2.2. Deep Learning Model
To define deep learning, a neural network must be defined first.
Neural network is a graph of nodes with weights and activation functions. Nodes
are stacked up one by one and consist of a layer. On the network, each layer is
partially or completely associated with the previous layer. These simple
functions can be learned to better recognize increasingly complex input signals
as layers are added one by one. The ultimate goal of this learning is to
properly adjust the connection between each node on the network, its weight,
and the value of each node so that it can be associated with a specific value
when a value is entered. Deep learning forms a variety of architectures and
builds up these layers countless times. Neural networks themselves are
algorithms that were created decades ago. However, attention has been focused
on various machine learning algorithms and has declined. However, the recent
use of large datasets (BigData), the development of powerful hardware (Cluster
and GPUs), and the advent of new learning algorithms have enabled us to learn a
large neural network that goes far beyond the performance of many traditional
machine learning methods.
For typical machine learning techniques, the limitations of
computational performance did not allow good results as data grew. However,
deep learning is advantageous as data and information grows, and rather
requires much larger datasets than conventional machine learning algorithms.
Deep Neural Network (DNN) has now become the standard for computer vision,
voice processing and some natural language processing operations, providing
better learning performance compared to models learned by existing users
manually adjusting parameters, and has been actively introduced in other areas
of machine learning.
The DNN model is very effective when a lot of data needs to be
processed and nonlinear, such as weather forecasting problems. According to the
papers on various DNN algorithm models, MPL models are used the most and are
considered to have high accuracy[13].
- 7 -
The difference between Hadoop and the traditional Relational
Database Management System (RDBMS) is that Hadoop can store structured,
semi-formal, unstructured and all-format data, but RDBMS can only store
structured data. In terms of speed, RDBMS also adopts the Schema On Write
function, which has the disadvantage of being very slow when data is recorded.
On the other hand, Hadoop adopts the Schema On Read function, which has the
advantage of being very fast when data is written. Hadoop also has the
advantage of being able to form a cluster using common hardware rather than
specific equipment, because it uses clusters by connecting equipment in
parallel. These advantages can be highly useful in economic terms. Because the
cost of upgrading a computer's memory or hard disk is exponentially increased
compared to the performance of the component. However, many companies are now
adopting the technology because parallel architectures can translate computer
performance improvements into an arithmetic growth. Hadoop also has an
advantage in terms of network efficiency because it has a philosophy of sending
light code to where data exists and processing it.
| 