Deep Neural Networks for New Product Form Design
Chun-Chun Wei
1
, Chung-Hsing Yeh
2
, Ian Wang
3
, Bernie Walsh
3
and Yang-Cheng Lin
4
1
Department of Digital Media Design, National Taipei University of Business, Taoyuan, 324, Taiwan
2
Faculty of Information Technology, Monash University, Clayton, Victoria 3800, Australia
3
Department of Design, Monash Art Design and Architecture, Monash University, Caulfield East, Victoria 3145, Australia
4
Department of Industrial Design, National Cheng Kung University, Tainan, 701, Taiwan
Keywords: Artificial Intelligence, Consumer-oriented Expert System, Deep Learning, Neural Networks, Product Form
Design.
Abstract: Neural Networks (NNs) are non-linear models and are widely used to model complex relationships, thus
being well suited to formulate the product design process for matching design form elements to consumers’
affective preferences. In this paper, we construct 36 deep NN models, using one to four hidden layers with
three different dropout ratios and three widely used rules for determining the number of neurons in the
hidden layer(s). As a result of extensive experiments, the NN model using one hidden layer with 140 hidden
neurons has the highest predicting accuracy rate (80%) and is used to help product designers determine the
optimal form combination for new fragrance bottle design.
1 INTRODUCTION
Artificial intelligence (AI) is a technique which
enables machines to mimic human behavior, and is
defined as an innovative approach to reasoning and
learning the human mind in an uncertainty and
imprecision environment (Lin and Yeh, 2015). The
aim of AI is to exploit the tolerance for imprecision,
uncertainty, approximate reasoning to achieve
tractability, low solution cost, and close resemblance
with humanlike decision-making (Zheng et al.,
2017). Deep learning is a subfield of machine
learning, which both fall under the broad category of
AI. Moreover, deep learning is usually used behind
the most humanlike AI as it structures algorithms in
layers to create an artificial neural network that can
learn and make intelligent decisions on its own
(Deep Learning Studio, 2019; Azure Machine
Learning Studio, 2019).
Chan et al. (2018) have revealed that consumers
are not only concerned with the functionality and
reliability of products, but are also concerned with
product affections (e.g. texture, shape, color, style,
etc.) that are related to the emotional feelings and
impressions of the products (Chan et al., 2018).
Affect is defined as consumers’ psychological
responses (or emotional feelings) to the perceptual
design details of the products (Lin and Wei, 2017).
In order to be successful in a competitive market,
products need to appeal to consumers on an affective
level, and further capture their affective preferences
(Chan et al., 2018; Lin and Wei, 2017). However,
consumers’ affective preferences are often a black
box and cannot be precisely described (Lin and Wei,
2017; Lai et al., 2005; Lin et al., 2014). How to
accurately capture consumers’ affective preferences
and then transforms them into design elements is
thus a major challenge for product designers (Wang,
2011). What specific techniques can be used to help
product designers achieve this goal and design
products that match consumers’ affective
preferences?
To address this challenging design research issue,
we develop a consumer-oriented expert system
based on deep neural networks (DNNs) for new
product form design (Negnevitsky, 2002). To collect
data required for training and testing DNNs, we
conduct a consumer-oriented experiment on
fragrance bottle form design due to its wide variety
of appearances and appropriate for verifying the
consumer-oriented expert system (Lin and Wei,
2017; Lin and Chen, 2016; Zheng and Lin, 2017;
Lin et al., 2018). The expert system can be used to
help product designers determine the optimal form
Wei, C., Yeh, C., Wang, I., Walsh, B. and Lin, Y.
Deep Neural Networks for New Product Form Design.
DOI: 10.5220/0007933506530657
In Proceedings of the 16th International Conference on Informatics in Control, Automation and Robotics (ICINCO 2019), pages 653-657
ISBN: 978-989-758-380-3
Copyright
c
 2019 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
653

combination of a new product design that best
matches consumers’ affective preferences.
2 A CONSUMER-ORIENTED
EXPERIMENT ON
FRAGRANCE BOTTLE FORM
DESIGN
In our previous studies (Lin and Wei, 2017; Lin and
Chen, 2016; Zheng and Lin, 2016; Lin et al., 2018;
Chen, 2015), we investigated 617 various world-
famous fragrances and their own fragrance bottle
forms with 75 different brands. After performing a
preliminary assessment by a focus expert group, the
multidimensional scaling analysis and the cluster
analysis were used to choose the representative
experimental samples. Finally, 28 fragrance bottle
forms were selected to conduct the morphological
analysis which was used to identify the product form
elements, as shown in Figure 1. As a result of the
morphological analysis, nine product form elements
and 30 associated product form types were identified,
as shown in Table 1. The nine product form
elements are “Shape of bottle top (x
1
)”, “Connection
of bottle top and body (x
2
)”, “Shape of bottle
shoulder (x
3
)”, “Shape of bottle body (x
4
)”, “Shape
of bottle bottom (x
5
)”, “Ratio of bottle width and
length (x
6
)”, “Transparency of bottle top (x
7
)”,
“Transparency of bottle body (x
8
)”, and “Texture of
bottle body (x
9
)”. Please refer to our previous studies
for details (Lin and Wei, 2017; Lin and Chen, 2016;
Zheng and Lin, 2016; Lin et al., 2018).
Figure 1: The product form elements of fragrance bottle.
(“1” indicating “Bottle Top”, “2” meaning “Connection of
Bottle Top and Body”, “3” showing “Bottle Shoulder”,
“4” representing “Bottle Body”, and “5” being “Bottle
Bottom”, respectively).
Table 1: The morphological analysis.
Product Form
Element
Product Form Type
Type 1 Type 2 Type 3 Type 4 Type 5 Type 6
Shape of
bottle top
(x
1
)
Arch T-shape Rectangle
Extraordi
-nary-
shape
Cylinder None
Connection of
bottle top and
body
(x
2
)
With-
connection
None
Shape of
bottle
shoulder
(x
3
)
0-40
0
40
0
-70
0
Above
70
0
Others
Shape of
bottle body
(x
4
)
Symmetri
c-al
curves
Symmet-
rical lines
Parallels
Irregular-
lines
Shape of
bottle bottom
(x
5
)
Obtuse-
angled
(> 90
0
)
Acute-
angled
(< 90
0
)
Right-
angled
(= 90
0
)
Ratio of width
and length
(x
6
)
1:1 1:1-1:3
Above
1:3
Transparency
of bottle top
(x
7
)
Transpar-
ent
Opaque
Transparency
of bottle body
(x
8
)
Transpar-
ent
Opaque
Texture of
bottle body
(x
9
)
Patterned
Geomet-
ric
Streaked None
According to the morphological analysis, the
fragrance bottle sample can be coded using the value
of 1, 2, 3, 4, 5, or 6, if it has a particular design form
type for each of its nine product form elements, as
shown in Table 2. For each selected fragrance bottle
sample, the first column of Table 2 shows the
fragrance bottle sample number and Columns 2-10
show the corresponding type number for each of its
nine product form elements. Table 2 provides a data
set for training and testing DNNs to develop the
consumer-oriented expert system described in the
following sections.
3 DEEP NEURAL NETWORKS
(DNNs)
Machine learning uses algorithms to parse data,
learn from that data, and make informed decisions
based on what it has learned, while deep learning is
a particular kind of machine learning that is inspired
by the functionality of our brain cells called neurons
which led to the concept of neural networks (NNs)
(Deep Learning Studio, 2019; Azure Machine
Learning Studio, 2019). NNs are non-linear models
and are widely used to examine the complex
relationship between input variables (features) and
output variables (labels) (Negnevitsky, 2002).
ICINCO 2019 - 16th International Conference on Informatics in Control, Automation and Robotics
654

Table 2: Data for training and testing deep neural
networks.
No. X
1
X
2
X
3
X
4
X
5
X
6
X
7
X
8
X
9
1
1 2 4 1 1 1 2 1 4
2
4 1 4 1 1 1 2 2 4
3
6 2 4 3 3 1 2 1 4
4
4 1 4 1 1 1 2 1 1
5
4 1 1 3 1 1 1 1 1
6
3 1 4 3 3 2 2 1 4
7
4 2 2 1 2 2 1 1 3
8
4 1 2 1 1 2 2 1 4
9
1 2 3 1 1 2 2 1 4
10
3 2 3 3 3 3 2 1 4
11
4 1 1 2 1 2 2 2 4
12
5 2 4 4 2 3 1 1 4
13
3 1 1 3 3 2 2 1 2
14
2 1 1 3 3 2 2 1 4
15
1 2 3 1 2 2 2 2 3
16
4 2 4 4 1 2 2 2 4
17
5 1 4 1 1 2 1 1 4
18
2 1 1 3 3 2 1 1 4
19
1 1 4 1 2 2 1 1 2
20
4 1 1 3 3 2 1 1 4
21
3 2 3 3 3 2 2 1 1
22
3 2 3 2 3 2 2 1 4
23
5 2 4 2 1 2 2 1 4
24
5 2 1 3 3 2 2 1 4
25
4 1 1 3 3 2 2 1 4
26
5 2 2 3 1 2 2 1 4
27
1 1 4 1 1 2 2 1 1
28
4 2 1 4 2 2 2 1 1
NNs are thus well suited to formulate the product
design process for matching design form elements
(modelled as features) to consumers’ affective
preferences (modelled as labels) (Cross, 2000;
Nelson and Illingworth, 1991). For example, Lai et
al. (2005) used an NN model and a grey prediction
(GP) model in conjunction with a grey relational
analysis (GRA) model to help product designers
determine the best combination of form elements for
achieving a desirable product image (Lai et al.,
2005).
However, most studies in relation to affective
design use NN models, while only few studies pay
attention to DNNs (Chan et al., 2018; Lin et al.,
2007; Lin et al., 2012). In other words, most studies
use a three-layer NN that includes one input layer,
one output layer, and one single hidden layer.
Although this three-layer NN may produce a good
outcome, it cannot effectively model the complex
consumers’ affective preferences as a black box. In
addition, the availability of social big data is
valuable to support product design decision-making
and to fulfill consumer requirements in developing
new products (Chan et al., 2018; Jin et al., 2016). As
such, it is desirable for researchers or product
designers to consider using DNNs with multiple
hidden layers to develop a consumer-oriented expert
system for new product form design.
In this paper, we use the Deep Learning Studio
(DLS) software developed by Deep Cognition Inc.
for its powerful utility and convenient
implementation (Deep Learning Studio, 2019).
There are three main parts to develop a model or
system in the DLS, including (1) Create a DNN
model (e.g. get data, prepare the data, and define
features), (2) Train the model (e.g. choose and apply
a learning algorithm, transfer function, loss function,
and parameters tuning), and (3) Predict (e.g.
evaluate or track performance).
In this paper, we use the multilayered
feedforward DNNs trained with the backpropagation
learning algorithm, as it is an effective and the most
popular supervised learning algorithm (Negnevitsky,
2002; Nelson and Illingworth, 1991). As mentioned
above, the 30 product form types of nine product
form elements are used as the input variables
(features). Therefore, there are 30 neurons in the
input layer. In the output layer, we use the “Sexy”
image word to represent the consumers’ affective
preference as the output variable (label). 250
participants (with ages ranging from 25 to 50) are
recruited to assess the form (look) of the 28
fragrance bottle samples on the “Sexy” image scale
(Lin and Wei, 2017; Lin et al., 2018). A 5-point
Likert scale is used, ranging from 1 (the lowest) to 5
(the highest), so there are five neurons in the output
layer. As a pilot study for using DNNs, 36 (=4*3*3)
DNN models are developed with one to four hidden
layers with three different hidden neurons and
dropout ratios. We use three widely used rules for
determining the number of neurons in the hidden
layer(s), as follows:
HN1: (The number of input neurons + the number of
output neurons) (a)
HN2: (The number of input neurons + the number of
output neurons)*2 (b)
HN3: (The number of input neurons + the number of
output neurons)*4 (c)
In this paper, the number of the total input data is
randomly divided into two sets for the training data
and the testing data with the ratio of 8:2. The
transfer function is Relu for all layers, except the
Softmax function (Deep Learning Studio, 2019;
Azure Machine Learning Studio, 2019) is adopted
Deep Neural Networks for New Product Form Design
655

between the last one hidden layer and the output
layer. The loss function used is Crossentropy, while
the optimizer used is Adadelta (Deep Learning
Studio, 2019). Moreover, the default values of DLS
are used for other parameters such as learning rate
and momentum. Figure 2 shows a DNN model with
two hidden layers as an illustration.
Figure 2: The DNN model with two hidden layers.
4 RESULTS AND DISCUSSION
To evaluate the performance of the 36 DNN models
developed in terms of their prediction ability, Table
3 shows their loss and accuracy values. As shown in
Table 3, two DNN models out of these 36 DNN
models have the highest test accuracy (0.8), i.e. the
Den1_Dp1_HN3 model and the Den3_Dp2_HN1
model. This result indicates that these two models
have an accuracy rate of 80% for predicting the
value of the Sexy image about fragrance bottles.
That is, they are more suitable for modeling the
consumers’ preference about fragrance bottles.
According to the experimental result, these two
models have the same accuracy rate for predicting
the consumers’ preference. For further analysis, the
Den1_Dp1_HN3 model is a three-layer model (one
hidden layer) with the dropout ratio of 0.10 and 140
hidden layer’s neurons, while the Den3_Dp2_HN1
model is a five-layer model (three hidden layers)
with the dropout ratio of 0.25 and 35 hidden layer’s
neurons. It is evident that the Den3_Dp2_HN1
model is deep (in terms of hidden layers), whereas
the Den1_Dp1_HN3 model is wide (in terms of
hidden neurons on the hidden layer). As such, a
“deep” neural network architecture is not always
performing better. In other words, in some design
settings, a “wide” neural network architecture may
have a better performance.
Table 3: The loss and accuracy values of 36 dnn models.
Den1
HN1 HN2 HN3
Drop Train Test Train Test Train Test
Dp1
Loss 0.82 1.03 0.16 1.01 0.17 0.96
Accu. 0.80 0.50 1.00 0.50 1.00
0.80
Dp2
Loss 0.65 1.00 0.41 0.98 0.30 0.93
Accu. 0.90 0.70 1.00 0.70 1.00 0.50
Dp3
Loss 0.88 1.27 0.92 1.13 0.75 0.97
Accu. 0.80 0.50 0.70 0.60 0.90 0.60
Den2
HN1 HN2 HN3
Drop Train Test Train Test Train Test
Dp1
Loss 0.62 1.22 0.10 1.06 0.02 1.25
Accu. 0.80 0.50 1.00 0.60 1.00
0.70
Dp2
Loss 1.08 1.11 0.57 1.04 0.12 0.90
Accu. 0.40
0.70
0.90 0.50 1.00
0.70
Dp3
Loss 1.60 1.35 0.99 1.20 0.52 0.95
Accu. 0.30 0.50 0.70 0.50 0.80 0.60
Den3
HN1 HN2 HN3
Drop Train Test Train Test Train Test
Dp1
Loss 0.44 1.28 0.03 1.21 0.00 1.41
Accu. 0.90 0.60 1.00 0.60 1.00 0.70
Dp2
Loss 1.39 0.83 0.28 1.01 0.08 1.41
Accu. 0.40
0.80
1.00 0.70 1.00 0.60
Dp3
Loss 1.61 1.51 1.38 1.35 0.36 1.01
Accu. 0.20 0.30 0.20 0.40 0.90 0.50
Den4
HN1 HN2 HN3
Drop Train Test Train Test Train Test
Dp1
Loss 0.35 1.39 0.03 1.52 0.01 1.95
Accu. 1.00 0.50 1.00 0.50 1.00 0.60
Dp2
Loss 1.45 1.54 0.27 1.14 0.14 2.27
Accu. 0.50
0.70
1.00
0.70
0.90 0.60
Dp3
Loss 1.47 1.56 1.60 1.37 1.46 1.14
Accu. 0.30 0.40 0.30 0.60 0.40 0.50
Note: “Den1” indicates one hidden layer in the
architecture of the DDN model, and “Den2” indicates two
hidden layers, and so on. In addition, “Drop.” means
“dropout”, and Dp1 has the dropout ratio of 0.10, Dp2 is
0.25, and Dp3 is 0.50. “Accu.” means “accuracy”.
With its highest predicting accuracy rate, the
Den1_Dp1_HN3 model is used to develop the
consumer-oriented expert system, which can model
the consumers’ preference about fragrance bottles.
All possible combinations of design form elements
can be input to the Den1_Dp1_HN3 model for
generating their associated image values. As a result,
a consumer-oriented expert system can be generated,
consisting of 20,736 (=6*2*4*4*3*3*2*2*3)
different combinations of design form elements,
together with their associated “Sexy” image values.
Product designers can specify a desirable image
value for a new fragrance bottle design, and the
consumer-oriented expert system can then work out
the optimal combination of design form elements.
5 CONCLUSIONS
In this paper, we have constructed 36 DNN models
for developing a consumer-oriented expert system
for fragrance bottle design. Extensive experiments
on the performance of the DNN models have shown
ICINCO 2019 - 16th International Conference on Informatics in Control, Automation and Robotics
656

that a “deep” neural network architecture is not
always performing better. In other words, in some
design settings, a “wide” neural network architecture
may have an equivalent or better performance. The
consumer-oriented expert system developed consists
of 20,736 different combinations of design form
elements. With the expert system, product designers
can easily specify a desirable image value into the
system to work out the optimal combination of
design form elements for a new fragrance bottle
design.
ACKNOWLEDGEMENTS
This research was supported by the Ministry of
Science and Technology, Taiwan under Grant
MOST 105-2221-E-141-007, and the Hierarchical
Green-Energy Materials Research Center in National
Cheng Kung University, Taiwan.
REFERENCES
Azure Machine Learning Studio, 2019. https://azure.micro
soft.com/
Chan, K. Y., Kwong, C. K., Wongthongtham, P., Jiang,
H., Fung, C. K., Abu-Salih, B., ... & Jain, P., 2018.
Affective design using machine learning: a survey and
its prospect of conjoining big data. International
Journal of Computer Integrated Manufacturing, pp. 1-
19.
Chen, Y.-T., 2015. The Study of Relationship between the
Perfume’s Form and Scent Image, Master Thesis,
Department of Arts and Design. National Dong Hwa
University, Hualien, Taiwan.
Cross, N., 2000. Engineering Design Methods, Strategies
for Product Design. John Wiley and Sons, Chichester,
UK.
Deep Learning Studio, 2019. https://deepcognition.ai/
Jin, J., Liu, Y., Ji, P., & Liu, H., 2016. Understanding big
consumer opinion data for market-driven product
design. International Journal of Production
Research, 54(10), pp. 3019-3041.
Lai, H.-H., Lin, Y.-C., & Yeh, C.-H., 2005. Form design
of product image using grey relational analysis and
neural network models. Computers & Operations
Research, 32(10), pp. 2689-2711.
Lin, Y.-C., Chen, C.-C., & Yeh, C.-H., 2014. Intelligent
decision support for new product development: A
consumer-oriented approach. Applied Mathematics &
Information Sciences, 8(6), pp. 2761-2768.
Lin, Y.-C., & Chen, Y.-T., 2016. Artificial intelligent
models for new product design: An application study,
IEEE Society, pp. 1134-1139.
Lin, Y.-C., Lai, H.-H., & Yeh, C.-H., 2007. Consumer-
oriented product form design based on fuzzy logic: A
case study of mobile phones. International Journal of
Industrial Ergonomics, 37(6), pp. 531-543.
Lin, Y.-C., & Wei, C.-C., 2017. A hybrid consumer-
oriented model for product affective design: An aspect
of visual ergonomics. Human Factors and Ergonomics
in Manufacturing & Service Industries, 27(1), pp. 17-
29.
Lin, Y.-C., Wei, C.-C., & Chen, Y.-T., 2018. Emotional
design: A multisensory evaluation to visual and
olfactory perceptions of consumers. In 2018 IEEE
International Conference on Applied System Invention
(ICASI), pp. 1292-1295.
Lin, Y.-C., & Yeh C.-H., 2015. Grey relational analysis
based artificial neural networks for product design: A
comparative study, Scitepress, pp. 653-658.
Lin, Y.-C., Yeh, C.-H., Wang, C.-C., & Wei, C.-C., 2012.
Is the linear modeling technique good enough for
optimal form design? A comparison of quantitative
analysis models. The Scientific World Journal, 2012.
Negnevitsky, M., 2002. Artificial Intelligence, Addison-
Wesley, New York.
Nelson, M. M., & Illingworth, W. T., 1991. A practical
guide to neural nets.
Wang, K.-C., 2011. A hybrid Kansei engineering design
expert system based on grey system theory and
support vector regression. Expert Systems with
Applications, 38(7), pp. 8738-8750.
Zheng, F., Wei, C.-C., Lin, Y.-C., Du, J., & Yao, J., 2017.
Intelligent computing for vehicle form design: A case
study of sand making machine, Lecture Notes in
Computer Science, 10689, pp. 154-161.
Zheng, F., & Lin, Y.-C., 2017. A fuzzy TOPSIS expert
system based on neural networks for new product
design. In 2017 International Conference on Applied
System Innovation (ICASI), pp. 598-601.
Deep Neural Networks for New Product Form Design
657