IEEE Robotics & Automation Magazine - December 2022 - 132

few-shot tactile recognition tasks according to the N-way
K-shot protocol. In the process of metatraining, the training
set contains a labeled few-shot support set
{( ,),...,( ,)}
Dx yx yn
TS =
nn
11
Dx yx yi
TQ = {( ,),...,( ,)} equal to N × M (where M
n
i
qq
11
i
q
i
equal to N × K (where K is usuq
!
#
to
a vibration
ally a small value, such as one or five) and unlabeled query
set
can be any value). Among these, N represents the categories
of samples, and K and M represent the number of
samples for each type. For our problem, each sample in the
support set consists of input data xRDL
magnetic signal of length L for each magnetic feature, and
the label
or contact status. For the training stage, we sample a series
of tasks {, }TD D
yN
t
! 123
=
TS
N
TQ
N
{, ,,...,} means the shape of the object
to train the model. The purpose
of this model is to correctly classify N types of DTQ based
on the DTS.
Architecture of the Metric-Based
Metalearning Model
The flowchart of our model for contact surface recognition
is shown in Figure 6. The overall process of the algorithm
is similar to that of Pronet. First, we collected raw magnetic
contact information of multiple objects in different locations.
Then, we built the training set, testing set, and valid
set based on these data. Through feature extraction, DTS
and DTQ were both converted to feature maps. Finally, we
calculated the prototypical features of each category and
calculated the distance between the feature of the query set
and prototypical features to predict the classification of
new data.
py npn (| )
==
R
In the feature extraction stage, each x in DTQ is converted
P), where θ is the learnable
i
n
i
n
i "
p = 1
n
= i
q
K / fxn
( ,)!
xi yDTS
n
n
i
where K is the set of examples labeled with class k. Then,
dd (( ), )fx pi
n
, which is the distance between xq
n
i()i
(2)
(samples
in DTQ) and the prototype pn, is calculated. For this model,
we choose a squared Euclidean distance to metric the distance.
Considering that the distance represents the similarity,
we define the probability of the sample being in class N based
on the prototype as follows:
m
N
exp(( (),))
q
=1exp(( (),))
df xpn
i
i
q
i
.
df xpm
i
To minimize the distance in the feature space between the
sample and the corresponding prototype, the purpose of the
network is to minimize the cross-entropy loss, which is
defined as
Loss
1 =- gpy npn
=
lo (| ).
(4)
Then, we also expert-maximize the distance between each
different class of samples. When we calculate the prototype
(3)
to a P-dimension vector in the feature space by a convolutional
NN (CNN): ( ():fx x R
parameter of the CNN. Then, we calculate the prototype for
these N classes:
Magnetic Data
Task Building
Metatraining Set
Task 1
Task 2
Known Contact
Information
S1
S2
Q1
Q2
Feature
Extraction
(CNN)
1
2
n
Task N SN QN
Metatesting Set
Task 1
Task 2
S1
S2
Q1
Q2
Match
Recognize
Result
Metric
Embedding
Metric
Embedding
Task N SN QN
New Contact
Information
Figure 6. The flowchart of our method.
132 * IEEE ROBOTICS & AUTOMATION MAGAZINE * DECEMBER 2022
IEEE Robotics & Automation Magazine - December 2022

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - December 2022
