Page 346 - IJB-9-4

International Journal of Bioprinting Bioprinting with machine learning

attributes with enormous initial values from being larger the opposite direction of the data propagation in the neural
than those with smaller initial values. network model. The essence of neural network training is
The classification judgment rules in the algorithm to adjust the parameters of each layer iteratively according
are often majority voting . The most class value of the to the error between the actual output and the expected
[35]
[41]
k nearest training pieces of the input data determines its output . In the training stage, the gradient of the loss
category. In the regression task, the average value of k function is calculated layer by layer through the back
nearest neighbors can be used as the predictive value. propagation algorithm, the output error of each layer in the
network is fed back to the upper layer through the iterative
3.2. Artificial neural network method, and then the weight of the network is updated
Compared with traditional machine learning algorithms layer by layer, so that the theoretical value is closer to the
such as the k-nearest neighbor, a neural network has better sample value. After several iterations, the output error of
generalization ability, and it can automatically extract the the objective function is reduced to the minimum value,
characteristics of sample data for classification, regression, and the trained model is obtained .
[42]
and nonlinear operations. The simplest artificial neural Gradient refers to the expression that expresses the
network consists of three layers, an input layer, a hidden partial derivative of a parameter in a function of multiple
layer, and an output layer. The data comes in from the variables in the form of a vector. In the geometric sense,
input layer, and after being processed by the hidden layer, it represents the place where the function changes and
the output layer outputs the final result . The most critical increases the most. The gradient descent algorithm is a
[36]
section is the hidden layer, where most of the data feature search-based optimization method, which is a common
extraction and generalization work is done. An artificial method to minimize the loss function. Its mathematical
neural network containing multiple hidden layers is often theory is the chain derivative rule of compound function.
called a deep neural network, and the corresponding The gradient descent algorithm mainly obtains the gradient
learning process is called deep learning . of the objective function for all variables in the process of
[37]
The data sequence from the input layer through the training the neural network. In the operation of the neural
hidden layer and finally output in the output layer is network gradient descent, the weight is updated through
considered to be positive. Such a neural network model the negative gradient direction, and the effective optimal
is called a feedforward neural network. In the process way of the objective function can be obtained .
[43]
of feedforward neural network model training, the There are two commonly used gradient descent
parameters of each neuron are constantly modified until algorithms: batch gradient descent algorithm and random
the final model can better fit the data in the training gradient descent algorithm. The difference between them
set. Feedforward neural network usually uses the back mainly lies in the number of samples used to obtain gradient
propagation algorithm based on gradient descent to update and update parameters. The former uses all samples, while
neuron parameters . the latter only selects one sample at random. Random
[38]
3.2.1. Feedforward neural network gradient descent algorithm has advantages in training
The feedforward neural network is a classical neural speed because only one sample is randomly selected to
network model, which is also called multilayer perceptron. update parameters, while the batch gradient descent
Feedforward refers to the one-way propagation of algorithm has more advantages in convergence speed and
[44]
parameters from the input end to the output end. The can converge to the local optimal point faster .
model itself and the output of the model do not form a 3.3. Convolutional neural network
directed ring . There is no feedback connection, and the The simple network structure of the artificial neural network
[39]
information always flows to the output end. When there is model is the reason for the loss of spatial information in
a feedback connection in the network, the model does not vector space, the difficulty of multi-parameter training,
belong to the feedforward neural networks. In the forward and the problems of network overfitting. However, a
propagation stage, the input feature vector is processed, convolutional neural network can better solve the defects
and the output value of each node is calculated. If there is a of artificial neural networks. Its main characteristics are
deviation between the real value of the output layer and the local connection and parameter sharing, and it is easier to
expected value, the error is propagated back . optimize the network by reducing the number of weights,
[40]
3.2.2. Back propagation algorithm thus reducing the risk of model overfitting. Convolutional
A back propagation refers to the order in which neurons neural networks have a significant improvement in large
update parameters first from the output layer, then from image processing performance compared with artificial
[45]
the hidden layer, and finally from the input layer, which is neural networks .

Volume 9 Issue 4 (2023) 338 https://doi.org/10.18063/ijb.739