Shallow and Deep Learning Principles

Fundamental aspects of modeling, simulation, optimization, and prediction are available in many sources of literature in the form of scientific papers, books and electronic media that provide living and dynamic foundations for their learning leading to real-life applications. The main purpose of this chapter is not to briefly reflect the content of this book and prepare interested reader directly for most advanced learning principles, but to revise basic knowledge and information taken from the formal education system on the basis of scientific, philosophical, and logical foundations. Shallow learning concepts are presented with their expansions to deep learning methodological modeling, simulation, optimization, and prediction problem solutions through artificial intelligence procedures to achieve rational results that encourage new and innovative directions. The shoots of each topic include categories as possible uncertainties, assumptions and logical rule foundations, dynamic learning principles, “shallow” and “deep” learning foundations and sustainability, safe transition from shallow learning to deep learning environment.

The main concern of this chapter is to provide a brief description of artificial intelligence (AI) in terms of historical evolution, types, education, human intelligence, methods, rationality, and related aspects. In the principles that trigger human intelligence, three stages are explained in detail; imagination, visualization, and idea generation. After these three stages of reflection, it is recommended that critical discussion and experimental validation are important cornerstones for scientific inferences. Based on the author’s experience, a number of important recommendations are proposed for future AI work. As will be explained in the next section, before mathematical equations, it is important to understand the solution mechanism of any problem linguistically based on philosophical and logical rule principles. Finally, some examples of misuse in AI studies are given.

In this chapter, it is strongly emphasized that the very fundamental subjects for scientific progress are science philosophy and logical principles, which should be priority courses for mathematical, physical, engineering, social, medical, financial, and all other disciplinary teaching and learning pair education system foundations. Intelligent functionality can be derived from these principles by considering various thinking alternatives as described in this chapter. The distinction between the subjects of general philosophy and philosophy of science is clearly explained with prejudices towards scientific aspects. In particular, logical deduction principles are mentioned in distinction between crisp (bivalent) and fuzzy logic fundamentals. It is stated that linguistic knowledge needs logical principles and fuzzy logic is preferred in ambiguous, vague, incomplete, and uncertainty real-life problem solutions. Common logic conjunctives and their applications in logical inferences are explained in detail. Sets and their overwhelming importance in fuzzy logic modeling are presented with attractive graphics, shapes, and figures. The fuzzy inference system modeling principles are explained by considering different types of the most used fuzzy sets. Some illuminating examples are presented throughout the text.

In this chapter, the types of uncertainty encountered in any scientific study are presented in terms of objective treatment procedures, especially probability and statistical methodologies. Accordingly, the theoretical aspects are not provided except practical insights that will give the reader the skills and abilities to interpret uncertainty modeling procedures, problem solution, and validation on the bases of approximate reasoning and logical rules that lead to numerical problem solutions after linguistic explanations. It has been shown that real-life uncertainties are avoided by a set of homogeneity, isotropy, stationarity, homoscedasticity, uniformity, linearity, and many other assumptions relative to the problem at hand. Basic probability and statistical parameters and methodologies are explained practically. Two main sources of uncertainty have been explained from epistemic, epistemological, and random perspectives. Model validation and efficiency measures are given in terms of a set of error formulations, and it is recommended that the percentage error remain within the limits of ±5% and a maximum of ±10%. The pitfalls in the classical regression analysis application are given with detailed explanations. Trend-setting methodological approaches are needed not only in machine depreciation measurements, but also in quality controls, financial fluctuations, and climate change impacts. In the last part of this chapter, well-known trend analysis methods are explained comparatively.

After linguistical philosophical and logical principles have been widely discussed, the principles of deriving a mathematical statement about a problem at hand are discussed broadly. For this purpose, especially Chap. 2 and the previous chapters are used as a basis for scientific rational propositions in terms of “If… .Then… .” statements. The establishment of all such relevant propositions about a problem provides the basis for translating this linguistical rational knowledge into symbolic expressions, which are mathematical models. The foundations of conceptual, analytical, empirical, and numerical models are equipped with verbal expressions, which are the front soldiers of the conquests of mathematical expressions. The importance of geometric concepts for mathematical formulation derivations including even differential equations is also mentioned. The modeling stages are explained in seven steps, first in the real field, then in the form of critical comments, and finally in the field of model itself. Square graphs are recommended for further model adjustments, if any, to avoid systematic errors with transferences and rotations. A full explanation of the most commonly used mathematical functions is explained in linguistic detail, followed by practical tips for their visual description and use in practical applications.

In general, genetic algorithm (GA) is more suitable for optimization (maximizing and minimizing) problems in numerical and random sequences with much better solutions in a short time than many other mathematical modeling alternatives. The first historical appearance of decimal number system is explained in detail. The application of GA mutation and crossover operations depend on random variability, so different random selections are conveniently explained to GA procedures. The major components of GA procedure are optimization, error minimization, fitness, target function, initial population, mutation, and crossover procedures. Although there are random elements in the method itself, GA can reach the absolute optimization solution in the shortest time. The GA method is easy to understand by everyone as it can reach the result with only arithmetic calculations without requiring detailed and heavy mathematics. However, for this, it is necessary to explain the verbal aspects of the subject within the framework of the rules of philosophy, logic, and rationality. In this chapter, the principles, logic, and similarities of GA philosophy with other classical methods are explained, and the reader ambition for the subject and self-development principles is taken into consideration by giving the necessary clues. Different application examples are given with numerical applications.

This chapter presents a comprehensive explanation of early artificial neural network (ANN) structural and procedural functionality in nonlinear problem solving as preliminary shallow learning aspects across artificial intelligence (AI) methodologies. ANNs are like the human brain in the form of parallel processing modeling inspired by the brain working system. They can process information quickly because they work in parallel and their hardware is easier to implement than other methods. One of the most important features of ANN, which includes many features beyond the models made with regression and stochastic methods, is that it does not require some assumptions about the event or data at the beginning. The connection of the ANN to biological system is explained with characteristic definitions. Various ANN modeling application disciplines and some practical applications are given. The principles of the basic ANN alternative perceptron are explained with numerical applications. The educational contents of ANN are explained to improve the philosophical and logical thinking principles of science. For this purpose, unsupervised and supervised alternatives of ANN procedures are shown for classification and regression purposes.

This chapter presents the principles of machine learning (ML) as the support for shallow and especially deep learning procedures’ software implementations. The main difference between ML and artificial intelligence (AI) is simply given from their deep learning methodologies. Although ML depends on the mathematical procedures mentioned in Chaps. 3 and 4, AI is stated to imitate and mimic human brain functions (Chaps. 6 and 7). Elements and aspects of human thinking for ML are described as shallow and deep learning skills and abilities. Cases of data reliability, model, no data, lack of data, measurement and sampling errors, misclassification, and missing data are discussed, and the relevant uncertainty reduction methodologies are presented. A roadmap is also provided for data reliability adjustment followed by model output prediction validation. Several loss function types are explained by their ease in finding the best (least error) mathematical or AI models. Unsupervised, supervised, and reinforced learning alternatives are explained comparatively. Both bivalent logic k-means and fuzzy logic c-means classification methodologies are explained with their philosophical, logical, and mathematical backgrounds.

It is stated that deep learning (DL) depends on a mixture of artificial intelligence (AI) and machine learning (ML) rules that encompasses all those suggested in the previous chapters. Two main versions of deep learning-enhanced artificial neural network (ANN), convolution neural network (CNN) and recurrent neural network (RNN), are explained in terms of model architecture. The first has feedforward procedures with a series of hidden convolution-pooling layers, followed by the fully connected layer, and then the output layer. CNN is relatively very powerful as a deep learning procedure from traditional shallow learning ANN models. Possibilities of regularization of CNN against over- or lower-fitting cases are mentioned. As for the RNN in the DL domain, they have back propagation layer possibilities to reach the final solution in a short time in the form of compressed or unfolded neural network architectures. In the text, training, testing, and prediction stages of CNN and RNN are explained comparatively.