An Empirical Model For Validity And Verification Of Ai Behavior: Overcoming Ai Hazards In Neural Networks

Rapid progress in machine learning and artificial intelligence (AI) has brought increasing attention to the potential impacts of AI technologies on society. This paper discusses hazards in machine learning systems, defined as unintended and harmful behavior that may emerge from poor design of real-world AI systems with a particular focus on ANN. The paper provides a review of previous work in these areas as well as suggesting research directions with a focus on relevance to cutting-edge AI systems with a focus on neural networks. Finally, the paper considers the high-level question of how to think most productively about the safety of forward-looking applications of AI.

zigzag-shaped function [2]. The question now is how we arrange neurons in ANN to make it easier for a learning algorithm to find those biases and weights.
For clarity and simplicity, the paper divides the most common ANN architectures based on three criteria: (1) number of layers, (2) flow of information and (3) neuron connectivity.

Learning Algorithm
Designing network architectures is a difficult task but training and teaching these networks are surely more difficult. To understand how ANN has been trained, it is better to start with a very simple one neuron example [26]. The principles which are used to teach a single neuron are also used to teach a whole network. However, a network level adds extra complexity which requires an additional step. Suppose you have a very simple neuron with one input and one output. You want to teach this neuron to do a certain task (for example to memorize a multiplication table for number 5). To teach this neuron, ANN researchers usually give it a so-called training set. A training set contains a number of different input values (1, 2, 3, 4, 5, 6 ...) paired with the correct output (5, 10, 15, 20, 25, 30 ...).
One note in ANN model of learning is how AI researchers are setting the value of learning rate. Actually, learning rate is one of many other parameters which are left free for human and outside of ANN's control. For example, (1) the number of layers, (2) the number of neurons in each layer, (3) the size of training set, (4) the activation function type, and (5) regularization parameter as well as (6) the learning rate are some of those free parameters which are called hyperparameters [24,18] Choosing the right values of hyper-parameters is left for a person who manages the ANN.
Bandura [18] criticizes those views of human learning which concentrate merely on neural patterns to interpret learning and argues that such views strip humans of agentic capabilities and a self-identity. In contrary, Bandura [18] conceives consciousness as an emergent property of brain activities which is not reducible solely to the property of neurons activity. In other words, the consciousness is higher-level force which is a result of lower-level neural activities but its properties are not limited to them. As clarified in this study, ANN design shows the need for consciousness force to manage and regulate ANN learning but this force does not occur as an emergent property of neural activity as Bandura proposes. Rather, it is a completely distinct entity which uses, guides and manages the neural activity and does not result from it. Therefore, overcoming hazards in the field AI becomes crucial to maximize societal benefit of AI given its significant expansion.

RESEARCH METHOD
Research has been developed and constructed based on a review of various books focusing on Russell and Norvig How can one enable meaningful human control over an AI system after it begins to operate? ("Ok, I built the system wrong, can I fix it?")

RESULTS
User-Centered Design (UCD) is a systematic approach that is used to overcome AI hazards during software development. UCD is typically divided into 5 main phases: User centered design focuses on software development using a top-down holistic approach. The main goal of UCD is to start by building a user navigation model first using multiple UCD tools and techniques in alignment with ANN development goal. The second major step will be to use this navigation model to define and design the ANN application structure using multiple prototyping techniques ( Figure 1).

Figure 1: Traditional SE (software engineering) vs UCD
As it can be seen in Figure 1, given the main departure from the traditional software development starting with application structure, the main focus of UCD is on the following different points: 1) Who are the users for ANN?
2) What do the users currently do?
3) What do the users want from the new software based on ANN?
The first step of identifying the users starts by meeting with different stakeholders affiliated with the training domain. Two main aspects need to be considered: i. Different types of users are researched and identified as "User Roles".
ii. Users can have different backgrounds, different experience, and different context to teach in. These differences are looked at using "User Personas" as a common UCD tool that is growing in popularity.
The second part, "what the users currently do" is done via different UCD tools. Several users can be met during brainstorming sessions or design workshops to get their input on their daily work practices. A series of interviews and questionnaires could also supplement the work resulting in potential categories for data analysis.
The third part, "what the users want" can be conducted with deeper analysis on the findings extracted in the previous part. This is where the navigation modeling is used as an innovative technique to envision the user needs in terms of a navigation structure that can be translated into an actual application structure.
After successful completion of the User Research phase, the high-level and detailed design phases can start. The following sections provide an overview of the various phases of UCD.

Phase 1 -User Research
During the User Research phase, focus groups with engineering and computing systems staff can be organized to understand the course design process used by the participants. Participants can also be asked to fill an electronic questionnaire about software design tools that they currently use to create and manage their software Data collected from the focus groups about software design process can be categorized as inputs, processing and decision-making, and output artifacts.

Phase 2 -High-level Design
Once the user research provided a relatively clear idea and understanding of domain-and user needs, this initial design phase provides a high-level design with concepts identification, conceptual modeling and early prototyping for ANN. The main goal of high-level design is to plot down schematic ideas and steps into visual graphs and models; an early blueprint of ANN. This can be done by investigating different options and provide design alternatives to ensure a broad view before identifying a good design. Doing this early on, at high-level, sketchy, paper-based only, and without going into details could help provide several solution alternatives at a very low cost. The high-level design sketches can be discussed with the users to make sure what they said in unstructured dialogs and vague ideas and imaginations can now be concretely captured in design artifacts for further validation and clarifications. At this stage, there are 2 tools that are most suitable for the development stage:

a)
Navigation Model is one of the essential methods of design. A significant challenge in complex software is not the contents of each screen, but how the user mentally builds a mental view of how all screens are connected (like a city road map), and how to navigate between hundreds of screens to accomplish their task. In this regard, an effective technique-elastic prototyping-can be used an implementation of a participatory design to help designers and users build a navigation model together, greatly reducing time and effort needed. b) Prototyping (PT) is extensively used in UCD to visualize and validate all otherwise vague ideas and unclear expectations at low cost and high effectiveness. There exist three main categories of prototyping: Paper (low-level) PT, low-fidelity electronic (medium level) PT, and high-fidelity, detailed PT. Paper prototypes are very inexpensive and help capture several initial ideas and concepts, and validate them. After explaining their needs, users often change their minds when they see them on paper. Therefore, multiple paper PT sessions gives a head start in validating what users actually mean and need. After initial concepts, once design ideas and directions were identified, a medium fidelity prototyping stage can start where a sketchy visualization of key screens without contents are provided to be gradually validated them and added with initial contents.

Phase 3 -Detailed Design
At this stage, the focus is on the main high level solution, including details from different perspectives such as main application features, auxiliary features, concrete navigation models, menu options, visual and interaction consistency across all screens, exceptions and error massages and recovery, reliability assurances and, help. This phase can proceed in parallel with development phase as more details are uncovered and technical problems arise. User interface mockups can be created with details of various user inputs that will be solicited through the course design process.

Phase 4 -Development and Development Support
As implementation of essential features starts, close collaboration between designers and software engineers (software architects and developers) is essential to ensure the consistency of design and to prevent any deviations.
Several technical problems require careful reconsideration of detailed design and even high-level design options. Iteration is a fundamental design approach that is extensively being used across the UCD process. Therefore, UCD is highly iterative and most of its phases are heavily overlapping to ensure design and development decisions are aligned at all times with the actual user needs.
This phase of the project includes identifying appropriate technologies to be used for the development of the ANN application, design of the back-end database schema, installation and configuration of the server-side and client-side technologies, and development of the user interface screens for login, registration, index, and creation of an instructional module and the connectivity of these web pages with the backend database.

Analysis of Technologies
The purpose of analyzing various technologies during this phase of the project is to ensure rapid development with the latest technologies in the field of software development and use open source technologies wherever feasible. Towards this end, an analysis of web application frameworks, version control systems, server side technologies and client side technologies can be performed.

System Architecture
A Model-View-Controller (MVC) architecture is suggested as the underlying web application framework. MVC is a software architecture pattern which separates the representation of information from the user's interaction with it. The recommended architecture can be described as follows: • The foundation is the Java Virtual Machine (JVM).
• There is a separation between the Java language and the JVM.
• The final layer of the architecture is the application layer. This layer follows the Model-• View-Controller (MVC) pattern.
• A controller handles requests and creates or prepares the response. A controller can generate the response directly or delegate to a view.
• A controller can have multiple public action methods, each of which maps to a URI.

DISCUSSION
There are many ways of responding to information hazards. In many cases, the best response is no response, i.e., to proceed as though no such hazard existed. The benefits of information may so far outweigh its costs that even when information hazards are fully accounted for, we still under-invest in the gathering and dissemination of information. Moreover, ignorance carries its own dangers which are oftentimes greater than those of knowledge. Information risks might simply be tolerated.
When mitigation is called for, it need not take the form of an active attempt to suppress information through measures such as bans, censorship, disinformation campaigns, encryption, or secrecy. One response option is simply to invest less in discovering and disseminating certain kinds of information. Somebody who is worried about the spoiler hazard of learning about the ending of a movie can simply refrain from reading reviews and plot summaries.
At the same time, however, we should recognize that knowledge and information frequently have downsides. Future scientific and technological advances, in particular, may create information which, misused, would cause tremendous harm-including, potentially, existential catastrophe.
It can also be hoped that new information technologies will bring about a vastly more transparent society, in which everybody (the watchmen included) are under constant surveillance; and that this universal transparency will prevent the worst potential misuses of the new technological powers that humanity will develop.

CONCLUSION
Even if our best policy is to form an unyielding commitment to unlimited freedom of thought, virtually limitless freedom of speech, an extremely wide freedom of inquiry, we should realize not only that this policy has costs but that perhaps the strongest reason for adopting such an uncompromising stance would itself be based on an information hazard; namely, norm hazard: the risk that precious yet fragile norms of truth-seeking and truthful reporting would be jeopardized if we permitted convenient exceptions in our own adherence to them or if their violation were in general too readily excused.
It is said that a little knowledge is a dangerous thing. It is an open question whether more knowledge is safer. Even if our best bet is that more knowledge is on average good, we should recognize that there are numerous cases in which more knowledge makes things worse.