MobileNetV2: Inverted Residuals and Linear Bottlenecks

Mark Sandler Andrew Howard Menglong Zhu Andrey Zhmoginov Liang-Chieh Chen
Google Inc.
{sandler, howarda, menglong, azhmogin, lcchen}@google.com
arXiv:1801.04381v4 [cs.CV] 21 Mar 2019

Abstract applications.
This paper introduces a new neural network architec-
In this paper we describe a new mobile architecture, ture that is specifically tailored for mobile and resource
MobileNetV2, that improves the state of the art perfor- constrained environments. Our network pushes the state
mance of mobile models on multiple tasks and bench- of the art for mobile tailored computer vision models,
marks as well as across a spectrum of different model by significantly decreasing the number of operations and
sizes. We also describe efficient ways of applying these memory needed while retaining the same accuracy.
mobile models to object detection in a novel framework Our main contribution is a novel layer module: the
we call SSDLite. Additionally, we demonstrate how inverted residual with linear bottleneck. This mod-
to build mobile semantic segmentation models through ule takes as an input a low-dimensional compressed
a reduced form of DeepLabv3 which we call Mobile representation which is first expanded to high dimen-
DeepLabv3. sion and filtered with a lightweight depthwise convo-
is based on an inverted residual structure where lution. Features are subsequently projected back to a
the shortcut connections are between the thin bottle- low-dimensional representation with a linear convolu-
neck layers. The intermediate expansion layer uses tion. The official implementation is available as part of
lightweight depthwise convolutions to filter features as TensorFlow-Slim model library in [4].
a source of non-linearity. Additionally, we find that it is This module can be efficiently implemented using
important to remove non-linearities in the narrow layers standard operations in any modern framework and al-
in order to maintain representational power. We demon- lows our models to beat state of the art along multiple
strate that this improves performance and provide an in- performance points using standard benchmarks. Fur-
tuition that led to this design. thermore, this convolutional module is particularly suit-
Finally, our approach allows decoupling of the in- able for mobile designs, because it allows to signifi-
put/output domains from the expressiveness of the trans- cantly reduce the memory footprint needed during in-
formation, which provides a convenient framework for ference by never fully materializing large intermediate
further analysis. We measure our performance on tensors. This reduces the need for main memory access
ImageNet [1] classification, COCO object detection [2], in many embedded hardware designs, that provide small
VOC image segmentation [3]. We evaluate the trade-offs amounts of very fast software controlled cache memory.
between accuracy, and number of operations measured
by multiply-adds (MAdd), as well as actual latency, and 2. Related Work
the number of parameters.
Tuning deep neural architectures to strike an optimal
balance between accuracy and performance has been
1. Introduction an area of active research for the last several years.
Both manual architecture search and improvements in
Neural networks have revolutionized many areas of training algorithms, carried out by numerous teams has
machine intelligence, enabling superhuman accuracy for lead to dramatic improvements over early designs such
challenging image recognition tasks. However, the drive as AlexNet [5], VGGNet [6], GoogLeNet [7]. , and
to improve accuracy often comes at a cost: modern state ResNet [8]. Recently there has been lots of progress
of the art networks require high computational resources in algorithmic architecture exploration included hyper-
beyond the capabilities of many mobile and embedded parameter optimization [9, 10, 11] as well as various
methods of network pruning [12, 13, 14, 15, 16, 17] and convolutions. Effectively depthwise separable convolu-
connectivity learning [18, 19]. A substantial amount of tion reduces computation compared to traditional layers
work has also been dedicated to changing the connectiv- by almost a factor of k 21 . MobileNetV2 uses k = 3
ity structure of the internal convolutional blocks such as (3 × 3 depthwise separable convolutions) so the compu-
in ShuffleNet [20] or introducing sparsity [21] and oth- tational cost is 8 to 9 times smaller than that of standard
ers [22]. convolutions at only a small reduction in accuracy [27].
Recently, [23, 24, 25, 26], opened up a new direc-
tion of bringing optimization methods including genetic
3.2. Linear Bottlenecks
algorithms and reinforcement learning to architectural Consider a deep neural network consisting of n layers
search. However one drawback is that the resulting net- Li each of which has an activation tensor of dimensions
works end up very complex. In this paper, we pursue the hi × wi × di . Throughout this section we will be dis-
goal of developing better intuition about how neural net- cussing the basic properties of these activation tensors,
works operate and use that to guide the simplest possible which we will treat as containers of hi × wi “pixels”
network design. Our approach should be seen as compli- with di dimensions. Informally, for an input set of real
mentary to the one described in [23] and related work. images, we say that the set of layer activations (for any
In this vein our approach is similar to those taken by layer Li ) forms a “manifold of interest”. It has been long
[20, 22] and allows to further improve the performance, assumed that manifolds of interest in neural networks
while providing a glimpse on its internal operation. Our could be embedded in low-dimensional subspaces. In
network design is based on MobileNetV1 [27]. It re- other words, when we look at all individual d-channel
tains its simplicity and does not require any special op- pixels of a deep convolutional layer, the information
erators while significantly improves its accuracy, achiev- encoded in those values actually lie in some manifold,
ing state of the art on multiple image classification and which in turn is embeddable into a low-dimensional sub-
detection tasks for mobile applications. space2 .
At a first glance, such a fact could then be captured
3. Preliminaries, discussion and intuition and exploited by simply reducing the dimensionality of
a layer thus reducing the dimensionality of the oper-
3.1. Depthwise Separable Convolutions ating space. This has been successfully exploited by
Depthwise Separable Convolutions are a key build- MobileNetV1 [27] to effectively trade off between com-
ing block for many efficient neural network architectures putation and accuracy via a width multiplier parameter,
[27, 28, 20] and we use them in the present work as well. and has been incorporated into efficient model designs
The basic idea is to replace a full convolutional opera- of other networks as well [20]. Following that intuition,
tor with a factorized version that splits convolution into the width multiplier approach allows one to reduce the
two separate layers. The first layer is called a depthwise dimensionality of the activation space until the mani-
convolution, it performs lightweight filtering by apply- fold of interest spans this entire space. However, this
ing a single convolutional filter per input channel. The intuition breaks down when we recall that deep convo-
second layer is a 1 × 1 convolution, called a pointwise lutional neural networks actually have non-linear per co-
convolution, which is responsible for building new fea- ordinate transformations, such as ReLU. For example,
tures through computing linear combinations of the in- ReLU applied to a line in 1D space produces a ’ray’,
put channels. where as in Rn space, it generally results in a piece-wise
Standard convolution takes an hi × wi × di in- linear curve with n-joints.
put tensor Li , and applies convolutional kernel K ∈ It is easy to see that in general if a result of a layer
Rk×k×di ×dj to produce an hi × wi × dj output ten- transformation ReLU(Bx) has a non-zero volume S,
sor Lj . Standard convolutional layers have the compu- the points mapped to interior S are obtained via a lin-
tational cost of hi · wi · di · dj · k · k. ear transformation B of the input, thus indicating that
Depthwise separable convolutions are a drop-in re- the part of the input space corresponding to the full di-
placement for standard convolutional layers. Empiri- mensional output, is limited to a linear transformation.
cally they work almost as well as regular convolutions In other words, deep networks only have the power of
but only cost: a linear classifier on the non-zero volume part of the
1 more precisely, by a factor k2 dj /(k2 + dj )
hi · wi · di (k 2 + dj ) (1) 2 Note that dimensionality of the manifold differs from the dimen-
sionality of a subspace that could be embedded via a linear transfor-
which is the sum of the depthwise and 1 × 1 pointwise mation.
 Input Output/dim=2 Output/dim=3 Output/dim=5 Output/dim=15 Output/dim=30
(a) Residual block (b) Inverted residual block

Figure 1: Examples of ReLU transformations of

low-dimensional manifolds embedded in higher-dimensional
spaces. In these examples the initial spiral is embedded into
an n-dimensional space using random matrix T followed by
ReLU, and then projected back to the 2D space using T −1 . Figure 3: The difference between residual block [8, 30]
In examples above n = 2, 3 result in information loss where and inverted residual. Diagonally hatched layers do not
certain points of the manifold collapse into each other, while use non-linearities. We use thickness of each block to
for n = 15 to 30 the transformation is highly non-convex. indicate its relative number of channels. Note how clas-
sical residuals connects the layers with high number of
(a) Regular (b) Separable channels, whereas the inverted residuals connect the bot-
tlenecks. Best viewed in color.

2. ReLU is capable of preserving complete informa-

tion about the input manifold, but only if the input
manifold lies in a low-dimensional subspace of the
(c) Separable with linear (d) Bottleneck with ex- input space.
bottleneck pansion layer

These two insights provide us with an empirical hint

for optimizing existing neural architectures: assuming
the manifold of interest is low-dimensional we can cap-
ture this by inserting linear bottleneck layers into the
Figure 2: Evolution of separable convolution blocks. The convolutional blocks. Experimental evidence suggests
diagonally hatched texture indicates layers that do not contain that using linear layers is crucial as it prevents non-
non-linearities. The last (lightly colored) layer indicates the linearities from destroying too much information. In
beginning of the next block. Note: 2d and 2c are equivalent Section 6, we show empirically that using non-linear
blocks when stacked. Best viewed in color. layers in bottlenecks indeed hurts the performance by
several percent, further validating our hypothesis3 . We
note that similar reports where non-linearity was helped
output domain. We refer to supplemental material for were reported in [29] where non-linearity was removed
a more formal statement. from the input of the traditional residual block and that
On the other hand, when ReLU collapses the chan- lead to improved performance on CIFAR dataset.
nel, it inevitably loses information in that channel. How- For the remainder of this paper we will be utilizing
ever if we have lots of channels, and there is a a structure bottleneck convolutions. We will refer to the ratio be-
in the activation manifold that information might still be tween the size of the input bottleneck and the inner size
preserved in the other channels. In supplemental ma- as the expansion ratio.
terials, we show that if the input manifold can be em-
bedded into a significantly lower-dimensional subspace 3.3. Inverted residuals
of the activation space then the ReLU transformation
preserves the information while introducing the needed The bottleneck blocks appear similar to residual
complexity into the set of expressible functions. block where each block contains an input followed
by several bottlenecks then followed by expansion [8].
To summarize, we have highlighted two properties
However, inspired by the intuition that the bottlenecks
that are indicative of the requirement that the manifold
actually contain all the necessary information, while an
of interest should lie in a low-dimensional subspace of
expansion layer acts merely as an implementation detail
the higher-dimensional activation space:
that accompanies a non-linear transformation of the ten-
sor, we use shortcuts directly between the bottlenecks.
1. If the manifold of interest remains non-zero vol-
ume after ReLU transformation, it corresponds to 3 We note that in the presence of shortcuts the information loss is

a linear transformation. actually less strong.

Figure 3 provides a schematic visualization of the differ- Input Operator Output
ence in the designs. The motivation for inserting short- h×w×k 1x1 conv2d , ReLU6 h × w × (tk)
cuts is similar to that of classical residual connections: h × w × tk h w
3x3 dwise s=s, ReLU6 s × s × (tk)
we want to improve the ability of a gradient to propagate h w h w 0
s × s × tk linear 1x1 conv2d s × s ×k
across multiplier layers. However, the inverted design is
considerably more memory efficient (see Section 5 for
Table 1: Bottleneck residual block transforming from k
details), as well as works slightly better in our experi-
to k 0 channels, with stride s, and expansion factor t.
ments.

Running time and parameter count for bottleneck

convolution The basic implementation structure is il- Table 1. The architecture of MobileNetV2 contains the
lustrated in Table 1. For a block of size h × w, ex- initial fully convolution layer with 32 filters, followed
pansion factor t and kernel size k with d0 input chan- by 19 residual bottleneck layers described in the Ta-
nels and d00 output channels, the total number of multi- ble 2. We use ReLU6 as the non-linearity because of
ply add required is h · w · d0 · t(d0 + k 2 + d00 ). Com- its robustness when used with low-precision computa-
pared with (1) this expression has an extra term, as in- tion [27]. We always use kernel size 3 × 3 as is standard
deed we have an extra 1 × 1 convolution, however the for modern networks, and utilize dropout and batch nor-
nature of our networks allows us to utilize much smaller malization during training.
input and output dimensions. In Table 3 we compare the With the exception of the first layer, we use constant
needed sizes for each resolution between MobileNetV1, expansion rate throughout the network. In our experi-
MobileNetV2 and ShuffleNet. ments we find that expansion rates between 5 and 10 re-
3.4. Information flow interpretation sult in nearly identical performance curves, with smaller
networks being better off with slightly smaller expan-
One interesting property of our architecture is that it sion rates and larger networks having slightly better per-
provides a natural separation between the input/output formance with larger expansion rates.
domains of the building blocks (bottleneck layers), and
the layer transformation – that is a non-linear function For all our main experiments we use expansion factor
that converts input to the output. The former can be seen of 6 applied to the size of the input tensor. For example,
as the capacity of the network at each layer, whereas the for a bottleneck layer that takes 64-channel input tensor
latter as the expressiveness. This is in contrast with tra- and produces a tensor with 128 channels, the intermedi-
ditional convolutional blocks, both regular and separa- ate expansion layer is then 64 · 6 = 384 channels.
ble, where both expressiveness and capacity are tangled
together and are functions of the output layer depth.
In particular, in our case, when inner layer depth
is 0 the underlying convolution is the identity function Trade-off hyper parameters As in [27] we tailor our
thanks to the shortcut connection. When the expansion architecture to different performance points, by using
ratio is smaller than 1, this is a classical residual con- the input image resolution and width multiplier as tun-
volutional block [8, 30]. However, for our purposes we able hyper parameters, that can be adjusted depending
show that expansion ratio greater than 1 is the most use- on desired accuracy/performance trade-offs. Our pri-
ful. mary network (width multiplier 1, 224 × 224), has a
This interpretation allows us to study the expressive- computational cost of 300 million multiply-adds and
ness of the network separately from its capacity and we uses 3.4 million parameters. We explore the perfor-
believe that further exploration of this separation is war- mance trade offs, for input resolutions from 96 to 224,
ranted to provide a better understanding of the network and width multipliers of 0.35 to 1.4. The network com-
properties. putational cost ranges from 7 multiply adds to 585M
MAdds, while the model size vary between 1.7M and
4. Model Architecture 6.9M parameters.
Now we describe our architecture in detail. As dis- One minor implementation difference, with [27] is
cussed in the previous section the basic building block that for multipliers less than one, we apply width multi-
is a bottleneck depth-separable convolution with resid- plier to all layers except the very last convolutional layer.
uals. The detailed structure of this block is shown in This improves performance for smaller models.
 Input Operator t c n s conv 1x1, Relu6
2
224 × 3 conv2d - 32 1 2
1122 × 32 bottleneck 1 16 1 1
1122 × 16 bottleneck 6 24 2 2 Dwise 3x3,
stride=s, Relu6
562 × 24 bottleneck 6 32 3 2
282 × 32 bottleneck 6 64 4 2
142 × 64 bottleneck 6 96 3 1 input
142 × 96 bottleneck 6 160 3 2
72 × 160 bottleneck 6 320 1 1 (a) NasNet[23] (b) MobileNet[27]
72 × 320 conv2d 1x1 - 1280 1 1
72 × 1280 avgpool 7x7 - - 1 - Add conv 1x1, Linear

1 × 1 × 1280 conv2d 1x1 - k - conv 1x1, Linear

Dwise 3x3,
stride=2, Relu6

Dwise 3x3, Relu6

Table 2: MobileNetV2 : Each line describes a sequence

Conv 1x1, Relu6
of 1 or more identical (modulo stride) layers, repeated Conv 1x1, Relu6

n times. All layers in the same sequence have the same input input

number c of output channels. The first layer of each Stride=1 block Stride=2 block

sequence has a stride s and all others use stride 1. All (c) ShuffleNet [20] (d) Mobilenet V2
spatial convolutions use 3 × 3 kernels. The expansion
factor t is always applied to the input size as described Figure 4: Comparison of convolutional blocks for dif-
in Table 1. ferent architectures. ShuffleNet uses Group Convolu-
tions [20] and shuffling, it also uses conventional resid-
Size MobileNetV1 MobileNetV2 ShuffleNet ual approach where inner blocks are narrower than out-
(2x,g=3) put. ShuffleNet and NasNet illustrations are from re-
112x112 64/1600 16/400 32/800 spective papers.
56x56 128/800 32/200 48/300
28x28 256/400 64/100 400/600K
14x14 512/200 160/62 800/310
7x7 1024/199 320/32 1600/156
1x1 1024/2 1280/2 1600/3
cient implementation of inference that uses for instance
max 1600K 400K 600K
TensorFlow[31] or Caffe [32], builds a directed acyclic
compute hypergraph G, consisting of edges represent-
ing the operations and nodes representing tensors of in-
Table 3: The max number of channels/memory (in
termediate computation. The computation is scheduled
Kb) that needs to be materialized at each spatial res-
in order to minimize the total number of tensors that
olution for different architectures. We assume 16-bit
needs to be stored in memory. In the most general case,
floats for activations. For ShuffleNet, we use 2x, g =
it searches over all plausible computation orders Σ(G)
3 that matches the performance of MobileNetV1 and
and picks the one that minimizes
MobileNetV2. For the first layer of MobileNetV2 and
ShuffleNet we can employ the trick described in Sec-  
tion 5 to reduce memory requirement. Even though X
ShuffleNet employs bottlenecks elsewhere, the non- M (G) = min max  |A| + size(πi ).
π∈Σ(G) i∈1..n
bottleneck tensors still need to be materialized due to the A∈R(i,π,G)

presence of shortcuts between the non-bottleneck ten-

sors. where R(i, π, G) is the list of intermediate tensors that
are connected to any of πi . . . πn nodes, |A| represents
the size of the tensor A and size(i) is the total amount
5. Implementation Notes of memory needed for internal storage during operation
i.
5.1. Memory efficient inference
For graphs that have only trivial parallel structure
The inverted residual bottleneck layers allow a partic- (such as residual connection), there is only one non-
ularly memory efficient implementation which is very trivial feasible computation order, and thus the total
important for mobile applications. A standard effi- amount and a bound on the memory needed for infer-
ence on compute graph G can be simplified: 77.5 96x96 V2 1.4
75.0 128x128
72.5 V2 1.0
  70.0 160x160
67.5 192x192
224x224

Accuracy, Top 1, %
65.0
X X
M (G) = max  |A| + |B| + |op| 62.5 NasNet
op∈G
A∈opinp B∈opout
60.0 MobileNetV1
57.5 ShuffleNet
55.0
(2) 52.5
50.0
Or to restate, the amount of memory is simply the max- 47.5
45.0
imum total size of combined inputs and outputs across 42.5
40.0
all operations. In what follows we show that if we treat 37.5
35.07.5 10 15 20 30 40 50 75 100 150 200 300 400 500 600
a bottleneck residual block as a single operation (and Multiply-Adds, Millions
treat inner convolution as a disposable tensor), the total
amount of memory would be dominated by the size of Figure 5: Performance curve of MobileNetV2 vs
bottleneck tensors, rather than the size of tensors that are MobileNetV1, ShuffleNet, NAS. For our networks we
internal to bottleneck (and much larger). use multipliers 0.35, 0.5, 0.75, 1.0 for all resolutions,
and additional 1.4 for for 224. Best viewed in color.
Bottleneck Residual Block A bottleneck block oper-
ator F(x) shown in Figure 3b can be expressed as a 72 72
71 71
composition of three operators F(x) = [A ◦ N ◦ B]x, 70 70

Top 1 Accuracy

Top 1 Accuracy
where A is a linear transformation A : Rs×s×k → 69 69
68 68 Shortcut between bottlenecks
Rs×s×n , N is a non-linear per-channel transformation: 67 Linear botleneck 67 Shortcut between expansions
0 0 Relu6 in bottleneck No residual
N : Rs×s×n → Rs ×s ×n , and B is again a linear 660 1 2 3
Step, millions
4 5 6 7 660 1 2 3
Step, millions
4 5 6 7
0 0
transformation to the output domain: B : Rs ×s ×n →
0 0 0 (a) Impact of non-linearity in (b) Impact of variations in
Rs ×s ×k . the bottleneck layer. residual blocks.
For our networks N = ReLU6 ◦ dwise ◦ ReLU6 ,
but the results apply to any per-channel transformation. Figure 6: The impact of non-linearities and various
Suppose the size of the input domain is |x| and the size types of shortcut (residual) connections.
of the output domain is |y|, then the memory required
to compute F (X) can be as low as |s2 k| + |s02 k 0 | +
O(max(s2 , s02 )).
creased cache misses. We find that this approach is the
The algorithm is based on the fact that the inner ten-
most helpful to be used with t being a small constant
sor I can be represented as concatenation of t tensors, of
between 2 and 5. It significantly reduces the memory
size n/t each and our function can then be represented
requirement, but still allows one to utilize most of the ef-
as
Xt ficiencies gained by using highly optimized matrix mul-
F(x) = (Ai ◦ N ◦ Bi )(x) tiplication and convolution operators provided by deep
i=1 learning frameworks. It remains to be seen if special
by accumulating the sum, we only require one interme- framework level optimization may lead to further run-
diate block of size n/t to be kept in memory at all times. time improvements.
Using n = t we end up having to keep only a single
channel of the intermediate representation at all times.
6. Experiments
The two constraints that enabled us to use this trick is
(a) the fact that the inner transformation (which includes 6.1. ImageNet Classification
non-linearity and depthwise) is per-channel, and (b) the
consecutive non-per-channel operators have significant Training setup We train our models using
ratio of the input size to the output. For most of the tra- TensorFlow[31]. We use the standard RMSPropOp-
ditional neural networks, such trick would not produce a timizer with both decay and momentum set to 0.9.
significant improvement. We use batch normalization after every layer, and the
We note that, the number of multiply-adds opera- standard weight decay is set to 0.00004. Following
tors needed to compute F (X) using t-way split is in- MobileNetV1[27] setup we use initial learning rate of
dependent of t, however in existing implementations we 0.045, and learning rate decay rate of 0.98 per epoch.
find that replacing one matrix multiplication with sev- We use 16 GPU asynchronous workers, and a batch size
eral smaller ones hurts runtime performance due to in- of 96.
Results We compare our networks against Params MAdds
MobileNetV1, ShuffleNet and NASNet-A models.
The statistics of a few selected models is shown in SSD[34] 14.8M 1.25B
Table 4 with the full performance graph shown in SSDLite 2.1M 0.35B
Figure 5.
Table 5: Comparison of the size and the computa-
6.2. Object Detection tional cost between SSD and SSDLite configured with
MobileNetV2 and making predictions for 80 classes.
We evaluate and compare the performance of
MobileNetV2 and MobileNetV1 as feature extractors
[33] for object detection with a modified version of the Network mAP Params MAdd CPU
Single Shot Detector (SSD) [34] on COCO dataset [2]. SSD300[34] 23.2 36.1M 35.2B -
We also compare to YOLOv2 [35] and original SSD SSD512[34] 26.8 36.1M 99.5B -
(with VGG-16 [6] as base network) as baselines. We do YOLOv2[35] 21.6 50.7M 17.5B -
not compare performance with other architectures such MNet V1 + SSDLite 22.2 5.1M 1.3B 270ms
as Faster-RCNN [36] and RFCN [37] since our focus is MNet V2 + SSDLite 22.1 4.3M 0.8B 200ms
on mobile/real-time models.
SSDLite: In this paper, we introduce a mobile Table 6: Performance comparison of MobileNetV2 +
friendly variant of regular SSD. We replace all the regu- SSDLite and other realtime detectors on the COCO
lar convolutions with separable convolutions (depthwise dataset object detection task. MobileNetV2 + SSDLite
followed by 1 × 1 projection) in SSD prediction lay- achieves competitive accuracy with significantly fewer
ers. This design is in line with the overall design of parameters and smaller computational complexity. All
MobileNets and is seen to be much more computation- models are trained on trainval35k and evaluated on
ally efficient. We call this modified version SSDLite. test-dev. SSD/YOLOv2 numbers are from [35]. The
Compared to regular SSD, SSDLite dramatically re- running time is reported for the large core of the Google
duces both parameter count and computational cost as Pixel 1 phone, using an internal version of the TF-Lite
shown in Table 5. engine.
For MobileNetV1, we follow the setup in [33]. For
MobileNetV2, the first layer of SSDLite is attached to
the expansion of layer 15 (with output stride of 16). The Both MobileNet models are trained and evalu-
second and the rest of SSDLite layers are attached on top ated with Open Source TensorFlow Object Detection
of the last layer (with output stride of 32). This setup is API [38]. The input resolution of both models is 320 ×
consistent with MobileNetV1 as all layers are attached 320. We benchmark and compare both mAP (COCO
to the feature map of the same output strides. challenge metrics), number of parameters and number
of Multiply-Adds. The results are shown in Table 6.
MobileNetV2 SSDLite is not only the most efficient
Network Top 1 Params MAdds CPU model, but also the most accurate of the three. No-
MobileNetV1 70.6 4.2M 575M 113ms
tably, MobileNetV2 SSDLite is 20× more efficient and
ShuffleNet (1.5) 71.5 3.4M 292M - 10× smaller while still outperforms YOLOv2 on COCO
ShuffleNet (x2) 73.7 5.4M 524M - dataset.
NasNet-A 74.0 5.3M 564M 183ms
MobileNetV2 72.0 3.4M 300M 75ms 6.3. Semantic Segmentation
MobileNetV2 (1.4) 74.7 6.9M 585M 143ms
In this section, we compare MobileNetV1 and
MobileNetV2 models used as feature extractors with
Table 4: Performance on ImageNet, comparison for dif- DeepLabv3 [39] for the task of mobile semantic seg-
ferent networks. As is common practice for ops, we mentation. DeepLabv3 adopts atrous convolution [40,
count the total number of Multiply-Adds. In the last 41, 42], a powerful tool to explicitly control the reso-
column we report running time in milliseconds (ms) for lution of computed feature maps, and builds five paral-
a single large core of the Google Pixel 1 phone (using lel heads including (a) Atrous Spatial Pyramid Pooling
TF-Lite). We do not report ShuffleNet numbers as effi- module (ASPP) [43] containing three 3 × 3 convolu-
cient group convolutions and shuffling are not yet sup- tions with different atrous rates, (b) 1 × 1 convolution
ported. head, and (c) Image-level features [44]. We denote by
output stride the ratio of input image spatial resolution Network OS ASPP MF mIOU Params MAdds
to final output resolution, which is controlled by apply- MNet V1 16 X 75.29 11.15M 14.25B
ing the atrous convolution properly. For semantic seg- 8 X X 78.56 11.15M 941.9B
mentation, we usually employ output stride = 16 or 8 MNet V2* 16 X 75.70 4.52M 5.8B
for denser feature maps. We conduct the experiments 8 X X 78.42 4.52M 387B
on the PASCAL VOC 2012 dataset [3], with extra anno- MNet V2* 16 75.32 2.11M 2.75B
tated images from [45] and evaluation metric mIOU. 8 X 77.33 2.11M 152.6B

To build a mobile model, we experimented with three ResNet-101 16 X 80.49 58.16M 81.0B
8 X X 82.70 58.16M 4870.6B
design variations: (1) different feature extractors, (2)
simplifying the DeepLabv3 heads for faster computa-
Table 7: MobileNet + DeepLabv3 inference strategy
tion, and (3) different inference strategies for boost-
on the PASCAL VOC 2012 validation set. MNet
ing the performance. Our results are summarized in
V2*: Second last feature map is used for DeepLabv3
Table 7. We have observed that: (a) the inference
heads, which includes (1) Atrous Spatial Pyramid Pool-
strategies, including multi-scale inputs and adding left-
ing (ASPP) module, and (2) 1 × 1 convolution as well
right flipped images, significantly increase the MAdds
as image-pooling feature. OS: output stride that con-
and thus are not suitable for on-device applications,
trols the output resolution of the segmentation map. MF:
(b) using output stride = 16 is more efficient than
Multi-scale and left-right flipped inputs during test. All
output stride = 8, (c) MobileNetV1 is already a pow-
of the models have been pretrained on COCO. The po-
erful feature extractor and only requires about 4.9 − 5.7
tential candidate for on-device applications is shown in
times fewer MAdds than ResNet-101 [8] (e.g., mIOU:
bold face. PASCAL images have dimension 512 × 512
78.56 vs 82.70, and MAdds: 941.9B vs 4870.6B), (d)
and atrous convolution allows us to control output fea-
it is more efficient to build DeepLabv3 heads on top of
ture resolution without increasing the number of param-
the second last feature map of MobileNetV2 than on the
eters.
original last-layer feature map, since the second to last
feature map contains 320 channels instead of 1280, and
by doing so, we attain similar performance, but require 7. Conclusions and future work
about 2.5 times fewer operations than the MobileNetV1
counterparts, and (e) DeepLabv3 heads are computa- We described a very simple network architecture that
tionally expensive and removing the ASPP module sig- allowed us to build a family of highly efficient mobile
nificantly reduces the MAdds with only a slight perfor- models. Our basic building unit, has several proper-
mance degradation. In the end of the Table 7, we identify ties that make it particularly suitable for mobile appli-
a potential candidate for on-device applications (in bold cations. It allows very memory-efficient inference and
face), which attains 75.32% mIOU and only requires relies utilize standard operations present in all neural
2.75B MAdds. frameworks.
For the ImageNet dataset, our architecture improves
the state of the art for wide range of performance points.
6.4. Ablation study
For object detection task, our network outperforms
Inverted residual connections. The importance of state-of-art realtime detectors on COCO dataset both in
residual connection has been studied extensively [8, terms of accuracy and model complexity. Notably, our
30, 46]. The new result reported in this paper is that architecture combined with the SSDLite detection mod-
the shortcut connecting bottleneck perform better than ule is 20× less computation and 10× less parameters
shortcuts connecting the expanded layers (see Figure 6b than YOLOv2.
for comparison). On the theoretical side: the proposed convolutional
block has a unique property that allows to separate the
Importance of linear bottlenecks. The linear bottle-
network expressiveness (encoded by expansion layers)
neck models are strictly less powerful than models with
from its capacity (encoded by bottleneck inputs). Ex-
non-linearities, because the activations can always op-
ploring this is an important direction for future research.
erate in linear regime with appropriate changes to bi-
ases and scaling. However our experiments shown in
Figure 6a indicate that linear bottlenecks improve per- Acknowledgments We would like to thank Matt
formance, providing support that non-linearity destroys Streeter and Sergey Ioffe for their helpful feedback and
information in low-dimensional space. discussion.
References [11] Jasper Snoek, Oren Rippel, Kevin Swersky, Ryan
Kiros, Nadathur Satish, Narayanan Sundaram, Md.
[1] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Mostofa Ali Patwary, Prabhat, and Ryan P. Adams.
Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Scalable bayesian optimization using deep neu-
Huang, Andrej Karpathy, Aditya Khosla, Michael
ral networks. In Francis R. Bach and David M.
Bernstein, Alexander C. Berg, and Li Fei-Fei. Im-
Blei, editors, Proceedings of the 32nd Interna-
agenet large scale visual recognition challenge.
tional Conference on Machine Learning, ICML
Int. J. Comput. Vision, 115(3):211–252, December 2015, Lille, France, 6-11 July 2015, volume 37
2015. 1 of JMLR Workshop and Conference Proceedings,
[2] Tsung-Yi Lin, Michael Maire, Serge Belongie, pages 2171–2180. JMLR.org, 2015. 1
James Hays, Pietro Perona, Deva Ramanan, Piotr
Dollár, and C Lawrence Zitnick. Microsoft COCO: [12] Babak Hassibi and David G. Stork. Second or-
Common objects in context. In ECCV, 2014. 1, 7 der derivatives for network pruning: Optimal brain
surgeon. In Stephen Jose Hanson, Jack D. Cowan,
[3] Mark Everingham, S. M. Ali Eslami, Luc Van and C. Lee Giles, editors, Advances in Neural In-
Gool, Christopher K. I. Williams, John Winn, and formation Processing Systems 5, [NIPS Confer-
Andrew Zisserma. The pascal visual object classes ence, Denver, Colorado, USA, November 30 - De-
challenge a retrospective. IJCV, 2014. 1, 8 cember 3, 1992], pages 164–171. Morgan Kauf-
mann, 1992. 2
[4] Mobilenetv2 source code. Available from
https://github.com/tensorflow/ [13] Yann LeCun, John S. Denker, and Sara A. Solla.
models/tree/master/research/slim/ Optimal brain damage. In David S. Touretzky,
nets/mobilenet. 1 editor, Advances in Neural Information Process-
[5] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. ing Systems 2, [NIPS Conference, Denver, Col-
Hinton. Imagenet classification with deep convolu- orado, USA, November 27-30, 1989], pages 598–
tional neural networks. In Bartlett et al. [48], pages 605. Morgan Kaufmann, 1989. 2
1106–1114. 1 [14] Song Han, Jeff Pool, John Tran, and William J.
[6] Karen Simonyan and Andrew Zisserman. Very Dally. Learning both weights and connec-
deep convolutional networks for large-scale image tions for efficient neural network. In Corinna
recognition. CoRR, abs/1409.1556, 2014. 1, 7 Cortes, Neil D. Lawrence, Daniel D. Lee, Masashi
Sugiyama, and Roman Garnett, editors, Advances
[7] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre in Neural Information Processing Systems 28: An-
Sermanet, Scott E. Reed, Dragomir Anguelov, Du- nual Conference on Neural Information Process-
mitru Erhan, Vincent Vanhoucke, and Andrew Ra- ing Systems 2015, December 7-12, 2015, Mon-
binovich. Going deeper with convolutions. In treal, Quebec, Canada, pages 1135–1143, 2015.
IEEE Conference on Computer Vision and Pattern 2
Recognition, CVPR 2015, Boston, MA, USA, June
7-12, 2015, pages 1–9. IEEE Computer Society, [15] Song Han, Jeff Pool, Sharan Narang, Huizi Mao,
2015. 1 Shijian Tang, Erich Elsen, Bryan Catanzaro, John
Tran, and William J. Dally. DSD: regulariz-
[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and ing deep neural networks with dense-sparse-dense
Jian Sun. Deep residual learning for image recog- training flow. CoRR, abs/1607.04381, 2016. 2
nition. CoRR, abs/1512.03385, 2015. 1, 3, 4, 8
[9] James Bergstra and Yoshua Bengio. Random [16] Yiwen Guo, Anbang Yao, and Yurong Chen. Dy-
search for hyper-parameter optimization. Journal namic network surgery for efficient dnns. In
of Machine Learning Research, 13:281–305, 2012. Daniel D. Lee, Masashi Sugiyama, Ulrike von
1 Luxburg, Isabelle Guyon, and Roman Garnett, ed-
itors, Advances in Neural Information Processing
[10] Jasper Snoek, Hugo Larochelle, and Ryan P. Systems 29: Annual Conference on Neural Infor-
Adams. Practical bayesian optimization of ma- mation Processing Systems 2016, December 5-10,
chine learning algorithms. In Bartlett et al. [48], 2016, Barcelona, Spain, pages 1379–1387, 2016.
pages 2960–2968. 1 2
[17] Hao Li, Asim Kadav, Igor Durdanovic, Hanan Mobilenets: Efficient convolutional neural net-
Samet, and Hans Peter Graf. Pruning filters for works for mobile vision applications. CoRR,
efficient convnets. CoRR, abs/1608.08710, 2016. abs/1704.04861, 2017. 2, 4, 5, 6
2
[28] Francois Chollet. Xception: Deep learning
[18] Karim Ahmed and Lorenzo Torresani. Connec- with depthwise separable convolutions. In The
tivity learning in multi-branch networks. CoRR, IEEE Conference on Computer Vision and Pattern
abs/1709.09582, 2017. 2 Recognition (CVPR), July 2017. 2
[19] Tom Veniat and Ludovic Denoyer. Learning time-
efficient deep architectures with budgeted super [29] Dongyoon Han, Jiwhan Kim, and Junmo Kim.
networks. CoRR, abs/1706.00046, 2017. 2 Deep pyramidal residual networks. CoRR,
abs/1610.02915, 2016. 3
[20] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and
Jian Sun. Shufflenet: An extremely efficient [30] Saining Xie, Ross B. Girshick, Piotr Dollár,
convolutional neural network for mobile devices. Zhuowen Tu, and Kaiming He. Aggregated
CoRR, abs/1707.01083, 2017. 2, 5 residual transformations for deep neural networks.
CoRR, abs/1611.05431, 2016. 3, 4, 8
[21] Soravit Changpinyo, Mark Sandler, and Andrey
Zhmoginov. The power of sparsity in convolu- [31] Martı́n Abadi, Ashish Agarwal, Paul Barham,
tional neural networks. CoRR, abs/1702.06257, Eugene Brevdo, Zhifeng Chen, Craig Citro,
2017. 2 Greg S. Corrado, Andy Davis, Jeffrey Dean,
Matthieu Devin, Sanjay Ghemawat, Ian Goodfel-
[22] Min Wang, Baoyuan Liu, and Hassan Foroosh.
low, Andrew Harp, Geoffrey Irving, Michael Isard,
Design of efficient convolutional layers using sin-
Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser,
gle intra-channel convolution, topological subdivi-
Manjunath Kudlur, Josh Levenberg, Dan Mané,
sioning and spatial ”bottleneck” structure. CoRR,
Rajat Monga, Sherry Moore, Derek Murray, Chris
abs/1608.04337, 2016. 2
Olah, Mike Schuster, Jonathon Shlens, Benoit
[23] Barret Zoph, Vijay Vasudevan, Jonathon Shlens, Steiner, Ilya Sutskever, Kunal Talwar, Paul Tucker,
and Quoc V. Le. Learning transferable archi- Vincent Vanhoucke, Vijay Vasudevan, Fernanda
tectures for scalable image recognition. CoRR, Viégas, Oriol Vinyals, Pete Warden, Martin Wat-
abs/1707.07012, 2017. 2, 5 tenberg, Martin Wicke, Yuan Yu, and Xiaoqiang
Zheng. TensorFlow: Large-scale machine learning
[24] Lingxi Xie and Alan L. Yuille. Genetic CNN. on heterogeneous systems, 2015. Software avail-
CoRR, abs/1703.01513, 2017. 2 able from tensorflow.org. 5, 6
[25] Esteban Real, Sherry Moore, Andrew Selle, [32] Yangqing Jia, Evan Shelhamer, Jeff Donahue,
Saurabh Saxena, Yutaka Leon Suematsu, Jie Tan, Sergey Karayev, Jonathan Long, Ross Girshick,
Quoc V. Le, and Alexey Kurakin. Large-scale Sergio Guadarrama, and Trevor Darrell. Caffe:
evolution of image classifiers. In Doina Pre- Convolutional architecture for fast feature embed-
cup and Yee Whye Teh, editors, Proceedings of ding. arXiv preprint arXiv:1408.5093, 2014. 5
the 34th International Conference on Machine
Learning, ICML 2017, Sydney, NSW, Australia, [33] Jonathan Huang, Vivek Rathod, Chen Sun, Men-
6-11 August 2017, volume 70 of Proceedings of glong Zhu, Anoop Korattikara, Alireza Fathi,
Machine Learning Research, pages 2902–2911. Ian Fischer, Zbigniew Wojna, Yang Song, Sergio
PMLR, 2017. 2 Guadarrama, et al. Speed/accuracy trade-offs for
[26] Barret Zoph and Quoc V. Le. Neural architec- modern convolutional object detectors. In CVPR,
ture search with reinforcement learning. CoRR, 2017. 7
abs/1611.01578, 2016. 2
[34] Wei Liu, Dragomir Anguelov, Dumitru Erhan,
[27] Andrew G. Howard, Menglong Zhu, Bo Chen, Christian Szegedy, Scott Reed, Cheng-Yang Fu,
Dmitry Kalenichenko, Weijun Wang, Tobias and Alexander C Berg. Ssd: Single shot multibox
Weyand, Marco Andreetto, and Hartwig Adam. detector. In ECCV, 2016. 7
[35] Joseph Redmon and Ali Farhadi. Yolo9000: [45] Bharath Hariharan, Pablo Arbeláez, Lubomir
Better, faster, stronger. arXiv preprint Bourdev, Subhransu Maji, and Jitendra Malik. Se-
arXiv:1612.08242, 2016. 7 mantic contours from inverse detectors. In ICCV,
2011. 8
[36] Shaoqing Ren, Kaiming He, Ross Girshick, and
Jian Sun. Faster r-cnn: Towards real-time object [46] Christian Szegedy, Sergey Ioffe, and Vincent Van-
detection with region proposal networks. In Ad- houcke. Inception-v4, inception-resnet and the im-
vances in neural information processing systems, pact of residual connections on learning. CoRR,
pages 91–99, 2015. 7 abs/1602.07261, 2016. 8
[47] Guido Montúfar, Razvan Pascanu, Kyunghyun
[37] Jifeng Dai, Yi Li, Kaiming He, and Jian Sun. R-
Cho, and Yoshua Bengio. On the number of linear
fcn: Object detection via region-based fully con-
regions of deep neural networks. In Proceedings of
volutional networks. In Advances in neural infor-
the 27th International Conference on Neural Infor-
mation processing systems, pages 379–387, 2016.
mation Processing Systems, NIPS’14, pages 2924–
7
2932, Cambridge, MA, USA, 2014. MIT Press. 13
[38] Jonathan Huang, Vivek Rathod, Derek Chow, [48] Peter L. Bartlett, Fernando C. N. Pereira, Christo-
Chen Sun, and Menglong Zhu. Tensorflow object pher J. C. Burges, Léon Bottou, and Kilian Q.
detection api, 2017. 7 Weinberger, editors. Advances in Neural Infor-
mation Processing Systems 25: 26th Annual Con-
[39] Liang-Chieh Chen, George Papandreou, Florian ference on Neural Information Processing Systems
Schroff, and Hartwig Adam. Rethinking atrous 2012. Proceedings of a meeting held December 3-
convolution for semantic image segmentation. 6, 2012, Lake Tahoe, Nevada, United States, 2012.
CoRR, abs/1706.05587, 2017. 7 9
[40] Matthias Holschneider, Richard Kronland- A. Bottleneck transformation
Martinet, Jean Morlet, and Ph Tchamitchian.
A real-time algorithm for signal analysis with In this section we study the properties of an operator
the help of the wavelet transform. In Wavelets: A ReLU(Bx), where x ∈ Rn represents an n-channel
Time-Frequency Methods and Phase Space, pages pixel, B is an m × n matrix and A is an n × m matrix.
289–297. 1989. 7 We argue that if m ≤ n, transformations of this form
can only exploit non-linearity at the cost of losing infor-
[41] Pierre Sermanet, David Eigen, Xiang Zhang, mation. In contrast, if n  m, such transforms can be
Michaël Mathieu, Rob Fergus, and Yann Le- highly non-linear but still invertible with high probabil-
Cun. Overfeat: Integrated recognition, localiza- ity (for the initial random weights).
tion and detection using convolutional networks. First we show that ReLU is an identity transforma-
arXiv:1312.6229, 2013. 7 tion for any point that lies in the interior of its image.
Lemma 1 Let S(X) = {ReLU(x)|x ∈ X}. If a vol-
[42] George Papandreou, Iasonas Kokkinos, and Pierre-
ume of S(X) is non-zero, then interior S(X) ⊆ X.
Andre Savalle. Modeling local and global defor-
mations in deep learning: Epitomic convolution,
multiple instance learning, and sliding window de- Proof: Let S 0 = interior ReLU(S). First we note that
tection. In CVPR, 2015. 7 if x ∈ S 0 , then xi > 0 for all i. Indeed, image of ReLU
does not contain points with negative coordinates, and
[43] Liang-Chieh Chen, George Papandreou, Iasonas points with zero-valued coordinates can not be interior
Kokkinos, Kevin Murphy, and Alan L Yuille. points. Therefore for each x ∈ S 0 , x = ReLU(x) as
Deeplab: Semantic image segmentation with deep desired. 
convolutional nets, atrous convolution, and fully It follows that for an arbitrary composition of inter-
connected crfs. TPAMI, 2017. 7 leaved linear transformation and ReLU operators, if it
preserves non-zero volume, that part of the input space
[44] Wei Liu, Andrew Rabinovich, and Alexander C. X that is preserved over such a composition is a lin-
Berg. Parsenet: Looking wider to see better. CoRR, ear transformation, and thus is likely to be a minor con-
abs/1506.04579, 2015. 7 tributor to the power of deep networks. However, this
is a fairly weak statement. Indeed, if the input mani- the positive channel (as predicted by initialization sym-
fold can be embedded into (n − 1)-dimensional mani- metries). For fully trained network, while the standard
fold (out of n dimensions total), the lemma is trivially deviation grew significantly, all but the two layers are
true, since the starting volume is 0. In what follows we still above the invertibility thresholds. We believe fur-
show that when the dimensionality of input manifold is ther study of this is warranted and might lead to helpful
significantly lower we can ensure that there will be no insights on network design.
information loss.
Since the ReLU(x) nonlinearity is a surjective func- Theorem 1 Let S be a compact n-dimensional subman-
tion mapping the entire ray x ≤ 0 to 0, using this nonlin- ifold of Rn . Consider a family of functions fB (x) =
earity in a neural network can result in information loss. ReLU(Bx) from Rn to Rm parameterized by m × n
Once ReLU collapses a subset of the input manifold to a matrices B ∈ B. Let p(B) be a probability density on
smaller-dimensional output, the following network lay- the space of all matrices B that satisfies:
ers can no longer distinguish between collapsed input
• P (Z) = 0 for any measure-zero subset Z ⊂ B;
samples. In the following, we show that bottlenecks with
sufficiently large expansion layers are resistant to infor- • (a symmetry condition) p(DB) = p(B) for any
mation loss caused by the presence of ReLU activation B ∈ B and any m × m diagonal matrix D with
functions. all diagonal elements being either +1 or −1.
Lemma 2 (Invertibility of ReLU) Consider an opera- Then, the average n-volume of the subset of S that is
tor ReLU(Bx), where B is an m × n matrix and x ∈ collapsed by fB to a lower-dimensional manifold is
Rn . Let y0 = ReLU(Bx0 ) for some x0 ∈ Rn , then
equation y0 = ReLU(Bx) has a unique solution with Nm,n V
V − ,
respect to x if and only if y0 has at least n non-zero val- 2m
ues and there are n linearly independent rows of B that
correspond to non-zero coordinates of y0 . where V = vol S and
m−n
X 
m
Proof: Denote the set of non-zero coordinates of y0 Nm,n ≡ .
k
as T and let yT and BT be restrictions of y and B k=0
to the subspace defined by T . If |T | < n, we have
yT = BT x0 where BT is under-determined with at Proof: For any σ = (s1 , . . . , sm ) with sk ∈
least one solution x0 , thus there are infinitely many so- {−1, +1}, let Qσ = {x ∈ Rm |xi si > 0} be a corre-
lutions. Now consider the case of |T | ≥ n and let the sponding quadrant in Rm . For any n-dimensional sub-
rank of BT be n. Suppose there is an additional solu- manifold Γ ⊂ Rm , ReLU acts as a bijection on Γ ∩ Qσ
tion x1 6= x0 such that y0 = ReLU(Bx1 ), then we have if σ has at least n positive values4 and contracts Γ ∩ Qσ
yT = BT x0 = BT x1 , which cannot be satisfied unless otherwise. Also notice that the intersection of BS with
x0 = x1 .  Rm \(∪σ Qσ ) is almost surely (n − 1)-dimensional. The
One of the corollaries of this lemma says that if m  average n-volume of S that is not collapsed by applying
n, we only need a small fraction of values of Bx to be ReLU to BS is therefore given by:
positive for ReLU(Bx) to be invertible. X
The constraints of the lemma 2 can be empirically EB [Vσ (B)], (3)
validated for real networks and real inputs and hence we σ∈Σn
can be assured that information is indeed preserved. We P
further show that with respect to initialization, we can be where Σn = {(s1 , . . . , sm )| k θ(sk ) ≥ n}, θ is a step
sure that these constraints are satisfied with high proba- function and Vσ (B) is a volume of the largest subset
bility. Note that for random initialization the conditions of S that is mapped by B to Qσ . Now let us calcu-
of lemma 2 are satisfied due to initialization symmetries. late EB [Vσ (B)]. Recalling that p(DB) = p(B) for
However even for trained graphs these constraints can be any D = diag(s1 , . . . , sm ) with sk ∈ {−1, +1}, this
empirically validated by running the network over valid average can be rewritten as EB ED [Vσ (DB)]. Notic-
inputs and verifying that all or most inputs are above ing that the subset of S mapped by DB to Qσ is
the threshold. On Figure 7 we show how this distribu- also mapped by B to D−1 Qσ , we immediately obtain
tion looks for different MobileNetV2 layers. At step 0 4 unless at least one of the positive coordinates for all x ∈ Γ ∩ Q
σ
the activation patterns concentrate around having half of is fixed, which would not be the case for almost all B and Γ = BS
 1000 1.0
1000 1.0
avg 0.9 avg
800 min min
max 0.8 800 0.8
Num positive filters max

Num positive filters

Fraction of filters
total

Fraction of filters
0.7 total
600 threshold 0.6 600 threshold 0.6

0.5
400 400 0.4
0.4
200 0.3 200 0.2
0.2
0 0.0
0 0.1 0 2 4 6 8 10 12 14 16 0 2 4 6 8 10 12 14 16
0 2 4 6 8 10 12 14 16 0 2 4 6 8 10 12 14 16 Layer N Layer N
Layer N Layer N

(a) At step 0 (b) Fully trained

Figure 7: Distribution of activation patterns. The x-axis is the layer index, and we show minimum/maximum/average
number of positive channels after each convolution with ReLU. y-axis is either absolute or relative number of chan-
nels. The “threshold” line indicates the ReLU invertibility threshold - that is the number of positive dimensions is
higher than the input space. In our case this is 1/6 fraction of the channels. Note how at the beginning of the train-
ing on Figure 7a the distribution is much more tightly concentrated around the mean. After the training has finished
(Figure 7b), the average hasn’t changed but the standard deviation grew dramatically. Best viewed in color.

Vσ [diag(σ 0 )B] = σ0 Vσ0 [B] = vol S and there-

P P
σ0
fore EBP [Vσ (B)] = 2−m vol S. Substituting this and
m−n m

|Σn | = k=0 k into Eq. 3 concludes the proof. 
Notice that for sufficiently large expansion layers
with m  n, the fraction of collapsed space Nm,n /2m
can be bounded by:

Nm,n mn+1
≥ 1− ≥ 1−2(n+1) log m−m ≥ 1−2−m/2
2m 2m n!
and therefore ReLU(Bx) performs a nonlinear transfor-
mation while preserving information with high proba-
bility.
We discussed how bottlenecks can prevent manifold
collapse, but increasing the size of the bottleneck expan-
sion may also make it possible for the network to repre-
sent more complex functions. Following the main re-
sults of [47], one can show, for example, that for any in-
teger L ≥ 1 and p > 1 there exist a network of L ReLU
layers, each containing n neurons and a bottleneck ex-
pansion of size pn such that it maps pnL input volumes
(linearly isomorphic to [0, 1]n ) to the same output re-
gion [0, 1]n . Any complex possibly nonlinear function
attached to the network output would thus effectively
compute function values for pnL input linear regions.

B. Semantic segmentation visualization re-

sults
Figure 8: MobileNetv2 semantic segmentation visualization results on PASCAL VOC 2012 val set. OS: output stride.
S: single scale input. MS+F: Multi-scale inputs with scales = {0.5, 0.75, 1, 1.25, 1.5, 1.75} and left-right flipped
inputs. Employing output stride = 16 and single input scale = 1 attains a good trade-off between FLOPS and accuracy.