IEEE Consumer Electronics Magazine - May/June 2023 - 33

Swarm Intelligence Optimization This
method only relies on sampling the objective
function, and then realizes the overall random
search through a certain search mechanism,
mainly including simulated annealing (SA), particle
swarm optimization (PSO), and genetic algorithm
(GA). In the edge-cloud collaborative
computing environment, for each preprocessed
DNN subtask, the optimal allocation of each subtask
is searched as the partition-offloading strategy
of the DNN model.1
Distributed Deep Learning Deep learning is
a branch of machine learning that emphasizes
learning from continuous layers. The distributed
deep learning algorithm9 closely combines distributed
learning and deep learning, and uses
multiple parallel DNNs to generate partition-offloading
decisions, providing a new way for DNN
partition-offloading decisions in an edge-cloud
collaborative computing environment.
Deep Reinforcement Learning (DRL) DRL10
reflects human learning by exploring and exploiting
feedback from the environment. In order to
achieve system optimization, previous works
usually convert the DNN offloading process into
a Markov decision process (MDP), and then
select the appropriate DRL algorithm to achieve
system optimization.
Meta Deep Reinforcement Learning Since
the sample complexity of meta-learning is
closely related to DRL, it is considered to combine
meta-learning and DRL in depth to obtain a
meta-deep reinforcement learning (meta-DRL)
algorithm.11 It can quickly adapt to new tasks
and new computing environments, and then
make partition-offloading decisions for the DNN
model faster in an edge-cloud collaborative
computing environment.
Deep Imitation Learning (DIL) DIL12 is a traditional
supervised learning method, which
includes offline training and online decision-making.
After the high-quality demonstration is generated,
the model can be trained offline, and
then the online decision can be made at a fast
online inference speed, which can obtain the
DNN partition-offloading plan very efficiently.
May/June 2023
DNN Partition-Uploading Plan
Direct Uploading Algorithm This approach
typically estimates the expected query execution
time for each possible partition point and, based
on that, finds the optimal partition point for the
DNN model. By partitioning the DNN model once,
we can obtain the DNN partition-uploading plan
(i.e., partition-offloading plan), and then directly
upload the corresponding DNN partition P3 (i.e.,
P1) to the edge/cloud server.13
IONN Algorithm The DNN partition-uploading
plan is obtained by using the shortest path
method and penalty factor method. IONN uses
up to eight penalty factors to find partitions P3:
1; 0:5; 0:25; ... ; 0:016; 0:0. However, partitions P3
are often very large, and there is a lack of opportunities
for fine-grained partitions P3 to provide
better performance.14
Efficiency-Based Algorithm The upload efficiency
and the shortest path method are combined
to find the next partition in each iteration,
where the upload efficiency acts as a penalty factor.
However, different from the fixed penalty
factor, the upload efficiency has higher flexibility
and selectivity. A greedy algorithm is designed
based on the upload efficiency, and the partition
P3 with the highest upload efficiency is repeatedly
selected to obtain a more fine-grained DNN
partition-uploading plan.15
Recursive Efficiency Algorithm This
approach recursively splits partitions obtained by
the efficient-based algorithm and creates a more
fine-grained DNN partition-uploading plan.15 Compared
with the efficient-based algorithm,itis obvious
that this algorithm further improves the query
performance of the DNNmodel.
GENERATION AND OPERATION OF
MIGRATION PLAN
Installation Phase: Generate Migration Plan
In this section, we will describe the installation
of applications in an edge-cloud collaborative
computing environment, briefly introduce
the role of applications, and discuss the flow of
information between applications.
From Figure 4, the applications Program1 and
Program3 installed on the client and the edge/
33
IEEE Consumer Electronics Magazine - May/June 2023

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