Analyzing the Performance of Multilayer Neural

Networks for Object Recognition

Pulkit Agrawal, Ross Girshick, and Jitendra Malik

University of California, Berkeley

{pulkitag,rbg,malik}@eecs.berkeley.edu

Abstract. In the last two years, convolutional neural networks (CNNs)

have achieved an impressive suite of results on standard recognition
datasets and tasks. CNN-based features seem poised to quickly replace
engineered representations, such as SIFT and HOG. However, compared
to SIFT and HOG, we understand much less about the nature of the
features learned by large CNNs. In this paper, we experimentally probe
several aspects of CNN feature learning in an attempt to help practition-
ers gain useful, evidence-backed intuitions about how to apply CNNs to
computer vision problems.

Keywords: convolutional neural networks, object recognition, empiri-

cal analysis.

1 Introduction
Over the last two years, a sequence of results on benchmark visual recognition
tasks has demonstrated that convolutional neural networks (CNNs) [6,14,18] will
likely replace engineered features, such as SIFT [15] and HOG [2], for a wide
variety of problems. This sequence started with the breakthrough ImageNet [3]
classification results reported by Krizhevsky et al. [11]. Soon after, Donahue et
al. [4] showed that the same network, trained for ImageNet classification, was an
effective blackbox feature extractor. Using CNN features, they reported state-
of-the-art results on several standard image classification datasets. At the same
time, Girshick et al. [7] showed how the network could be applied to object
detection. Their system, called R-CNN, classifies object proposals generated by
a bottom-up grouping mechanism (e.g., selective search [23]). Since detection
training data is limited, they proposed a transfer learning strategy in which the
CNN is first pre-trained, with supervision, for ImageNet classification and then
fine-tuned on the small PASCAL detection dataset [5]. Since this initial set of
results, several other papers have reported similar findings on a wider range of
tasks (see, for example, the outcomes reported by Razavian et al. in [17]).
Feature transforms such as SIFT and HOG afford an intuitive interpretation
as histograms of oriented edge filter responses arranged in spatial blocks. How-
ever, we have little understanding of what visual features the different layers of
a CNN encode. Given that rich feature hierarchies provided by CNNs are likely
to emerge as the prominent feature extractor for computer vision models over

D. Fleet et al. (Eds.): ECCV 2014, Part VII, LNCS 8695, pp. 329–344, 2014.

c Springer International Publishing Switzerland 2014
330 P. Agrawal, R. Girshick, and J. Malik

the next few years, we believe that developing such an understanding is an inter-
esting scientific pursuit and an essential exercise that will help guide the design
of computer vision methods that use CNNs. Therefore, in this paper we study
several aspects of CNNs through an empirical lens.

1.1 Summary of Findings

Effects of Fine-Tuning and Pre-training. Girshick et al. [7] showed that

supervised pre-training and fine-tuning are effective when training data is scarce.
However, they did not investigate what happens when training data becomes
more abundant. We show that it is possible to get good performance when
training R-CNN from a random initialization (i.e., without ImageNet supervised
pre-training) with a reasonably modest amount of detection training data (37k
ground truth bounding boxes). However, we also show that in this data regime,
supervised pre-training is still beneficial and leads to a large improvement in
detection performance. We show similar results for image classification, as well.

ImageNet Pre-training Does not Overfit. One concern when using super-
vised pre-training is that achieving a better model fit to ImageNet, for example,
might lead to higher generalization error when applying the learned features
to another dataset and task. If this is the case, then some form of regulariza-
tion during pre-training, such as early stopping, would be beneficial. We show
the surprising result that pre-training for longer yields better results, with di-
minishing returns, but does not increase generalization error. This implies that
fitting the CNN to ImageNet induces a general and portable feature represen-
tation. Moreover, the learning process is well behaved and does not require ad
hoc regularization in the form of early stopping.

Grandmother Cells and Distributed Codes. We do not have a good un-

derstanding of mid-level feature representations in multilayer networks. Recent
work on feature visualization, (e.g., [13,26]) suggests that such networks might
consist mainly of “grandmother” cells [1,16]. Our analysis shows that the rep-
resentation in intermediate layers is more subtle. There are a small number of
grandmother-cell-like features, but most of the feature code is distributed and
several features must fire in concert to effectively discriminate between classes.

Importance of Feature Location and Magnitude. Our final set of experi-

ments investigates what role a feature’s spatial location and magnitude plays in
image classification and object detection. Matching intuition, we find that spatial
location is critical for object detection, but matters little for image classification.
More surprisingly, we find that feature magnitude is largely unimportant. For
example, binarizing features (at a threshold of 0) barely degrades performance.
This shows that sparse binary features, which are useful for large-scale image
retrieval [8,24], come “for free” from the CNN’s representation.
 Analyzing the Performance of Multilayer Neural Networks 331

2 Experimental Setup

2.1 Datasets and Tasks

In this paper, we report experimental results using several standard datasets
and tasks, which we summarize here.

Image Classification. For the task of image classification we consider two

datasets, the first of which is PASCAL VOC 2007 [5]. We refer to this dataset
and task by “PASCAL-CLS”. Results on PASCAL-CLS are reported using the
standard average precision (AP) and mean average precision (mAP) metrics.
PASCAL-CLS is fairly small-scale with only 5k images for training, 5k im-
ages for testing, and 20 object classes. Therefore, we also consider the medium-
scale SUN dataset [25], which has around 108k images and 397 classes. We refer
to experiments on SUN by “SUN-CLS”. In these experiments, we use a non-
standard train-test split since it was computationally infeasible to run all of
our experiments on the 10 standard subsets proposed by [25]. Instead, we ran-
domly split the dataset into three parts (train, val, and test) using 50%, 10%
and 40% of the data, respectively. The distribution of classes was uniform across
all the three sets. We emphasize that results on these splits are only used to sup-
port investigations into properties of CNNs and not for comparing against other
scene-classification methods in the literature. For SUN-CLS, we report 1-of-397
classification accuracy averaged over all classes, which is the standard metric
for this dataset. For select experiments we report the error bars in performance
as mean ± standard deviation in accuracy over 5 runs (it was computationally
infeasible to compute error bars for all experiments). For each run, a different
random split of train, val, and test sets was used.

Object Detection. For the task of object detection we use PASCAL VOC
2007. We train using the trainval set and test on the test set. We refer to this
dataset and task by “PASCAL-DET”. PASCAL-DET uses the same set of im-
ages as PASCAL-CLS. We note that it is standard practice to use the 2007
version of PASCAL VOC for reporting results of ablation studies and hyperpa-
rameter sweeps. We report performance on PASCAL-DET using the standard
AP and mAP metrics. In some of our experiments we use only the ground-
truth PASCAL-DET bounding boxes, in which case we refer to the setup by
“PASCAL-DET-GT”.
In order to provide a larger detection training set for certain experiments,
we also make use of the “PASCAL-DET+DATA” dataset, which we define as
including VOC 2007 trainval union with VOC 2012 trainval. The VOC 2007
test set is still used for evaluation. This dataset contains approximately 37k
labeled bounding boxes, which is roughly three times the number contained in
PASCAL-DET.
332 P. Agrawal, R. Girshick, and J. Malik

2.2 Network Architecture and Layer Nomenclature

All of our experiments use a single CNN architecture. This architecture is the
Caffe [9] implementation of the network proposed by Krizhevsky et al. [11]. The
layers of the CNN are organized as follows. The first two are subdivided into
four sublayers each: convolution (conv), max(x, 0) rectifying non-linear units
(ReLUs), max pooling, and local response normalization (LRN). Layers 3 and
4 are composed of convolutional units followed by ReLUs. Layer 5 consists of
convolutional units, followed by ReLUs and max pooling. The last two layers are
fully connected (fc). When we refer to conv-1, conv-2, and conv-5 we mean the
output of the max pooling units following the convolution and ReLU operations
(also following LRN when applicable).1 For layers conv-3, conv-4, fc-6, and fc-7
we mean the output of ReLU units.

2.3 Supervised Pre-training and Fine-Tuning

Training a large CNN on a small dataset often leads to catastrophic overfitting.
The idea of supervised pre-training is to use a data-rich auxiliary dataset and
task, such as ImageNet classification, to initialize the CNN parameters. The
CNN can then be used on the small dataset, directly, as a feature extractor
(as in [4]). Or, the network can be updated by continued training on the small
dataset, a process called fine-tuning.
For fine-tuning, we follow the procedure described in [7]. First, we remove the
CNN’s classification layer, which was specific to the pre-training task and is not
reusable. Next, we append a new randomly initialized classification layer with
the desired number of output units for the target task. Finally, we run stochastic
gradient descent (SGD) on the target loss function, starting from a learning rate
set to 0.001 (1/10-th the initial learning rate used for training the network for
ImageNet classification). This choice was made to prevent clobbering the CNN’s
initialization to control overfitting. At every 20,000 iterations of fine-tuning we
reduce the learning rate by a factor of 10.

3 The Effects of Fine-Tuning and Pre-training on CNN

Performance and Parameters
The results in [7] (R-CNN) show that supervised pre-training for ImageNet clas-
sification, followed by fine-tuning for PASCAL object detection, leads to large
gains over directly using features from the pre-trained network (without fine-
tuning). However, [7] did not investigate three important aspects of fine-tuning:
(1) What happens if we train the network “from scratch” (i.e., from a random
initialization) on the detection data? (2) How does the amount of fine-tuning
data change the picture? and (3) How does fine-tuning alter the network’s pa-
rameters? In this section, we explore these questions on object detection and
image classification datasets.
1
Note that this nomenclature differs slightly from [7].
 Analyzing the Performance of Multilayer Neural Networks 333

Table 1. Comparing the performance of CNNs trained from scratch, pre-trained on

ImageNet, and fine-tuned. PASCAL-DET+DATA includes additional data from VOC
2012 trainval. (Bounding-box regression was not used for detection results.)
SUN-CLS PASCAL-DET PASCAL-DET+DATA
scratch pre-train fine-tune scratch pre-train fine-tune scratch pre-train fine-tune
40.4 ± 0.2 53.1 ± 0.2 56.8 ± 0.2 40.7 45.5 54.1 52.3 45.5 59.2

3.1 Effect of Fine-Tuning on CNN Performance

The main results of this section are presented in Table 1. First, we focus on the
detection experiments, which we implemented using the open source R-CNN
code. All results use features from layer fc-7.
Somewhat surprisingly, it’s possible to get reasonable results (40.7% mAP)
when training the CNN from scratch using only the training data from VOC
2007 trainval (13k bounding box annotations). However, this is still worse than
using the pre-trained network, directly, without fine-tuning (45.5%). Even more
surprising is that when the VOC 2007 trainval data is augmented with VOC
2012 data (an additional 25k bounding box annotations), we are able to achieve
a mAP of 52.3% from scratch. This result is almost as good as the performance
achieved by pre-training on ImageNet and then fine-tuning on VOC 2007 train-
val (54.1% mAP). These results can be compared to the 30.5% mAP obtained
by DetectorNet [21], a recent detection system based on the same network ar-
chitecture, which was trained from scratch on VOC 2012 trainval.
Next, we ask if ImageNet pre-training is still useful in the PASCAL-DET
+DATA setting? Here we see that even though it’s possible to get good per-
formance when training from scratch, pre-training still helps considerably. The
final mAP when fine-tuning with the additional detection data is 59.2%, which
is 5 percentage points higher than the best result reported in [7] (both without
bounding-box regression). This result suggests that R-CNN performance is not
data saturated and that simply adding more detection training data without any
other changes may substantially improve results.
We also present results for SUN image classification. Here we observe a similar
trend: reasonable performance is achievable when training from scratch, however
initializing from ImageNet and then fine-tuning yields significantly better per-
formance.

3.2 Effect of Fine-Tuning on CNN Parameters

We have provided additional evidence that fine-tuning a discriminatively pre-
trained network is very effective in terms of task performance. Now we look
inside the network to see how fine-tuning changes its parameters.
To do this, we define a way to measure the class selectivity of a set of filters.
Intuitively, we use the class-label entropy of a filter given its activations, above
a threshold, on a set of images. Since this measure is entropy-based, a low value
334 P. Agrawal, R. Girshick, and J. Malik

360

Measure of Class Selectivity (based on Entropy)

340

320

conv−1
300
conv−2
conv−3
conv−4
280 conv−5
fc−6
fc−7
260
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Fraction of Filters

Fig. 1. PASCAL object class selectivity plotted against the fraction of filters, for each
layer, before fine-tuning (dash-dot line) and after fine-tuning (solid line). A lower value
indicates greater class selectivity. Although layers become more discriminative as we go
higher up in the network, fine-tuning on limited data (PASCAL-DET) only significantly
affects the last two layers (fc-6 and fc-7).

Table 2. Comparison in performance when fine-tuning the entire network (ft) versus
only fine-tuning the fully-connected layers (fc-ft)

SUN-CLS PASCAL-DET PASCAL-DET+DATA

ft fc-ft ft fc-ft ft fc-ft
56.8 ± 0.2 56.2 ± 0.1 54.1 53.3 59.2 56.0

indicates that a filter is highly class selective, while a large value indicates that
a filter fires regardless of class. The precise definition of this measure is given in
the Appendix.
In order to summarize the class selectivity for a set of filters, we sort them
from the most selective to least selective and plot the average selectivity of the
first k filters while sweeping k down the sorted list. Figure 1 shows the class
selectivity for the sets of filters in layers 1 to 7 before and after fine-tuning (on
VOC 2007 trainval). Selectivity is measured using the ground truth boxes from
PASCAL-DET-GT instead of a whole-image classification task to ensure that
filter responses are a direct result of the presence of object categories of interest
and not correlations with image background.
Figure 1 shows that class selectivity increases from layer 1 to 7 both with
and without fine-tuning. It is interesting to note that entropy changes due to
fine-tuning are only significant for layers 6 and 7. This observation indicates
that fine-tuning only layers 6 and 7 may suffice for achieving good performance
when fine-tuning data is limited. We tested this hypothesis on SUN-CLS and
PASCAL-DET by comparing the performance of a fine-tuned network (ft) with
 Analyzing the Performance of Multilayer Neural Networks 335

Table 3. Performance variation (% mAP) on PASCAL-CLS as a function of pre-

training iterations on ImageNet. The error bars for all columns are similar to the one
reported in the 305k column.

layer 5k 15k 25k 35k 50k 95k 105k 195k 205k 305k
conv-1 23.0 24.3 24.4 24.5 24.3 24.8 24.7 24.4 24.4 24.4 ± 0.5
conv-2 33.7 40.4 40.9 41.8 42.7 43.2 44.0 45.0 45.1 45.1 ± 0.7
conv-3 34.2 46.8 47.0 48.2 48.6 49.4 51.6 50.7 50.9 50.5 ± 0.6
conv-4 33.5 49.0 48.7 50.2 50.7 51.6 54.1 54.3 54.4 54.2 ± 0.7
conv-5 33.0 53.4 55.0 56.8 57.3 59.2 63.5 64.9 65.5 65.6 ± 0.3
fc-6 34.2 59.7 62.6 62.7 63.5 65.6 69.3 71.3 71.8 72.1 ± 0.3
fc-7 30.9 61.3 64.1 65.1 65.9 67.8 71.8 73.4 74.0 74.3 ± 0.3

a network which was fine-tuned by only updating the weights of fc-6 and fc-7
(fc-ft). These results, in Table 2, show that with small amounts of data, fine-
tuning amounts to “rewiring” the fully connected layers. However, when more
fine-tuning data is available (PASCAL-DET+DATA), there is still substantial
benefit from fine-tuning all network parameters.

3.3 Effect of Pre-training on CNN Parameters

There is no single image dataset that fully captures the variation in natural
images. This means that all datasets, including ImageNet, are biased in some
way. Thus, there is a possibility that pre-training may eventually cause the CNN
to overfit and consequently hurt generalization performance [22]. To understand
if this happens, in the specific case of ImageNet pre-training, we investigated the
effect of pre-training time on generalization performance both with and without
fine-tuning. We find that pre-training for longer improves performance. This is
surprising, as it shows that fitting more to ImageNet leads to better performance
when moving to the other datasets that we evaluated.
We report performance on PASCAL-CLS as a function of pre-training time,
without fine-tuning, in Table 3. Notice that more pre-training leads to better
performance. By 15k and 50k iterations all layers are close to 80% and 90%
of their final performance (5k iterations is only ∼1 epoch). This indicates that
training required for generalization takes place quite quickly. Figure 2 shows
conv-1 filters after 5k, 15k, and 305k iterations and reinforces this observation.
Further, notice from Table 3 that conv-1 trains first and the higher the layer
is the more time it takes to converge. This suggests that a CNN, trained with
backpropagation, converges in a layer-by-layer fashion. Table 4 shows the inter-
action between varied amounts of pre-training time and fine-tuning on SUN-CLS
and PASCAL-DET. Here we also see that more pre-training prior to fine-tuning
leads to better performance.
336 P. Agrawal, R. Girshick, and J. Malik

(a) 5k Iterations (b) 15k Iterations (c) 305k Iterations

Fig. 2. Evolution of conv-1 filters with time. After just 15k iterations, these filters
closely resemble their converged state.

Table 4. Performance variation on SUN-CLS and PASCAL-DET using features from

a CNN pre-trained for different numbers of iterations and fine-tuned for a fixed number
of iterations (40k for SUN-CLS and 70k for PASCAL-DET)

50k 105k 205k 305k

SUN-CLS 53.0 ± 0.2 54.6 ± 0.1 56.3 ± 0.2 56.6 ± 0.2
PASCAL-DET 50.2 52.6 55.3 55.42

4 Are there Grandmother Cells in CNNs?

Neuroscientists have conjectured that cells in the human brain which only re-
spond to very specific and complex visual stimuli (such as the face of one’s grand-
mother) are involved in object recognition. These neurons are often referred to as
grandmother cells (GMC) [1,16]. Proponents of artificial neural networks have
shown great interest in reporting the presence of GMC-like filters for specific
object classes in their networks (see, for example, the cat filter reported in [13]).
The notion of GMC like features is also related to standard feature encodings
for image classification. Prior to the work of [11], the dominant approaches for
image and scene classification were based on either representing images as a bag
of local descriptors (BoW), such as SIFT (e.g., [12]), or by first finding a set
of mid-level patches [10,20] and then encoding images in terms of them. The
problem of finding good mid-level patches is often posed as a search for a set
of high-recall discriminative templates. In this sense, mid-level patch discovery
is the search for a set of GMC templates. The low-level BoW representation, in
contrast, is a distributed code in the sense that a single feature by itself is not dis-
criminative, but a group of features taken together is. This makes it interesting
to investigate the nature of mid-level CNN features such as conv-5.
For understanding these feature representations in CNNs, [19,26] recently pre-
sented methods for finding locally optimal visual inputs for individual filters.
However, these methods only find the best, or in some cases top-k, visual inputs
2
A network pre-trained from scratch, which was different from the one used in Section
3.1, was used to obtain these results. The difference in performance is not significant.
 Analyzing the Performance of Multilayer Neural Networks 337

aeroplane bicycle bird boat

1

0.5

bottle bus car cat

1

0.5

chair cow diningtable dog

Precision

0.5

horse motorbike person pottedplant

1

0.5

sheep sofa train tvmonitor

1

0.5

0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

Recall

Fig. 3. The precision-recall curves for the top five (based on AP) conv-5 filter responses
on PASCAL-DET-GT. Curves in red and blue indicate AP for fine-tuned and pre-
trained networks, respectively. The dashed black line is the performance of a random
filter. For most classes, precision drops significantly even at modest recall values. There
are GMC filters for classes such as bicycle, person, car, cat.

that activate a filter, but do not characterize the distribution of images that
cause an individual filter to fire above a certain threshold. For example, if it is
found that the top-10 visual inputs for a particular filter are cats, it remains
unclear what is the response of the filter to other images of cats. Thus, it is not
possible to make claims about presence of GMC like filters for cat based on such
analysis. A GMC filter for the cat class, is one that fires strongly on all cats and
nothing else. This criteria can be expressed as a filter that has high precision and
high recall. That is, a GMC filter for class C is a filter that has a high average
precision (AP) when tasked with classifying inputs from class C versus inputs
from all other classes.
First, we address the question of finding GMC filters by computing the AP of
individual filters (Section 4.1). Next, we measure how distributed are the feature
representations (Section 4.2). For both experiments we use features from layer
conv-5, which consists of responses of 256 filters in a 6 × 6 spatial grid. Using
max pooling, we collapse the spatial grid into a 256-D vector, so that for each
filter we have a single response per image (in Section 5.1 we show that this
transformation causes only a small drop in task performance).

4.1 Finding Grandmother Cells

For each filter, its AP value is calculated for classifying images using class labels
and filter responses to object bounding boxes from PASCAL-DET-GT. Then,
338 P. Agrawal, R. Girshick, and J. Malik

aeroplane bicycle bird boat

1
0.5
Fraction of complete performance

bottle bus car cat

1
0.5
0

chair cow diningtable dog

1
0.5
0

horse motorbike person pottedplant

1
0.5
0

sheep sofa train tvmonitor

1
0.5
0
0 50 100 150 200 250 300 0 50 100 150 200 250 300 0 50 100 150 200 250 300 0 50 100 150 200 250 300

Number of conv-5 filters

Fig. 4. The fraction of complete performance on PASCAL-DET-GT achieved by conv-5

filter subsets of different sizes. Complete performance is the AP computed by consider-
ing responses of all the filters. Notice, that for a few classes such as person and bicycle
only a few filters are required, but for most classes substantially more filters are needed,
indicating a distributed code.

for each class we sorted filters in decreasing order of their APs. If GMC filters
for this class exist, they should be the top ranked filters in this sorted list. The
precision-recall curves for the top-five conv-5 filters are shown in Figure 3. We
find that GMC-like filters exist for only for a few classes, such as bicycle, person,
cars, and cats.

4.2 How Distributed are the Feature Representations?

In addition to visualizing the AP curves of individual filters, we measured the

number of filters required to recognize objects of a particular class. Feature
selection was performed to construct nested subsets of filters, ranging from a
single filter to all filters, using the following greedy strategy. First, separate linear
SVMs were trained to classify object bounding boxes from PASCAL-DET-GT
using conv-5 responses. For a given class, the 256 dimensions of the learnt weight
vector (w) is in direct correspondence with the 256 conv-5 filters. We used the
magnitude of the i-th dimension of w to rank the importance of the i-th conv-5
filter for discriminating instances of this class. Next, all filters were sorted using
these magnitude values. Each subset of filters was constructed by taking the top-
k filters from this list.3 For each subset, a linear SVM was trained using only the
responses of filters in that subset for classifying the class under consideration.
3
We used values of k ∈ {1, 2, 3, 5, 10, 15, 20, 25, 30, 35, 40, 45, 50, 80, 100, 128, 256}.
 Analyzing the Performance of Multilayer Neural Networks 339

Table 5. Number of filters required to achieve 50% or 90% of the complete perfor-
mance on PASCAL-DET-GT using a CNN pre-trained on ImageNet and fine-tuned for
PASCAL-DET using conv-5 features
perf. aero bike bird boat bottle bus car cat chair cow table dog horse mbike person plant sheep sofa train tv
pre-train 50% 15 3 15 15 10 10 3 2 5 15 15 2 10 3 1 10 20 25 10 2
fine-tune 50% 10 1 20 15 5 5 2 2 3 10 15 3 15 10 1 5 15 15 5 2
pre-train 90% 40 35 80 80 35 40 30 20 35 100 80 30 45 40 15 45 50 100 45 25
fine-tune 90% 35 30 80 80 30 35 25 20 35 50 80 35 30 40 10 35 40 80 40 20

Fig. 5. The set overlap between the 50 most discriminative conv-5 filters for each class
determined using PASCAL-DET-GT. Entry (i, j) of the matrix is the fraction of top-50
filters class i has in common with class j (Section 4.2). Chance is 0.195. There is little
overlap, but related classes are more likely to share filters.

The variation in performance with the number of filters is shown in

Figure 4. Table 5 lists the number of filters required to achieve 50% and 90% of
the complete performance. For classes such as persons, cars, and cats relatively
few filters are required, but for most classes around 30 to 40 filters are required
to achieve at least 90% of the full performance. This indicates that the conv-5
feature representation is distributed and there are GMC-like filters for only a few
classes. Results using layer fc-7 are presented in the the supplementary material.
We also find that after fine-tuning, slightly fewer filters are required to achieve
performance levels similar to a pre-trained network.
Next, we estimated the extent of overlap between the filters used for dis-
criminating between different classes. For each class i, we selected the 50 most
discriminative filters (out of 256) and stored the selected filter indices in the set
Si . The extent of overlap between class i and j was evaluated by |Si ∩ Sj |/N ,
where N = |Si | = |Sj | = 50. The results are visualized in Figure 5. It can be
340 P. Agrawal, R. Girshick, and J. Malik

IHDWXUHPDS ELQDUL]H VSVKXIIOH VSPD[

        

)LOWHU          

        

        

)LOWHU          

        

        

)LOWHU1          

        

D E F G

Fig. 6. Illustrations of ablations of feature activation spatial and magnitude informa-

tion. See Sections 5.1 and 5.2 for details.

Table 6. Percentage non-zeros (sparsity) in filter responses of CNN

conv-1 conv-2 conv-3 conv-4 conv-5 fc-6 fc-7
87.5 ± 4.4 44.5 ± 4.4 31.8 ± 2.4 32.0 ± 2.7 27.7 ± 5.0 16.1 ± 3.0 21.6 ± 4.9

seen that different classes use different subsets of conv-5 filters and there is little
overlap between classes. This further indicates that intermediate representations
in the CNN are distributed.

5 Untangling Feature Magnitude and Location

The convolutional layers preserve the coarse spatial layout of the network’s in-
put. By layer conv-5, the original 227 × 227 input image has been progressively
downsampled to 6 × 6. This feature map is also sparse due to the max(x, 0)
non-linearities used in the network (conv-5 is roughly 27% non-zero; sparsity
statistics for all layers are given in Table 6). Thus, a convolutional layer en-
codes information in terms of (1) which filters have non-zero responses, (2) the
magnitudes of those responses, and (3) their spatial layout. In this section, we
experimentally analyze the role of filter response magnitude and spatial location
by looking at ablation studies on classification and detection tasks.

5.1 How Important is Filter Response Magnitude?

We can asses the importance of magnitude by setting each filter response x to 1
if x > 0 and to 0 otherwise. This binarization is performed prior to using the re-
 Analyzing the Performance of Multilayer Neural Networks 341

sponses as features in a linear classifier and leads to loss of information contained

in the magnitude of response while still retaining information about which filters
fired and where they fired. In Tables 7 and 8 we show that binarization leads to
a negligible performance drop for both classification and detection.
For the fully-connected layers (fc-6 and fc-7) PASCAL-CLS performance is
nearly identical before and after binarization. This is a non-trivial property
since transforming traditional computer vision features into short (or sparse)
binary codes is an active research area. Such codes are important for practical
applications in large-scale image retrieval and mobile image analysis [8,24]. Here
we observe that sparse binary codes come essentially “for free” when using the
representations learned in the fully-connected layers.

5.2 How Important is Response Location?

Now we remove spatial information from filter responses while retaining informa-
tion about their magnitudes. We consider two methods for ablating spatial infor-
mation from features computed by the convolutional layers (the fully-connected
layers do not contain explicit spatial information).
The first method (“sp-max”) simply collapses the p × p spatial map into a sin-
gle value per feature channel by max pooling. The second method (“sp-shuffle”)
retains the original distribution of feature activation values, but scrambles spa-
tial correlations between columns of feature channels. To perform sp-shuffle, we
permute the spatial locations in the p × p spatial map. This permutation is
performed independently for each network input (i.e., different inputs undergo
different permutations). Columns of filter responses in the same location move
together, which preserves correlations between features within each (shuffled)
spatial location. These transformations are illustrated in Figure 6.

Table 7. Effect of location and magnitude feature ablations on PASCAL-CLS

layer no ablation (mAP) binarize (mAP) sp-shuffle (mAP) sp-max (mAP)

conv-1 25.1 ± 0.5 17.7 ± 0.2 15.1 ± 0.3 25.4 ± 0.5

conv-2 45.3 ± 0.5 43.0 ± 0.6 32.9 ± 0.7 40.1 ± 0.3
conv-3 50.7 ± 0.6 47.2 ± 0.6 41.0 ± 0.8 54.1 ± 0.5
conv-4 54.5 ± 0.7 51.5 ± 0.7 45.2 ± 0.8 57.0 ± 0.5
conv-5 65.6 ± 0.6 60.8 ± 0.7 59.5 ± 0.4 62.5 ± 0.6
fc-6 71.7 ± 0.3 71.5 ± 0.4 - -
fc-7 74.1 ± 0.3 73.7 ± 0.4 - -

Table 8. Effect of location and magnitude feature ablations on PASCAL-DET

no ablation (mAP) binarize (mAP) sp-max (mAP)
conv-5 47.6 45.7 25.4
342 P. Agrawal, R. Girshick, and J. Malik

For image classification, damaging spatial information leads to a large differ-

ence in performance between original and spatially-ablated conv-1 features, but
with a gradually decreasing difference for higher layers (Table 7). In fact, the
performance of conv-5 after sp-max is close to the original performance. This
indicates that a lot of information important for classification is encoded in the
activation of the filters and not necessarily in the spatial pattern of their acti-
vations. Note, this observation is not an artifact of small number of classes in
PASCAL-CLS. On ImageNet validation data, conv-5 features and conv-5 after
sp-max result into accuracy of 43.2 and 41.5 respectively. However, for detection
sp-max leads to a large drop in performance. This may not be surprising since
detection requires spatial information for precise localization.

6 Conclusion
To help researchers better understand CNNs, we investigated pre-training and
fine-tuning behavior on three classification and detection datasets. We found that
the large CNN used in this work can be trained from scratch using a surprisingly
modest amount of data. But, importantly, pre-training significantly improves
performance and pre-training for longer is better. We also found that some of
the learnt CNN features are grandmother-cell-like, but for the most part they
form a distributed code. This supports the recent set of empirical results showing
that these features generalize well to other datasets and tasks.

Acknowledgments. This work was supported by ONR MURI N000141010933.

Pulkit Agrawal is partially supported by a Fulbright Science and Technology
fellowship. We thank NVIDIA for GPU donations. We thank Bharath Hariharan,
Saurabh Gupta and João Carreira for helpful suggestions.

References
1. Barlow, H.: Single units and sensations: A neuron doctrine for perceptual psychol-
ogy? Perception (1972)
2. Dalal, N., Triggs, B.: Histograms of oriented gradients for human detection. In:
CVPR, pp. 886–893 (2005)
3. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A Large-
Scale Hierarchical Image Database. In: CVPR 2009 (2009)
4. Donahue, J., Jia, Y., Vinyals, O., Hoffman, J., Zhang, N., Tzeng, E., Darrell, T.:
Decaf: A deep convolutional activation feature for generic visual recognition. arXiv
preprint arXiv:1310.1531 (2013)
5. Everingham, M., Van Gool, L., Williams, C.K.I., Winn, J., Zisserman, A.: The
pascal visual object classes (voc) challenge. IJCV 88(2) (2010)
6. Fukushima, K.: Neocognitron: A self-organizing neural network model for a mech-
anism of pattern recognition unaffected by shift in position. Biological Cybernet-
ics 36(4), 193–202 (1980)
7. Girshick, R., Donahue, J., Darrell, T., Malik, J.: Rich feature hierarchies for accu-
rate object detection and semantic segmentation. In: CVPR (2014)
 Analyzing the Performance of Multilayer Neural Networks 343

8. Gong, Y., Lazebnik, S.: Iterative quantization: A procrustean approach to learning

binary codes. In: 2011 IEEE Conference on Computer Vision and Pattern Recog-
nition (CVPR), pp. 817–824. IEEE (2011)
9. Jia, Y.: Caffe: An open source convolutional architecture for fast feature embedding
(2013), http://caffe.berkeleyvision.org/
10. Juneja, M., Vedaldi, A., Jawahar, C.V., Zisserman, A.: Blocks that shout: Distinc-
tive parts for scene classification. In: Proceedings of the IEEE Conf. on Computer
Vision and Pattern Recognition (CVPR) (2013)
11. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep con-
volutional neural networks. In: NIPS (2012)
12. Lazebnik, S., Schmid, C., Ponce, J.: Beyond bags of features: Spatial pyramid
matching for recognizing natural scene categories. In: 2006 IEEE Computer Society
Conference on Computer Vision and Pattern Recognition, vol. 2, pp. 2169–2178.
IEEE (2006)
13. Le, Q., Ranzato, M., Monga, R., Devin, M., Chen, K., Corrado, G., Dean, J.,
Ng, A.: Building high-level features using large scale unsupervised learning. In:
International Conference in Machine Learning (2012)
14. LeCun, Y., Boser, B., Denker, J.S., Henderson, D., Howard, R.E., Hubbard, W.,
Jackel, L.D.: Backpropagation applied to handwritten zip code recognition. Neural
Computation 1(4) (1989)
15. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. International
Journal of Computer Vision 60, 91–110 (2004)
16. Quiroga, R.Q., Reddy, L., Kreiman, G., Koch, C., Fried, I.: Invariant visual rep-
resentation by single neurons in the human brain. Nature 435(7045), 1102–1107
(2005), http://www.biomedsearch.com/nih/Invariant-visual-
representation-by-single/15973409.html
17. Razavian, A.S., Azizpour, H., Sullivan, J., Carlsson, S.: Cnn features off-the-shelf:
an astounding baseline for recognition. CoRR abs/1403.6382 (2014)
18. Rumelhart, D.E., Hinton, G.E., Williams, R.J.: Learning internal representations
by error propagation. Parallel Distributed Processing 1, 318–362 (1986)
19. Simonyan, K., Vedaldi, A., Zisserman, A.: Learning local feature descriptors using
convex optimisation. IEEE Transactions on Pattern Analysis and Machine Intelli-
gence (2014)
20. Singh, S., Gupta, A., Efros, A.A.: Unsupervised discovery of mid-level discrim-
inative patches. In: Fitzgibbon, A., Lazebnik, S., Perona, P., Sato, Y., Schmid,
C. (eds.) ECCV 2012, Part II. LNCS, vol. 7573, pp. 73–86. Springer, Heidelberg
(2012), http://arxiv.org/abs/1205.3137
21. Szegedy, C., Toshev, A., Erhan, D.: Deep neural networks for object detection. In:
NIPS (2013)
22. Torralba, A., Efros, A.A.: Unbiased look at dataset bias. In: 2011 IEEE Confer-
ence on Computer Vision and Pattern Recognition (CVPR), pp. 1521–1528. IEEE
(2011)
23. Uijlings, J., van de Sande, K., Gevers, T., Smeulders, A.: Selective search for object
recognition. IJCV (2013)
24. Weiss, Y., Torralba, A., Fergus, R.: Spectral hashing. In: Advances in Neural In-
formation Processing Systems, pp. 1753–1760 (2009)
25. Xiao, J., Hays, J., Ehinger, K.A., Oliva, A., Torralba, A.: Sun database: Large-scale
scene recognition from abbey to zoo. In: CVPR, pp. 3485–3492 (2010)
26. Zeiler, M.D., Fergus, R.: Visualizing and understanding convolutional networks.
CoRR abs/1311.2901 (2013)
344 P. Agrawal, R. Girshick, and J. Malik

Appendix: Estimating a Filter’s Discriminative Capacity

To measure the discriminative capacity of a filter, we collect filter responses from

a set of N images. Each image, when passed through the CNN produces a p × p
heat map of scores for each filter in a given layer (e.g., p = 6 for a conv-5 filter
and p = 1 for an fc-6 filter). This heat map is vectorized into a vector of scores of
length p2 . With each element of this vector we associate the image’s class label.
Thus, for every image we have a score vector and a label vector of length p2 each.
Next, the score vectors from all N images are concatenated into an N p2 -length
score vector. The same is done for the label vectors.
Now, for a given score threshold τ , we define the class entropy of a filter to be
the entropy of the normalized histogram of class labels that have an associated
score ≥ τ . A low class entropy means that at scores above τ , the filter is very class
selective. As this threshold changes, the class entropy traces out a curve which
we call the entropy curve. The area under the entropy curve (AuE), summarizes
the class entropy at all thresholds and is used as a measure of discriminative
capacity of the filter. The lower the AuE value, the more class selective the filter
is. The AuE values are used to sort filters in Section 3.2.