Semantic IoT: The Key to Realizing IoT Value
95
Table 2
(continued)
ML algorithms
Attributes
Pros.
Cons.
Deep neural network (DNN)
• Best-in-class performance in
multiple domains
•
Suitable for data-mining and large
feature sets
•
Reduce the need for feature
engineering, selection, and
optimization which consume much
time
•
Can adapt to new problems easily
• Maximum utilization of unstructured
data
• Elimination of the need for feature
engineering
• Ability to deliver high-quality results
• Elimination of unnecessary costs
• Elimination of the need for data
labeling
•
Not suitable for small feature sets
•
Needs thousands of samples to
perform satisfactorily
•
Computationally expensive to train
(requires GPU in most cases due to its
requirement of large data handling)
•
No strong theory or properly defined
mathematical model to guide for
determination of DNN topology,
training method, favor, and
hyper-parameters. Learning is spread
over the hidden layers and hence is a
black-box approach
•
It is not possible to comprehend what
is being learned