| Summary: | Background. The research considers the problem of statistical analysis of small
samples by synthesizing new statistical criteria. Materials and methods. It is proposed to
perform the operation of differentiation of random data of a small sample before calculations.
Results. The probability of errors of the first and second kind of classical hi-squared
criterion with a small sample of 16 experiments is 0.33, which is unacceptable for practice.
The new statistical criterion under the same conditions reduces the probability of errors to
0.075, which is already quite acceptable for a number of neural network biometrics applications.
Conclusions. It is generally believed that the operation of differentiating random sample data should lead to a significant loss of stability of calculations. This article shows a
situation that is an exception to the general rule. The synthesized statistical criterion has a
significantly lower probability of errors compared to the classical chi-squared criterion
when solving the problem of neural network separation of normal data and data with a uniform
distribution.
|
|---|