Introduction

Artificial intelligence (AI) alignment is an open and important problem within AI for applications, and there is no solution that can fully address it. With the advances of deep learning, a raised question is: When facing the property of the black box for the powerful machine learning tool like deep learning, how can we scientifically propose appropriate approaches for AI alignment? 

A reasonable answer to this question is important under the era of deep learning, since this question intrinsically/philosophically sets our focus on investigating the potential methodology of scientifically proposing appropriate approaches for AI alignment when facing the problem of the black box for deep learning.

Referring to our previous 22 research works, including articles, preprints and reports, we present the concept of discovering scientific paradigms (SPs) for artificial intelligence (AI) alignment to reveal the potential methodology of scientifically proposing appropriate approaches for AI alignment. The illustration for the presented concept of discovering SPs for AI alignment is shown as Fig. 1, and the basic information and the relations of our previous 22 research works are shown as Table 1 and Fig. 2.

Figure 1. Illustration for the concept of discovering scientific paradigms for artificial intelligence alignment. Solid-line parts: the explicit perception-driven fashion for the usual paradigm of developing the artificial intelligence module based on machine learning, data, and computing resource. Dotted-line parts: the implicit cognition-driven fashion for constraining the usual paradigm of developing the artificial intelligence module with scientific paradigms discovered by logical reasoning to produce the more aligned artificial intelligence module. Evolvement of Machine Learning Architecture Based on Data and Computing Resource: usually, a selected ML architecture is evolved based on an amount of data and sufficient computing resources to produce the artificial intelligence module. Artificial intelligence module: a functional component that can map the input data (Input) into the correspondingly desired output data (Desired Output); Input and Desired Output can be multimodality.

Table 1. Detailed information of our previous 22 research works.

Figure 2. Relations of our previous 22 research works. No.? is associated with our previous 22 research works listed in Table 1.

Specifically, by providing the connections for some AI-related terminologies, including machine learning, AI alignment, SP, and logical reasoning, we fuse these terminologies into a unifying concept to present the concept of discovering SPs for AI alignment. By summarizing what the benefits of the concept of discovering SPs for AI alignment are, we reveal that discovering SPs by logical reasoning can just be the potential methodology of scientifically proposing appropriate approaches for AI alignment. 

To make the presented concept of discovering SPs for AI alignment realistically concrete in practice, we illustrate some example SPs and their applications for AI alignment in real-world scenarios. To make the revealed potential methodology, which is discovering SPs by logical reasoning, of scientifically proposing appropriate approaches for AI alignment realistically concrete in practice, we investigate how the former illustrated example SPs for applications in real-world scenarios were discovered by logical reasoning specifically.

Additionally, we present comprehensive discussions to highlight the intrinsic properties and expected potentials of discovering SPs for AI alignment.

For the first time, the concept of discovering SPs for AI alignment, which fuses machine learning (ML), AI alignment, SP, and logical reasoning (LR) into a unifying concept for the AI realm, was formally and systematically demonstrated to reveal the potential methodology of scientifically proposing appropriate approaches for AI alignment. The contributions are:

1. Presenting the concept of discovering SPs for AI alignment by providing the connections for some AI-related terminologies, including ML, AI alignment, SP, and LR.
2. Revealing that discovering SPs by logical reasoning can be the potential methodology of scientifically proposing appropriate approaches for AI alignment by summarizing what the benefits of the concept of discovering SPs for AI alignment are.
3. Illustrating some example SPs and their applications for AI alignment in real-world scenarios to make the presented concept of discovering SPs for AI alignment realistically concrete in practice.
4. Investigating how the illustrated example SPs for applications in real-world scenarios were discovered by LR specifically to make the revealed potential methodology, which is discovering SPs by LR, of scientifically proposing appropriate approaches for AI alignment realistically concrete in practice.
5. Presenting comprehensive discussions to highlight the intrinsic properties and expected potentials of the concept of discovering SPs for AI alignment.
6. Establishing a new practical knowledge system for scientifically addressing the AI alignment problem.
7. Providing a tutorial for systematically constructing the new practical knowledge system for AI alignment.

Establishing a new practical knowledge system for scientifically addressing the AI alignment problem and also providing a tutorial for systematically constructing the new practical knowledge system for AI alignment, we hope potential AI practitioners can benefit from this to scientifically invent better AI technologies for specific application fields and also to systematically propose their own practical knowledge systems for implementing better future AI.

More information can be found at: http://dx.doi.org/10.13140/RG.2.2.15945.52320