6 | CONCLUSION AND FUTURE SCOPES
This paper provides a detailed review of arguments and factors affecting
designing the filters for the most promising architectures based on the
convolutional neural network. The paper introduces the convolutional
neural network, its layers, and signal flow, followed by a brief
overview of hyperparameters like filter size, number of filters,
activation function, and more.
The primary purpose of this paper is to shed light on the factors and
supporting arguments made in promising studies for designing filters.
The review starts with the filter initialization and its importance and
specifies different types of the same. Each of them is briefly explained
with strong and weak points. Due to the different nature of learning,
the filter designing study is divided into supervised and unsupervised
learning groups. The filter designing parameters are discussed in detail
for promising supervised methods like AlexNet, ResNet, VGG, and similar
variants with subsequent versions. The relevance of these parameters on
input data, objective functions, application types, computational power,
and other parameters are noted and critically compared. We have surveyed
and reviewed the studies on unsupervised methods like AE, K-means, SOM,
and SSL with the same objective, and the arguments are concluded. Having
lack of mathematical backing, filter designing is mainly summarized as a
data-dependent process. Deep learning is still in its infancy, and
having open questions like optimum filter designing is a tremendous
opportunity for the current algorithms.