upon a competitive baseline, SN-GAN. surements as an encoding-decoding problem and formulate
the model as an autoencoder (Mousavi et al., 2015; Kulkarni
et al., 2016; Mousavi et al., 2017; Grover & Ermon, 2018;
4.3. CS-SGANs Mousavi et al., 2018; Lu et al., 2018). Another category
of methods are designed to mimic principled iterative CS
We now experimentally assess our approach to use latent
algorithms using specialised network architectures (Metzler
optimisation in semi-supervised GANs, CS-SGAN. We il-
et al., 2017; Sun et al., 2016). In contrast, our framework
lustrate this extension with the MNIST dataset, and leave it
maintains the separation of measurements from generation
to future work to study other applications. We keep all the
but still uses generic neural networks. Therefore, both the
hyper-parameters the same as in section 4.2, except chang-
measurements and latent representation of the generator can
ing the number of measurements to 11 for the 10 MNIST
be flexibly optimised for different, even adversarial, objec-
classes and 1 class reserved for generated samples. Samples
tives, while taking advantage of powerful neural network
from CS-SGAN can be seen in Figure 5 (left). Figure 5
architectures.
(right) illustrates this with T-SNE (Maaten & Hinton, 2008)
computed from 2000 random samples, where class labels Moreover, we can train measurement functions with proper-
are colour-coded. The latent space formed separated regions ties that are difficult or impossible to obtain from random or
representing different digits. It is impossible to obtain such hand-crafted projections, thus broadening the range of prob-
clustered latent space in typical conditional GANs (Mirza & lems that can be solved by minimising error measurements
Osindero, 2014; Miyato & Koyama, 2018), where labels are online. In other words, learning the measurement can be
supplied as separate inputs while the random source only used as a useful stepping stone for learning complex tasks
provides label-independent variations. In contrast, in our where the cost function is difficult to design directly. Our ap-
model the labels are distilled into latent representation via proach can also be interpreted as training implicit generative
optimisation, leading to a more interpretable latent space. models, where explicit minimisation of divergences is re-
placed by statistical tests (Mohamed & Lakshminarayanan,
5. Discussion 2016). We have illustrated this idea in the context of rela-
tively simple tasks, but anticipate that complex tasks such
We present a novel framework for combining compressed as style transfer (Zhu et al., 2017), in areas already seen
sensing and deep neural networks. In this framework we the applications of CS, as well as applications including
trained both the measurement and generation functions, as MRI (Lustig et al., 2007) and unsupervised anomaly de-
well as the latent optimisation (i.e., reconstruction) proce- tection (Schlegl et al., 2017), may further benefit from our
dure itself via meta-learning. Inspired by Bora et al. (2017), approach.
 Deep Compressed Sensing

ACKNOWLEDGMENTS Duarte, M. F., Davenport, M. A., Takhar, D., Laska, J. N.,

Sun, T., Kelly, K. F., and Baraniuk, R. G. Single-pixel
We thank Shakir Mohamed and Jonathan Hunt for insight- imaging via compressive sampling. IEEE signal process-
ful discussions. We also appreciate the feedback from the ing magazine, 25(2):83–91, 2008.
anonymous reviewers.
Finn, C., Abbeel, P., and Levine, S. Model-agnostic meta-
References learning for fast adaptation of deep networks. arXiv
preprint arXiv:1703.03400, 2017.
Arjovsky, M., Chintala, S., and Bottou, L. Wasserstein gan.
arXiv preprint arXiv:1701.07875, 2017. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B.,
Warde-Farley, D., Ozair, S., Courville, A., and Bengio,
Azadi, S., Olsson, C., Darrell, T., Goodfellow, I., and Y. Generative adversarial nets. In Advances in neural
Odena, A. Discriminator rejection sampling. In In- information processing systems, pp. 2672–2680, 2014.
ternational Conference on Learning Representations,
2019. URL https://openreview.net/forum? Grover, A. and Ermon, S. Uncertainty autoencoders: Learn-
id=S1GkToR5tm. ing compressed representations via variational informa-
tion maximization. arXiv preprint arXiv:1812.10539,
Balduzzi, D., Racaniere, S., Martens, J., Foerster, J., Tuyls, 2018.
K., and Graepel, T. The mechanics of n-player differen-
tiable games. arXiv preprint arXiv:1802.05642, 2018. Hand, P. and Voroninski, V. Global guarantees for enforcing
deep generative priors by empirical risk. arXiv preprint
Baraniuk, R. G. Compressive sensing [lecture notes]. IEEE arXiv:1705.07576, 2017.
signal processing magazine, 24(4):118–121, 2007.
Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., and
Bishop, C. M. Pattern recognition and machine learning Hochreiter, S. Gans trained by a two time-scale update
(information science and statistics) springer-verlag new rule converge to a local nash equilibrium. In Advances in
york. Inc. Secaucus, NJ, USA, 2006. Neural Information Processing Systems, pp. 6626–6637,
2017.
Bojanowski, P., Joulin, A., Lopez-Pas, D., and Szlam, A.
Optimizing the latent space of generative networks. In Hu, Z., Yang, Z., Salakhutdinov, R., and Xing, E. P. On
Dy, J. and Krause, A. (eds.), Proceedings of the 35th In- unifying deep generative models. In International Confer-
ternational Conference on Machine Learning, volume 80 ence on Learning Representations, 2018. URL https:
of Proceedings of Machine Learning Research, pp. 600– //openreview.net/forum?id=rylSzl-R-.
609, Stockholmsmssan, Stockholm Sweden, 10–15 Jul
Ioffe, S. and Szegedy, C. Batch normalization: Accelerating
2018. PMLR. URL http://proceedings.mlr.
deep network training by reducing internal covariate shift.
press/v80/bojanowski18a.html.
arXiv preprint arXiv:1502.03167, 2015.
Bora, A., Jalal, A., Price, E., and Dimakis, A. G. Com- Johnson, W. B. and Lindenstrauss, J. Extensions of lipschitz
pressed sensing using generative models. arXiv preprint mappings into a hilbert space. Contemporary mathemat-
arXiv:1703.03208, 2017. ics, 26(189-206):1, 1984.
Bourlard, H. and Kamp, Y. Auto-association by multilayer Kingma, D. P. and Welling, M. Auto-encoding variational
perceptrons and singular value decomposition. Biological bayes. arXiv preprint arXiv:1312.6114, 2013.
cybernetics, 59(4-5):291–294, 1988.
Krizhevsky, A. and Hinton, G. Learning multiple layers
Candes, E. J., Romberg, J. K., and Tao, T. Stable signal of features from tiny images. Technical report, Citeseer,
recovery from incomplete and inaccurate measurements. 2009.
Communications on Pure and Applied Mathematics: A
Journal Issued by the Courant Institute of Mathematical Kulkarni, K., Lohit, S., Turaga, P., Kerviche, R., and Ashok,
Sciences, 59(8):1207–1223, 2006. A. Reconnet: Non-iterative reconstruction of images from
compressively sensed measurements. In Proceedings of
Dhar, M., Grover, A., and Ermon, S. Modeling sparse devi- the IEEE Conference on Computer Vision and Pattern
ations for compressed sensing using generative models. Recognition, pp. 449–458, 2016.
In International Conference on Machine Learning, pp.
1222–1231, 2018. Kurach, K., Lucic, M., Zhai, X., Michalski, M., and Gelly, S.
The gan landscape: Losses, architectures, regularization,
Donoho, D. L. Compressed sensing. IEEE Transactions on and normalization. arXiv preprint arXiv:1807.04720,
information theory, 52(4):1289–1306, 2006. 2018.
 Deep Compressed Sensing

Liu, Z., Luo, P., Wang, X., and Tang, X. Deep learning face Odena, A., Olah, C., and Shlens, J. Conditional image
attributes in the wild. In Proceedings of International synthesis with auxiliary classifier gans. In Proceedings of
Conference on Computer Vision (ICCV), 2015. the 34th International Conference on Machine Learning-
Volume 70, pp. 2642–2651. JMLR. org, 2017.
Lu, X., Dong, W., Wang, P., Shi, G., and Xie, X. Convcsnet:
A convolutional compressive sensing framework based on Radford, A., Metz, L., and Chintala, S. Unsupervised rep-
deep learning. arXiv preprint arXiv:1801.10342, 2018. resentation learning with deep convolutional generative
adversarial networks. arXiv preprint arXiv:1511.06434,
Lustig, M., Donoho, D., and Pauly, J. M. Sparse mri: The 2015.
application of compressed sensing for rapid mr imaging.
Magnetic Resonance in Medicine: An Official Journal Rezende, D. J., Mohamed, S., and Wierstra, D. Stochastic
of the International Society for Magnetic Resonance in backpropagation and approximate inference in deep gen-
Medicine, 58(6):1182–1195, 2007. erative models. arXiv preprint arXiv:1401.4082, 2014.

Maaten, L. v. d. and Hinton, G. Visualizing data using Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Rad-
t-sne. Journal of machine learning research, 9(Nov): ford, A., and Chen, X. Improved techniques for training
2579–2605, 2008. gans. In Advances in Neural Information Processing
Systems, pp. 2234–2242, 2016.
MacKay, D. J. Information theory, inference and learning
Schlegl, T., Seeböck, P., Waldstein, S. M., Schmidt-Erfurth,
algorithms. Cambridge university press, 2003.
U., and Langs, G. Unsupervised anomaly detection with
Mao, X., Li, Q., Xie, H., Lau, R. Y., Wang, Z., and generative adversarial networks to guide marker discov-
Paul Smolley, S. Least squares generative adversarial ery. In International Conference on Information Process-
networks. In Proceedings of the IEEE International Con- ing in Medical Imaging, pp. 146–157. Springer, 2017.
ference on Computer Vision, pp. 2794–2802, 2017.
Schmidhuber, J. Evolutionary principles in self-referential
Metzler, C., Mousavi, A., and Baraniuk, R. Learned d- learning, or on learning how to learn: the meta-meta-
amp: Principled neural network based compressive image ... hook. PhD thesis, Technische Universität München,
recovery. In Advances in Neural Information Processing 1987.
Systems, pp. 1772–1783, 2017. Sun, J., Li, H., Xu, Z., et al. Deep admm-net for com-
Mirza, M. and Osindero, S. Conditional generative adver- pressive sensing mri. In Advances in neural information
sarial nets. arXiv preprint arXiv:1411.1784, 2014. processing systems, pp. 10–18, 2016.
Tao, C., Chen, L., Henao, R., Feng, J., and Duke, L. C.
Miyato, T. and Koyama, M. cgans with projection discrimi-
Chi-square generative adversarial network. In Dy, J.
nator. arXiv preprint arXiv:1802.05637, 2018.
and Krause, A. (eds.), Proceedings of the 35th Interna-
Miyato, T., Kataoka, T., Koyama, M., and Yoshida, Y. Spec- tional Conference on Machine Learning, volume 80 of
tral normalization for generative adversarial networks. In Proceedings of Machine Learning Research, pp. 4887–
International Conference on Learning Representations, 4896, Stockholmsmssan, Stockholm Sweden, 10–15 Jul
2018. URL https://openreview.net/forum? 2018. PMLR. URL http://proceedings.mlr.
id=B1QRgziT-. press/v80/tao18b.html.

Mohamed, S. and Lakshminarayanan, B. Learn- Tibshirani, R. Regression shrinkage and selection via the
ing in implicit generative models. arXiv preprint lasso. Journal of the Royal Statistical Society. Series B
arXiv:1610.03483, 2016. (Methodological), pp. 267–288, 1996.

Mousavi, A., Patel, A. B., and Baraniuk, R. G. A deep learn- Villani, C. Optimal transport: old and new, volume 338.
ing approach to structured signal recovery. In 2015 53rd Springer Science & Business Media, 2008.
Annual Allerton Conference on Communication, Control, Weiss, Y., Chang, H. S., and Freeman, W. T. Learning
and Computing (Allerton), pp. 1336–1343. IEEE, 2015. compressed sensing. In Snowbird Learning Workshop,
Allerton, CA. Citeseer, 2007.
Mousavi, A., Dasarathy, G., and Baraniuk, R. G. Deepcodec:
Adaptive sensing and recovery via deep convolutional Williams, R. J. and Zipser, D. A learning algorithm for con-
neural networks. arXiv preprint arXiv:1707.03386, 2017. tinually running fully recurrent neural networks. Neural
computation, 1(2):270–280, 1989.
Mousavi, A., Dasarathy, G., and Baraniuk, R. G. A data-
driven and distributed approach to sparse signal repre- Yann, L., Corinna, C., and Burges, C. The mnist
sentation and recovery. In International Conference on database of handwritten digits. URL http://yhann. lecun.
Learning Representations, 2018. com/exdb/mnist, 1998.
 Deep Compressed Sensing

Zhu, J.-Y., Park, T., Isola, P., and Efros, A. A. Unpaired

image-to-image translation using cycle-consistent adver-
sarial networks. arXiv preprint, 2017.