IEEE Consumer Electronics Magazine - May/June 2023 - 32

DNN Migration in IoTs
refer to the abovementioned algorithm as DNN
partition-offloading algorithm.
DNNpartition-uploading plan: Due to the large
size of the DNN model, if we directly upload the
full DNN model to the cloud/edge server, this will
lead to high transmission delay. Therefore, we
consider performing DNN queries while uploading
DNN partitions P3. The DNN uploading plan
refers to determining the sequence of uploading
DNN partition to the edge/cloud server.
Existing DNN partition-uploading methods
are mainly applied to the situation from a single
local device to a single edge/cloud server. Most
of the pioneering methods combine the shortest
path, and after multiple iterations, the partitioning
plan is generated while obtaining the partition-uploading
plan.
It should be noted that in the DNN partition
generated in the DNN partition-offloading plan,
the DNN subtask deployed to the edge/cloud
server is called P1; the DNN subtask deployed on
the client is called P2. The DNN partition generated
in the DNN partition-uploading plan is a further
division of the DNN partition P1 generated
in the DNN partition-offloading plan. In other
words, the DNN partition-uploading plan builds
the upload sequence by dividing the DNN partition
P1 into a finer-grained DNN partition, namely
P3. Obviously, the DNN partition P1 is composed
of the DNN partition P3.
DNN Model Preprocessing:
Inception Module Preprocessing For the
inception modules, it is common to consider splitting
each inception module and executing them
together on both local and edge/cloud servers. To
avoid excessive data transmission on the wireless
network, the inception module is partitioned only
once, which can be considered as a minimum cut
problem, where algorithms, such as Boykov-Kolmogorov
maximum flow algorithm and topological
sorting algorithm, can be applied. Then, the
inception module is cut into two disjoint subgraphs,
with the first one processed locally and
the second one offloaded to the edge/cloud
server.3 In addition, when the difference between
thepredecessor'sout-degreeand thesuccessor's
in-degree under the inception module is 1, two
adjacent layers are merged into a new layer. The
data dependency between the predecessor and
32
successor will disappear after preprocessing.1
Some scholars have also proposed to directly
treat the inception module as a whole.4
Convolutional Layer Preprocessing In
view of the computationally intensive characteristics
of the convolutional layer, the traditional
convolution process is accelerated by
low-rank decomposition schemes, and then the
original kernel set in the convolutional layer is
approximated by two low-rank decomposition
kernels. In another way, the feature map of the
convolutional layer is divided into blocks in
space, and a related feature mapping block is
assigned to an edge server, so that the feature
map of the convolutional layer can be independently
run on the edge server. In addition, the
introduction of a 1 1 convolution kernel can
effectively enhance the nonlinear expression
ability of the network and greatly reduce the
number of model parameters.
Fully Connected Layer Preprocessing For
the storage-intensive fully connected layer, the
number of parameters in the fully connected
layer is actually the size of the weight matrix.
One method is to find a low-rank weight matrix
to approximate the original weight matrix,
thereby reducing the number of parameters of
each fully connected layer, and then reducing
the required storage space.5 Another method is
to use network pruning methods to sparse the
network, reduce the overfitting network, and
then improve the generalization ability of the
network.6
DNN Partition-Offloading Plan
One-Step Segmentation It mainly refers to the
processing of dividing the DNN model once and
then offloading one of the partitions to the edge/
cloud server. Some work considers the offload
performance of partitioning at each candidate
point, and selects the point with the optimal
delay/energy consumption performance for partitioning.7
Furthermore, some scholars have also
proposed a binary search method, and then
proved that the optimal partition-offloading decision
follows a one-climb policy. Based on this, the
Gibbs sampling algorithm was proposed to obtain
the optimal partition-offloading decision.8
IEEE Consumer Electronics Magazine
IEEE Consumer Electronics Magazine - May/June 2023

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - May/June 2023
