See discussions, stats, and author profiles for this publication at: [Link]

net/publication/364640176

Visualizing and Understanding Convolutional Networks

Preprint · October 2022

DOI: 10.13140/RG.2.2.12182.22080

CITATION READS
1 2,185

2 authors, including:

Md Hafizur Rahman Masum

Frankfurt University of Applied Sciences
6 PUBLICATIONS 3 CITATIONS

SEE PROFILE

All content following this page was uploaded by Md Hafizur Rahman Masum on 01 November 2022.

The user has requested enhancement of the downloaded file.

 Visualizing and Understanding Convolutional
Networks
Md Hafizur Rahman Masum
Frankfurt University of Applied Sciences
Frankfurt am Main, Germany
[Link]@[Link]
Matriculation Number: 1354904

Abstract—Convolutional Neural Networks (CNNs) are capable Zeiler et al. suggest employs a multi-layered Deconvolutional
of performing impressively working on computer vision tasks Network (deconvnet), as presented by Zeiler et al. (2011). By
of all kinds, including object identification, picture recognition, obscuring some of the input picture, it can also be possible to
image retrieval, etc. These successes are a result of CNN’s
exceptional capacity to learn input information using several deep do a sensitivity analysis on the classifier output and determine
layers of neuronal structures and iterative training methods. On which aspects of the scene are crucial for categorization.[1]
the other hand, from a human visual standpoint, these learnt
characteristics are challenging to recognize and interpret, making Zeiler et al. examine many designs using these tools, starting
it difficult to comprehend the CNN’s internal workings.[1] In
order to enhance CNN interpretability, the CNN visualization with the architecture of (Krizhevsky et al., 2012), and they
is effectively used as a qualitative analysis method that converts eventually find ones that surpass their ImageNet findings. After
internal information into visually observable patterns. A number that, they simply retrain the softmax classifier on top of the
of visualization techniques are described in this paper, including model to see whether it can be applied to different datasets.
Deconvolutional Networks (DeconNet), Activation Maximization, In contrast to the unsupervised pre-training techniques made
Network Dissection and Network Inversion.[2] The performance
contribution of different model layers are also investigated famous by some researchers, this is a type of supervised
through an ablation study. pre-training. In related work , the generalizability of convnet
Index Terms—Convolutional Neural Network, Training Image, characteristics is also investigated[5] [6] [7] [1]
DeconNet, Activation Maximization, Network Dissection, Net-
work Inversion

I. I NTRODUCTION II. R ELATED W ORK

There has been excellent performance demonstrated by
convolutional networks in tasks such as Classification of It is standard practice to visualize features to obtain insight
handwritten digits and detection of faces from the early about the network, however this is often restricted to the
1990s when they were introduced by LeCun et al. [3]. A first layer when projections to pixel space are allowed. This
number of studies have shown that they are also capable is not the case at deeper levels, and there are few ways
of performing well on more demanding tasks requiring to interpret activity.[8]Gradient descent in picture space is
visual [Link] the ImageNet 2012 classification used to optimize the activation of each unit by locating the
benchmark, (Krizhevsky et al., 2012) demonstrate record- best stimulus.[1] This entails careful setup and provides no
breaking accuracy with their convnet model.[4] The 2nd knowledge of the unit’s invariances. Le et al. (2010), building
place model had error rates of 26.1%, compared to 16.4% on Berkes & Wiskott’s concept, demonstrate how the Hessian
for their model. There are several factors contributing to this of a given unit may be determined numerically around the
renewed interest in convnet models, including the size of optimal response, which Is motivated by the latter’s short-
the training set will grow as more examples are labelled, coming, providing some understanding of invariances.[9] [10]
the GPU implementations will become more efficient, and The issue is that a straightforward quadratic approximation
the regularization strategies will improve, such as Dropout.[4]. does a poor job of capturing the invariances for higher layers
due to their enormous complexity. Contrarily,( Zeiler et al.)
Although these advancements are encouraging, little is method offers a way to reveal which patterns from the training
known about how these complex models work or why they data activate the feature map in a non-parametric approach of
perform so well. A visualization technique is introduced by invariance. (Donahue et al., 2013) Activation patches at higher
(Zeiler et al.) that reveals the input stimulus that excites model layers are visualized as dataset patches. (Zeiler et al.)
individual feature maps across layers. As well as observing visualizations are distinct because they aren’t simply cropping
how the model evolves during training, it allows us to of the input pictures; instead, they’re top-down projections that
diagnose potential errors. In order to return the feature show structures inside each patch that cause a user to think in
activations to the input pixel space, the visualization method new ways about a specific feature map. [11] [1]
 III. V ISUALIZATION T ECHNIQUES The Network Inversion was later used to visualize CNN [23]
A visualization of CNNs has been developed in order to [24] in two variants: (1) Regularizer based Network Inversion:
interpret their overall functioning mechanism. A few repre- It is proposed by Mahendran et al., which uses gradient descent
sentative visualization methods are described here, including with a regularization term to reconstruct the image from each
Deconvolutional Networks (DeconNet), Activation Maxi- layer. (2)Dosovitskiy et al. [24] proposed the UpconvNet-
mization, Network Dissection, and Network Inversion.[2] based Network Inversion: A convolutional neural network
(UpconvNet) is trained to reconstruct the image. Based on
A. Visualization by Activation Maximization the specific activation of one entire layer’s feature maps, both
An activation maximization method (AM) is proposed algorithms primarily strive to reconstruct the original input
to determine what neurons prefer to input into each image. Since there is no need to train a second dedicated
layer. Neurons can learn features based on their preferred network, Regularizer-based Network Inversion is simpler to
input. As a result of this learning process, an input pattern is implement. A Network Inversion based on UpconvNet, on the
synthesized that can cause the maximal activation of a neuron. other hand, is able to visualize more actual data at a higher
This input pattern is synthesized by iterative changing the layer of abstraction.
pixels of the CNN’s input to maximize the neuron’s activation.
C. Visualization by Network Dissection
Erhan et al. proposed the fundamental algorithm behind Convolutional neural networks or multiple neurons can
the AM in 2009 [8]. From the MNIST digit dataset [12], The be correlated with specific semantic concepts by evaluating
Deep Belief Nets[4] and Stacked Denoising AutoEncoders their relationship. The previous section demonstrated how to
were used to visualize the preferred input patterns for the visualize the visual patterns capturing a single neuron or layer
hidden neurons [13]. Using this method, Simonyan et al. by using various visualization techniques. There is still a gap
maximized the activation of neurons in the last layer of CNNs between visually perceptible patterns and semantic concepts
[14]. Similar visualized patterns have also been synthesized that are comprehensible.
by Google for its inception network [15]. In a large-scale
application of the AM, Yosinksi et al. visualized all layers In order to understand color, textures, materials, parts, ob-
of a CNN with arbitrary neurons. There have been a number jects, and scenes, Bau et al. proposed the Network Dissection,
of optimization works conducted recently to improve the in which each convolutional neuron is directly associated with
interpretability and diversity of the visualized patterns. All a particular semantic concept. By seeking neuron responses
of these works have demonstrated that the AM is capable that are strongly correlated with semantic concepts, we can
of interpreting neural interests and identifying hierarchical measure the correlation between neurons and semantic con-
features learned by CNNs. cepts [25]. With the Borden dataset, heterogeneous image data
is labeled with local semantic concepts. In “Figure. 1” We see
This method of visualization reveals that CNNs are able several Broden examples. There are six categories of semantic
to identify faces, wheels, and bottles without our guidance. concepts highlighted with red boxes in “Figure. 1”. There are
Similarly, CNNs offer a hierarchical feature extraction method many semantic categories. For example, the object category
that mimics the hierarchical organization of the visual cortex. includes plants, trains, etc. Each example in “Figure. 1”.
Additionally, this visualization method suggests that neurons has a lower right corner. It is also possible to identify the
extract features in a localized manner rather than a distributed semantic corresponding neuron in “Figure. 1”. Furthermore,
manner, where neurons match patterns based on their local- black masks cover image content not related to the assigned
ization.[1] semantics. Network Dissection is responsible for generating
these black masks.[25]
B. Visualization by Network Inversion
An arbitrary layer can be used to reconstruct an image that
illustrates the CNN layer-level feature for a given input image.
When layer-level activation is examined instead of individual
neuron activation, a comprehensive feature representation of
all activation patterns on a layer is revealed. Accordingly, net-
work inversion-based visualization allows analyzing activation
information at the layer level rather than from the perspective
of a single neuron as in the aforementioned visualization meth-
ods. Before CNNs were used to visualize, Network Inversion
was originally proposed to study traditional computer vision
representations. Histograms of Oriented Gradients (HOGs) are Fig. 1. AlexNet neurons are activated by Broden images.[20]
one type of representation. [16] [17] [18], the Scale Invariant
Feature Transform (SIFT) [19], Local Binary Descriptors The Network Inversion works to gradually establish con-
(LBD)[20], and the Bag of Visual Word Descriptors [21] [22]. nections between semantic concepts and the many different
component levels of CNN. One semantic concept and one [3]. Traditional fully-connected networks make up the top
individual neuron were shown to be correlated by the Network several layers of the network, and a softmax classifier is the
Inversion’s basic algorithm. A single neuron can be attributed topmost layer [1]. The model utilized in several of our tests
to each semantic concept, which was the foundation for is seen in “Figure. 2”.
this correlation [26]. Later, more Network Inversion research
showed that the feature representation could be distributed,
suggesting that a semantic concept might be represented by
a combination of different neurons [27]. Accordingly, Fong
et al. proposed another Network Inversion method, called
Net2Vec, which uses neuron combinations to visualize seman-
tic concepts [28], following the paradigm of[25]. The Net2Vec
Network Inversion method proposed by Fong et al. makes
it possible to see semantic concepts based on combinations
of neurons [25]. Both approaches offer thorough visualization Fig. 2. The architecture of our 8 layer convnet model. The input is a 224 by
outcomes for understanding CNN hidden neurons. 224 crop of a picture (with three color planes). Using a stride of 2 in both x and
y, this is convolved with 96 separate first layer filters, each measuring 7 by 7,
each in the color red. Following that, the obtained feature maps are subjected
D. Visualization by Deconvolutional Network to a rectified linear function (not shown), (ii) pooled (max within 3x3 zones,
Decide which convolutional layer is activated by a specific using stride 2), and (iii) contrast adjusted across feature maps to produce
96 distinct 55 by 55 element feature maps. Layers 2,3,4,5 repeat similar
pattern from a given input image. CNN visualization based on procedures. The last two layers are completely linked and receive vectorized
deconvolutional networks (DeconvNet) describes CNNs from features from the top convolutional layer (6 6 256 = 9216 dimensions) as
the perspective of the input image, as opposed to activation input. A C-way softmax function, where C is the number of classes, makes
up the final layer. The shape of all filters and feature maps is square [1].
maximization. The convolutional layers are activated by
selective patterns in the input image. An image dimension Zeiler et al. use a sizable set of N labeled pictures fx; yg
is projected onto low-dimensional neuronal feature maps to train these models, where the real class is identified by
to reconstruct patterns. Convolutions and pooling layers are the discrete variable label yi. In order to compare yi with
reversed by an unpooling layer in a DeconvNet structure xi, they use a cross-entropy loss function that is suitable for
to accomplish this projection process. DeconvNet-based classifying images. The network’s parameters, such as the
visualization not only analyzes neuronal interests but also filters in the convolutional layers, weight matrices in the fully
demonstrates simple image-level feature analysis. connected layers, and biases, are trained by backpropagating
the derivative of the loss with respect to the parameters present
In most cases, Zeiler et al. lead research related to De- throughout the network and updating the parameters using
convNet structures. DeconvNet was first proposed in [1] with stochastic gradient [Link] section VI provides comprehen-
the objective of reconstructing the natural image through the sive training information[1].
projection of a highly diverse set of low-dimension feature
maps into high dimensions. Using the DeconvNet structure, V. V ISUALIZATION WITH A D ECONVNET
they decomposed an image hierarchically in [1], capturing Interpreting the feature activity in intermediary levels
information at both low-level edges and high-level object parts is necessary to comprehend how a convnet works. By
as well as in between. Eventually, they applied the The method demonstrating what input pattern initially generated a certain
of interpreting CNN hidden features is an effective way to activation in the feature maps, we(Zeiler et. al.) provide a
visualize CNNs using DeconvNet structure. unique method to map these activities back to the input pixel
In this paper I described Visualization by Deconvolutional space. Using a deconvolutional network (deconvnet), they
Network. carry out this mapping (Zeiler et al., 2011). A deconvnet
can be compared to a convnet model that employs the same
IV. A PPROACH building blocks (filtering, pooling) but operates in the other
Throughout the research, we employ typical fully super- direction, mapping pixels to features instead. Deconvnets
vised convnet models according to (LeCun et al., 1989) and were suggested as a method of doing unsupervised learning
(Krizhevsky et al., 2012) [12] [7]. These models convert a in (Zeiler et al., 2011). Here, they serve just as a test for
color 2D input picture (xi) into a probability vector (yi) over a previously trained convnet and are not employed in any
the C distinct classes by means of a number of layers. As an learning capacities[1].
example, the first layer consists of convolution of the output of
the previous layer with learned filters (relu(x) = max(x; 0)); (ii) As shown in “Figure. 3” (Top), a deconvnet is coupled to
passing the responses through a rectified linear function (max each layer of a convnet to provide a continuous path back
pooling over local neighborhoods is optional); (iii) normalizing to the picture pixels. An input picture is first shown to the
the responses across feature maps by using a local contrast convnet, and features are calculated across all of the layers.
operation (optional). See (Krizhevsky et al., 2012) and for Setting all other activations in the layer to zero and feeding
further information on these operations (Jarrett et al., 2009) the feature maps into the associated deconvnet layer allows
us to study a specific convnet activation. In order to rebuild VI. T RAINING D ETAILS
the activity in the layer underneath that gave birth to the In Section 4, a big convnet model will be displayed.
selected activation, we then unpool, (ii) correct, and (iii) filter This section describes such model. The architecture utilized
the data in turn. Then, until input pixel space is reached, this by (Krizhevsky et al., 2012) for ImageNet classification is
is repeated[1]. depicted in “Figure. 2”. One distinction is that our (Zeiler
et al.) model uses dense connections instead of the sparse
connections utilized in Krizhevsky’s layers 3, 4, which
were chosen since the model was spread across two GPUs.
Following examination of the visualizations in “Figure. 7”,
more significant distinctions pertaining to layers 1 and 2 were
made, as detailed in Section VII-A.

The model was trained using the 1.3 million photos from
the ImageNet 2012 training set, which were distributed
across 1000 distinct classes. To prepare each RGB image, the
lowest dimension was increased to 256, the center 256x256
region was cropped. After subtracting the per-pixel mean
(across all pictures), ten distinct sub-crops of size 224 by
224 were employed, with or without horizontal flips. These
Fig. 3. Top: deconvnet and convnet layers (right). The deconvnet reconstructs sub-crops included the corners and the center of the image.
the convnet layer underneath. Bottom: Unpooling in deconvnet utilizing Stochastic gradient descent with a mini-batch size of 128, a
switches that record the local max in each pooling region (colored zones) momentum term of 0.9, and an initial learning rate of 10−2 ,
during convnet pooling. [1]
was used to update the parameters. 10−2 When the validation
error reaches a plateau, we manually anneal the learning
Unpooling: Max pooling is not invertible in the convnet, rate during training. The completely linked layers (6 and 7)
but by storing the maxima in each pooling zone in switch employ dropout (Hinton et al., 2012) at a rate of 0.5. Biases
variables, we may approximate its inverse. Switches are used are set to 0 and all weights are initialized at 10−2 .[4]
in deconvnet during unpooling to place reconstructions from
the layer above appropriately while maintaining stimulus A couple of the first layer filters become more prominent
structure. “Figure. 3” (Bottom) shows the technique. when they are seen during training, as illustrated in “Figure. 8
(a)”. To counteract this, each convolutional layer’s filter whose
Rectification: Relu non-linearities are used by the convnet RMS value exceeds a predetermined radius of 101 has its
to correct the feature maps, making sure they are always RMS value renormalized to this fixed radius. This is essential,
positive. We run the reconstructed signal via a relu non- especially in the model’s first layer because the input photos
linearity in order to get accurate feature reconstructions at approximately fall inside the [-128,128] range. We construct
each layer (which also need to be positive). numerous distinct crops and flips of each training sample,
as in (Krizhevsky et al., 2012), to increase the size of the
training set. Using an implementation based on (Krizhevsky
Filtering: To convolve the feature maps from the preceding
et al., 2012), we ended training after 70 epochs, which on a
layer, the convnet employs learnt filters. The deconvnet inverts
single GTX580 GPU took about 12 days [7] [1].
this by applying transposed copies of the same filters on the
corrected maps rather than the output of the layer below. In VII. C ONVNET V ISUALIZATION
actuality, this entails flipping each filter both horizontally and In this section, the deconvnet model is utilized to display
vertically. the feature activations on the ImageNet validation set by
using the model discussed in Section VI of this article.
A switch setting is used when projecting down from a
higher level based on the maximum pooling in the convnet. Feature Visualization: After training is finished, the
The reconstruction that was obtained from a single activa- model’s feature visualizations are shown in “Figure. 6”.
tion consequently resembled a very small percentage of the However, it provide the top 9 activations rather than just the
original input image. Based on the switch settings used in strongest activation for a particular feature map. The multiple
the input image, the structures were weighted based on their structures that excite a particular feature map are shown by
contribution to feature activation. As a result of the model’s projecting each independently into pixel space, demonstrating
discriminative training, it is implicitly obvious which elements the feature map’s invariance to input deformations. They also
of the input picture are discriminative. Because there is no display the relevant picture patches with these visualizations.
generating process involved, it should be noted that these These differ more from visualizations, which just highlight
projections are not samples from the model[1]. the discriminant structure inside each patch. For example,
in layer 5, row 1, column 2, the patches don’t appear to
have much in common; yet, the visualizations reveal that this
feature map focuses on the grass in the background rather
than the things in the front. [1].

The projections from each layer demonstrate how the

network’s characteristics are organized hierarchically. Corners
and other edge/color conjunctions are handled by layer 2.
Layer 3 features more intricate invariances that capture
comparable textures, such as text (R2,C4) and mesh patterns
(Row 1, Col 1). Significant variety is seen in Layer 4,
however it is more class-specific: dog faces (R1,C1); bird’s
legs (R4,C2). Layer 5 displays whole objects with noticeable
posture change, such as dogs (R1,C11) and keyboards Fig. 5. Examines the model’s vertical translation, scaling, and rotation
invariance (rows a-c respectively). Col. 1: 5 illustrations of transformations in
(R1,C11) (R4)[1]. action. Cols. 2 and 3 show the Euclidean distance between the feature vectors
in layers 1 and 7, respectively, from the original and modified pictures. Col.
Feature Evolution during Training: In “Figure. 4” depicts 4: As a picture is altered, the likelihood that each label is accurate.[1]
the progression of the most significant activation (across all
training instances) inside of a particular feature map when the
map is projected back to pixel space. A shift in the picture
that causes the strongest activation causes abrupt jumps in
appearance. Within a few epochs, the model’s lowest layers
may be observed to converge. However, the development of
the top layers takes place after a sizable number of epochs
(40–50), highlighting the necessity of letting the models train
until they have fully converged.

Fig. 4. Training-driven evolution of a subset of model characteristics.

Different blocks display each layer’s features. It displays a random selection
of characteristics at epochs [1,2,5,10,20,30,40,64]. The graphic depicts the
strongest activation (across all training samples) for a specific feature map
in pixel space. The figure’s color contrast is boosted, and it’s best viewed
electronically.[1]

Feature invariance: In “Figure. 5”, five sample images

with varied degrees of translation, rotation, and scaling are
shown. As a comparison to the untransformed features, the top
and bottom feature vectors of the model are examined. Small
changes have a significant influence in the model’s top feature
layer, where translation and scaling are quasilinear and have a
less dramatic effect. The network output is resilient to scaling
and translation. With the exception of objects with rotational
symmetry, the output is often not rotation-invariant (e.g., the
entertainment center).
Fig. 6. Fully trained model characteristics For layers 2–5, it presents the
A. Architecture Choosing top 9 activation’s in a random subset of validation feature maps projected to
pixel space. The reconstructions aren’t model samples; they’re validation set
Visualization can help you understand how a trained model patterns that activate a feature map. It shows picture patches for each feature
works, and it can also help you choose designs that will map. Notice how each feature map is tightly clustered, how there is more
work. When looking at Krizhevsky et al.’s first and second variation at higher layers, and how distinguishing parts of the picture, like
dog eyes and noses (layer 4, row 1, cols 1), are made bigger. The electronic
levels “Figure. 5 (b) & (d)”, a number of issues become clear. format is preferred.[1]
Since the first layer filters don’t cover enough of the midrange
frequencies, they combine information at very high and very
low frequencies. In addition, the aliasing artifacts brought on
by the first layer convolutions’ use of a long stride 4 are visible
in the second layer visualization. In order to address these
issues, we decreased the size of the first layer filter from
11x11 to 7x7 and (ii) made the convolution stride 2 rather
than 4. According to “Figure. 5 (c) & (d)”, this new design
preserves significantly more information in the first and second
layer features “Figure. 5 (e)”. More significantly, it enhances
classification performance, as seen in Section ImageNet 2012.

Fig. 8. This test example shows how the top five feature maps ((b) & (c))
and classifier output ((d) & (e)) change as gray squares are systematically
drawn over different parts of the scene (1st column). To determine the total
activation for each gray scale position, we record it in one feature map in layer
5 (the one with the strongest response in the unoccluded image). Feature
Fig. 7. (a) Features of the first layer without clipping A single feature map projections for the input image (black square) and projections of the
dominates.(b) The first layer from Krizhevsky et al. (2012). first layer feature map from other images are shown in (c). A dog’s face is the most
characteristics The reduced stride (2 versus 4) and filter size (7x7 vs. 11x11) prominent feature in the first row example. In (b) the blue area in the feature
provide more distinctive and less ”dead” features. (c) Krizhevsky et al.’s map indicates decreased activity when this area is covered-up. It shows the
second layer characteristics (e) Visuals on the second layer. (d)No aliasing probability of a correct class as a function of the location of the gray square
artifacts are seen in these. [1] (d). The probability of pomeranian” drops significantly when the dog’s face
is obscured. E.g. in the first row, the likelihood drops significantly. (f): the
most probable label based on occluder position. It predicts pomeranian” in
most locations, but tennis ball” if the dog’s face, but not the ball, is obscured.
B. Occlusion Sensitivity The wheel is most sensitive to the classifier in the second example since the
Image classification techniques make you wonder if the text on the car is the strongest feature in layer 5. There are multiple objects
in the third example. Since layer 5 uses multiple feature maps, it is sensitive
model really knows where an object is or if it just uses to the dog (blue region in (d)) due to the strongest feature in layer 5 picking
context. “Figure. 8” answers this question by systematically out the faces. [1]
obscuring the input picture with grey squares and monitoring
the classifier’s output. The examples indicate the model local-
izes objects in the scene, since the likelihood of the proper
class lowers when The item is obscured. “Figure. 8” shows
how the top convolution layer’s strongest feature map and its
activity (across all locations) change depending on where the
occluder is. When the occluder covers a visualization zone,
feature map activity drops dramatically. This demonstrates that
the visualization matches the picture structure that drives the
feature map, verifying “Figure. 4” & “Figure. 7”[1].
C. Correspondence Analysis
Deep models are different from many earlier recognition
systems in that there is no explicit way to build a link between Fig. 9. Images used in tests on correspondence Col. 1: The first image Cols.
2, 3, and 4: The nose, right eye, and left eye, respectively, are blocked. Other
different parts of objects in different photos. Deep Models columns show instances of arbitrary occlusions.[1]
may implicitly compute them, though. We take five randomly
selected frontal dog photos and mask out the identical section
of the face in each (e.g. all the left eyes, see Figure 9). A architecture they recommended on the validation set. On the
smaller Hamming distance value means that the same parts of ImageNet 2012 validation set, they get an error rate that is
the same object are closer together in more than one picture within 1% of what is reported.
and that the change caused by the masking operation is more
consistent. For example, closing the left eye always changes The performance of our model after the structural modi-
how the feature is shown[1]. fications described in Section 4.1 (77 filters in layer 1 and
stride 2 convolutions in layers 1 & 2) is next examined. By
VIII. E XPERIMENTS
exceeding their single model result by 1:7%, the model in
A. ImageNet 2012 “Figure. 2” greatly beats the architecture of Krizhevsky et al.
This dataset has 1.3 million training examples, 50 thousand (2012). (test top-5). It acquires a test error of 14:8% when
validation examples, and 100 thousand test examples, it combines several models, which is the best documented
spread across 1,000 categories. Zeiler et al. attempt to performance on this dataset1 (despite only using the 2012
replicate Krizhevsky et al. (2012)’s findings using the precise training set). We(Zeiler et al.) see that this error is nearly half
that of the best non-convnet submission for the ImageNet 2012 these visualizations may be used to diagnose model issues
classification challenge, which had a 26:2% error rate (Gunji and enhance outcomes, such as the outstanding ImageNet
et al., 2012)[29]. 2012 result[7]. Researchers later showed by testing the model
with occlusions that the model is not only sensitive to local
B. Varying ImageNet Model Sizes structures in an image but also to broad scene context. Having
By changing the thickness of the layers or deleting a minimal depth to the network, as opposed to any specific
them altogether, Zeiler et al. investigate the architecture of region, is crucial to the model’s performance, according to an
(Krizhevsky et al., 2012) in the beginning. Each time, the ablation research on it[1].
model is retrained using the updated architecture. Remove
R EFERENCES
the completely linked layers (6,7), and the inaccuracy is
only slightly increased. Given that they include the bulk of [1] M. D. Zeiler and R. Fergus, “Visualizing and under-
model parameters, this is unexpected. Two of the intermediate standing convolutional networks,” in European confer-
convolutional layers being removed also changes the error ence on computer vision, Springer, 2014, pp. 818–833.
rate just little. However, a model with only 4 layers and [2] Z. Qin, F. Yu, C. Liu, and X. Chen, “How convolutional
significantly inferior performance is produced by deleting both neural network see the world-a survey of convolutional
the intermediate convolution layers and the fully linked layers. neural network visualization methods,” arXiv preprint
This would imply that achieving high performance depends on arXiv:1804.11191, 2018.
the model’s entire depth. In “Figure. 2”. Performance is not [3] K. Jarrett, K. Kavukcuoglu, M. Ranzato, and Y. LeCun,
significantly affected by changing the amount of completely “What is the best multi-stage architecture for object
linked layers (similar to the model from Krizhevsky et al., recognition?” In 2009 IEEE 12th international confer-
2012). However, enlarging the middle convolution layers can ence on computer vision, IEEE, 2009, pp. 2146–2153.
result in a useful speed improvement. However, expanding the [4] G. E. Hinton, N. Srivastava, A. Krizhevsky, I. Sutskever,
completely linked layers while concurrently increasing these and R. R. Salakhutdinov, “Improving neural networks
leads to over-fitting[7][1]. by preventing co-adaptation of feature detectors,” arXiv
preprint arXiv:1207.0580, 2012.
C. Feature Generalization [5] G. E. Hinton, S. Osindero, and Y.-W. Teh, “A fast learn-
The studies above demonstrate the significance of the ing algorithm for deep belief nets,” Neural computation,
convolutional component of the Zeiler et al. ImageNet model vol. 18, no. 7, pp. 1527–1554, 2006.
in achieving cutting-edge performance. The visualizations [6] P. Vincent, H. Larochelle, Y. Bengio, and P.-A. Man-
in ”Figure. 6”, which display the intricate invariances learnt zagol, “Extracting and composing robust features with
in the convolutional layers, support this. The capacity of denoising autoencoders,” in Proceedings of the 25th
these feature extraction layers to generalize to additional international conference on Machine learning, 2008,
datasets, including Caltech-101 , Caltech-256 and PASCAL pp. 1096–1103.
VOC 2012, is presently being investigated [30] [31]. To do [7] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Im-
this, they train a new softmax classifier (with the necessary agenet classification with deep convolutional neural
number of classes) on top of their ImageNet-trained model, networks,” Communications of the ACM, vol. 60, no. 6,
keeping layers 1 through 7 fixed. They accomplish this pp. 84–90, 2017.
using the training pictures from the newly acquired dataset. [8] D. Erhan, Y. Bengio, A. Courville, and P. Vincent,
The softmax can be trained fast from a limited number of “Visualizing higher-layer features of a deep network,”
instances, as is the case for some datasets, because it has a University of Montreal, vol. 1341, no. 3, p. 1, 2009.
small number of parameters. [9] P. Berkes and L. Wiskott, “On the analysis and interpre-
tation of inhomogeneous quadratic forms as receptive
Before retraining the Imagenet models, they(Zeiler et al.) fields,” Neural computation, vol. 18, no. 8, pp. 1868–
identified these few overlap images with normalized corre- 1895, 2006.
lation and removed them from the Imagenet training set to [10] Y. LeCun, B. Boser, J. S. Denker, et al., “Backprop-
prevent train/test contamination.[1]. agation applied to handwritten zip code recognition,”
Neural computation, vol. 1, no. 4, pp. 541–551, 1989.
IX. C ONCLUSION [11] J. Donahue, Y. Jia, O. Vinyals, et al., “Decaf: A
In various methods, researchers investigated large convo- deep convolutional activation feature for generic visual
lutional neural network models that have been trained for recognition,” in International conference on machine
image categorization. First, researchers offered a fresh method learning, PMLR, 2014, pp. 647–655.
for visualizing the model’s activities. This demonstrates that [12] Y. LeCun, “The mnist database of handwritten digits,”
the traits are not random, puzzling patterns. Instead, as we [Link] lecun. com/exdb/mnist/, 1998.
move up the layers, these exhibit numerous qualities that
are intuitively appealing, including compositionality, rising
invariance, and class discrimination. It also demonstrated how
[13] P. Vincent, H. Larochelle, I. Lajoie, Y. Bengio, P.-A. International Journal of Computer Vision, vol. 126,
Manzagol, and L. Bottou, “Stacked denoising autoen- no. 5, pp. 476–494, 2018.
coders: Learning useful representations in a deep net- [27] P. Agrawal, R. Girshick, and J. Malik, “Analyzing the
work with a local denoising criterion.,” Journal of performance of multilayer neural networks for object
machine learning research, vol. 11, no. 12, 2010. recognition,” in European conference on computer vi-
[14] K. Simonyan, A. Vedaldi, and A. Zisserman, “Deep sion, Springer, 2014, pp. 329–344.
inside convolutional networks: Visualising image clas- [28] R. Fong and A. Vedaldi, “Net2vec: Quantifying and
sification models and saliency maps,” arXiv preprint explaining how concepts are encoded by filters in deep
arXiv:1312.6034, 2013. neural networks,” in Proceedings of the IEEE confer-
[15] A. Mordvintsev, C. Olah, and M. Tyka, “Inceptionism: ence on computer vision and pattern recognition, 2018,
Going deeper into neural networks,” 2015. pp. 8730–8738.
[16] N. Dalal and B. Triggs, “Histograms of oriented gra- [29] A. G. Howard, “Some improvements on deep convolu-
dients for human detection,” in 2005 IEEE computer tional neural network based image classification,” arXiv
society conference on computer vision and pattern preprint arXiv:1312.5402, 2013.
recognition (CVPR’05), Ieee, vol. 1, 2005, pp. 886–893. [30] G Griffin, A Holub, and P Perona, “The caltech 256
[17] P. F. Felzenszwalb, R. B. Girshick, D. McAllester, and (technical report),” Caltech Computation and Neural
D. Ramanan, “Object detection with discriminatively Systems, 2006.
trained part-based models,” IEEE transactions on pat- [31] L. Fei-Fei, R. Fergus, and P. Perona, “One-shot learn-
tern analysis and machine intelligence, vol. 32, no. 9, ing of object categories,” IEEE transactions on pat-
pp. 1627–1645, 2010. tern analysis and machine intelligence, vol. 28, no. 4,
[18] R. B. Girshick, P. F. Felzenszwalb, and D. McAllester, pp. 594–611, 2006.
“Discriminatively trained deformable part models, re-
lease 5,” 2012.
[19] D. G. Lowe, “Distinctive image features from scale-
invariant keypoints,” International journal of computer
vision, vol. 60, no. 2, pp. 91–110, 2004.
[20] E. d’Angelo, A. Alahi, and P. Vandergheynst, “Be-
yond bits: Reconstructing images from local binary
descriptors,” in Proceedings of the 21st International
Conference on Pattern Recognition (ICPR2012), IEEE,
2012, pp. 935–938.
[21] G. Csurka, C. Dance, L. Fan, J. Willamowski, and C.
Bray, “Visual categorization with bags of keypoints,”
in Workshop on statistical learning in computer vision,
ECCV, Prague, vol. 1, 2004, pp. 1–2.
[22] J. Sivic and A. Zisserman, “Video google: A text
retrieval approach to object matching in videos,” in
Computer Vision, IEEE International Conference on,
IEEE Computer Society, vol. 3, 2003, pp. 1470–1470.
[23] A. Mahendran and A. Vedaldi, “Understanding deep im-
age representations by inverting them,” in Proceedings
of the IEEE conference on computer vision and pattern
recognition, 2015, pp. 5188–5196.
[24] N. Mayer, E. Ilg, P. Hausser, et al., “A large dataset to
train convolutional networks for disparity, optical flow,
and scene flow estimation,” in Proceedings of the IEEE
conference on computer vision and pattern recognition,
2016, pp. 4040–4048.
[25] D. Bau, B. Zhou, A. Khosla, A. Oliva, and A. Torralba,
“Network dissection: Quantifying interpretability of
deep visual representations,” in Proceedings of the IEEE
conference on computer vision and pattern recognition,
2017, pp. 6541–6549.
[26] A. Gonzalez-Garcia, D. Modolo, and V. Ferrari, “Do se-
mantic parts emerge in convolutional neural networks?”

View publication stats