Understanding Narrow AI (ANI)

Hello everyone!

Artificial Narrow Intelligence (ANI) is defined as a version of AI designed to excel at a single task, such as tracking weather updates, generating scientific data reports by analyzing raw data or playing games like poker, chess, etc.

This article explains the fundamentals of Narrow AI, its key advantages and challenges and the best development principles for Narrow Artificial Intelligence.

It’s me again, Neo — Admin — Community Manager of Optimus Finance and Growth Marketing of LECLE Vietnam.

Understanding Narrow AI (ANI)

1. What is Narrow AI (ANI)?

Artificial Narrow Intelligence (ANI) refers to a version of AI designed to excel at a single task, such as tracking weather updates, generating scientific data reports by analyzing raw data or playing games like poker, chess, etc.

Artificial Narrow Intelligence systems are programmed to perform one task at a time by extracting information from a specific dataset. In other words, these systems do not exceed the assigned task.

Unlike General Artificial Intelligence, Artificial Narrow Intelligence lacks self-awareness, consciousness, emotions, and true human-level intelligence. Although such systems may appear complex and intelligent, they operate within predefined parameters, constraints, and contexts.

Machine intelligence that we encounter daily is a part of the same Narrow AI. Examples include Google Assistant, Siri, Google Translate, and other natural language processing tools. Although these tools can interact with us, process and understand human language, they are referred to as Weak AI because they lack the flexibility or autonomy to think freely like humans.

Strong AI and Weak AI

Speaking of Siri, it is not a conscious machine. Instead, it is simply a tool that performs tasks. When we converse with Siri, it processes human language, queries systems like Google Search and provides results.

When someone poses abstract questions like how to solve a personal problem or confronts an experience challenging for tools like Alexa or Google Assistant, they often respond vaguely, lacking logic or providing links to online articles that might address the issue.

On the contrary, when we ask a basic question like ‘what is the outside temperature,’ we often receive an accurate answer from virtual assistants like Siri. This happens because answering such basic questions falls within the scope of the intelligence Siri is designed to perform.

Furthermore, even something as complex as self-driving cars falls under the category of Weak AI, as they are trained to navigate with the assistance of an annotated driving dataset. A typical self-driving car comprises multiple essential Narrow AI systems to ensure smooth movement in highly complex urban environments.

2. Advantages and Challenges of Narrow AI (ANI)

AI and intelligent machines today fall under the category of Weak AI. However, this does not diminish the value of Narrow AI, as it stands as one of the most significant innovations and intellectual achievements of humanity.

2.1. Advantages

2.1.1. Facilitates faster decision making

Narrow AI systems expedite the decision-making process by processing data and completing tasks much faster than humans.

Consequently, they enhance overall productivity and efficiency, thereby improving the quality of life. For instance, Artificial Narrow Intelligence systems like IBM’s Watson assist doctors in making rapid data-driven decisions by leveraging the power of AI. This has made the healthcare field better, faster, and safer than ever before.

IBM Watson

2.1.2. Relieves humans from mundane tasks

The advancements in ANI have ensured that humans are liberated from many tedious, monotonous, and repetitive tasks. It has made our daily lives more convenient, from ordering food online with Siri’s assistance to reducing the effort of analyzing large datasets to achieve results.

Additionally, technologies like self-driving cars have freed us from the stress and burden of being stuck in traffic for extended periods, providing more leisure time to engage in activities or pursue our hobbies

2.1.3. Serves as a building block for the development of more intelligent AI

ANI systems serve as a foundation for the eventual development of more intelligent AI versions such as artificial general intelligence and super artificial intelligence. Voice recognition technology allows computers to convert sound into text with significant accuracy, while computer vision enables the recognition and classification of objects in video streams. Currently, Google is using artificial intelligence to automatically caption millions of videos on YouTube.

Currently, AI combined with computer vision has been used for screen unlocking and online friend tagging. Simultaneously, the automotive industry is exploring the field of ‘affective AI,’ where systems can learn non-verbal expressions (emotions, sentiments) and encourage drowsy truck drivers to stay alert and focused. All these foundational technologies are paving the way for future artificial intelligence versions with consciousness and self-awareness.

2.1.4. Performs single tasks better than humans

ANI systems can perform specific tasks much better than humans. For instance, Narrow AI system programmed to detect cancer from X-ray or ultrasound images can quickly identify a cancerous mass in an image with significant accuracy compared to a professional radiologist.

Another example is predictive maintenance systems used in manufacturing plants. These systems collect and analyze sensor data in real-time to predict whether a machine is likely to malfunction. ANI automates this task. The entire process takes place faster and with virtually no individuals or groups that can match the speed and accuracy.

Overall, the performance, speed and accuracy of ANI far surpass that of humans. However, AI community faces many significant challenges in expanding the scope of ANI.

2.2. Challenges

2.2.1. Absence of explainable AI

Explainable AI is simply a type of AI that must be both capable of effective operation and able to provide transparent explanations for its decisions.

One essential requirement for the advancement of AI is the practical creation of AI that is less of a black box. This implies that we need to position ourselves better to understand what is happening within neural networks.

Current AI systems, such as recommendation systems for books, are effectively utilizing black-box methods. Deep learning algorithms used in such cases examine millions of data points as input and associate specific features to produce results.

This fundamental process is self-guided and challenging for programmers and experts in the field to explain.

2.2.2. Need for impenetrable security

Neural networks are widely used in ANI. However, it is crucial to understand that AI is quite fragile — prone to noise injection and system deception. For instance, an attacker could infiltrate the software system of autonomous vehicles and alter the AI program code so that the program might mistake a bus on the road for an elephant or something else.

This could lead to severe consequences and impacts. A hacker could also take control of the entire network of autonomous vehicles operating in an area and ultimately jeopardize a billion-dollar investment.

2.2.3. Need to learn from small data

To further develop, AI needs to be more adept at learning from less data. AI should have the ability to transfer knowledge from one neural network to others by leveraging previously acquired knowledge.

AI combines both learning and reasoning. Although contemporary AI has made significant strides in learning and accumulating knowledge, applying reasoning to that knowledge remains a challenge. For instance, a customer service chatbot for a retailer may answer questions related to store hours, product prices, and the store’s cancellation policy.

2.2.4. Prone to bias

Current AI systems are prone to bias as they often yield inaccurate results without proper explanations. Complex AI models are continuously trained on large datasets containing biases or inaccurate information. As a result, a model trained on a biased dataset may treat inaccurate information as reliable and make skewed predictions.

Furthermore, because Narrow AI lacks the ability for common sense or an awareness of fairness and justice, addressing bias in the training process requires significant planning and design.

2.2.5. Subject to human failings

Narrow AI heavily relies on humans to perform tasks. Therefore, it is susceptible to human weaknesses, such as setting overly ambitious business goals or prioritizing tasks inaccurately.

Imagine a situation where a human defines a task incorrectly. In this case, regardless of how long the machine operates or how many calculations it performs, the end result will still be a flawed conclusion.

Thus, the dependence of Narrow AI on humans can pose a significant challenge for experts in this field.

3. The best approaches to developing Narrow AI (ANI)

The development of AI is making significant contributions to improving the lives of people worldwide, spanning across business operations, aviation, manufacturing, healthcare, and education. As AI systems become an indispensable part of every industry, discussions on the best ways to integrate fairness, explainability, privacy, and security into these systems are expanding.

3.1. Have a human-centric design approach

The real impact of predictions, recommendations, and decisions made by AI systems can be assessed through the actual user experiences with the system. The following processes can be considered to control the development of your ANI:

• Use reinforcement and support as needed. While addressing certain user scenarios, AI system may consider providing a single answer in situations where a solution could cater to multiple user groups and diverse use cases. It may be appropriate to suggest multiple options for end-users in different scenarios.
• Before full deployment, negative feedback can be integrated into the design process. This can be followed by direct testing and iteration on small-scale traffic.
• Engage in a diverse user group and review various usage scenarios to enable you to incorporate feedback throughout the project development cycle. This practice considers multiple user perspectives while building AI projects, thereby enhancing the number of people who can benefit from the technology.

3.2. Consider metrics to assess training and monitoring

To understand the trade-off between various errors and the user experience of an AI system, one should consider multiple key metrics instead of opting for a single one.

• Metrics may include feedback from user surveys, performance tracking variables for the overall system, control factors for product health in the short and long term such as user click-through rates, and quantities of falsification and errors monitored across various categories of AI products.
• It must be ensured that the selected metrics are based on the context and goals of AI system. For instance, a fire alarm system should have high recall values, regardless of whether the system occasionally returns false alarms.

3.3. Ensure periodic examination of raw data

Analyzing raw data can help you gain a deeper understanding of how machine learning (ML) models operate because they reflect the data on which they are trained. In cases involving sensitive data, you can focus on understanding the input data while still respecting privacy:

• By examining raw data, you can determine whether the data contains missing values or inaccurate labels. You can ascertain whether the information sampled is representative of all users (i.e., users of all ages) in your AI system.
• Identifying performance issues during training and deployment is a challenging task. Therefore, during the training phase, you should look for potential biases and address them immediately, including adjusting training data or restructuring the objective function.
• The data bias should be addressed by carefully analyzing the raw data fed into AI system.

3.4. Consider the limitations of the AI dataset and model

Understanding the limitations of the dataset and artificial intelligence model is crucial to monitoring the gaps in Narrow AI.

• Do not use AI models for correlation detection to make inferences. For instance, your AI model might be aware that most buyers of running shoes are overweight. However, this does not imply that users purchasing a pair of running shoes will become overweight.
• Do not use correlation-detection AI models for making inferences. For instance, your AI model may know that most buyers of running shoes are overweight. However, this does not mean that users purchasing a pair of running shoes will become overweight.
• Machine learning models primarily operate on training data. Therefore, it’s important to clarify the scope and coverage of the training process.
• Limitations should be communicated to end users.

3.5. Ensure the AI system works as intended by performing tests

To ensure that AI system is designed to operate correctly and reliably as expected, you must implement quality testing practices.

• Conducting integration testing allows you to understand how individual ML components interact with other system components.
• Perform repeated user testing to incorporate user needs throughout the AI product development cycle.
• Build quality testing activities into the system to prevent immediate feedback in the event of unintended system errors. For instance, if a critical feature suddenly stops working for AI model’s predictions, AI system may fail to generate an accurate output.

3.6. Regularly monitor and update the AI system post-deployment

Regular monitoring ensures that AI model reviews performance in the real world and integrates feedback from users to update AI system.

• AI products must have a clear roadmap that allows time to address and resolve any issues.
• Addressing issues in both the short term and long term is crucial for AI systems. Short-term fixes may provide quick solutions; however, in the long run, they may not work well. Therefore, balancing short-term and long-term solutions may be better than focusing on just one aspect.
• Understanding that updates can impact the overall quality of the system and user experience is crucial. Therefore, before implementing updates, you must analyze and understand the differences between candidate models and the deployed model.

3.7. Ensure fairness

Today, AI systems are employed in industrial sectors to carry out important tasks such as predicting the severity of medical conditions and cross-referencing records with job or marriage partners.

The risk here is that any unfairness in decision-making systems that are computerized in this way can have wide-ranging impacts. Therefore, as AI permeates society, it is crucial to design a fair and comprehensive model for everyone.

• Analyze how the technology will impact different users and use cases over time.
• Define goals to enable your AI system to function somewhat for diverse use cases. This may include designing certain features provided in different languages for ‘A’ or specifically tailored for different age groups of ‘B’.
• The structure of the objective function and underlying algorithms reflects the fairness goals of AI system.
• Regularly monitor the system to check for biases that machine learning models or algorithms may learn over time.
• Evaluate user experience in real-world use cases, contexts, and scenarios using TensorFlow Model Analysis tools.

3.8. Consider interpretability

Narrow AI systems have improved our lives as automated prediction and decision-making have become commonplace. This can involve various examples, from music recommendations to monitoring patients’ vital signs.

Despite the prevalence of Narrow AI in various fields, interpretability is crucial for understanding and trusting AI systems. The following practical methods for interpretation can be considered before, during, and after designing and training AI models.

• The AI team should collaborate closely with domain experts (e.g., healthcare, marketing, finance) to identify necessary explanatory features.
• Identify post-training interpretability options. Also, determine whether you have access to the internals of ML model (i.e., black-box or white-box).
• Determine whether you can analyze training or test data. For instance, when working with sensitive and private data, you may not have access to the necessary input data for investigation.
• Evaluate whether your AI model provides too much transparency, which may create vectors for abuse from external parties.
• Provide explanations about the interpretability of AI system to the appropriate model users; technical details can be conveyed to industry experts and academics, while ordinary users may be provided with visual representations (charts, graphs, and statistical figures) or concise descriptions.

3.9. Ensure privacy

Machine learning models are programmed to learn from training data and make subsequent predictions on input data. In some cases, both the training and input data can be sensitive. Here, it is crucial to consider the privacy implications when handling sensitive data. This may involve legal and regulatory requirements, social ethics, and user expectations.

• Conduct testing using metrics such as visibility measures or membership inference evaluations to determine whether AI model unintentionally memorizes or discloses sensitive data. Moreover, metrics can also be used for regression testing later in the model maintenance process.
• To understand the balance and identify optimal model settings, experimentation with variables and parameters can be considered to minimize data (such as synthesis, outlier thresholds, and random factors).

3.10. Provide security

The security of an AI system involves determining whether the system operates as intended, regardless of how attackers may attempt to interfere. Addressing the security issues of an AI system before relying on it completely is essential for safety-critical applications.

• Determine all possible attack vectors by constructing a robust threat model. For instance, a threat model should be able to identify a vulnerability in AI system that allows an attacker to manipulate the input to machine learning model, making it susceptible to attacks.
• If the system encounters an error, you need to identify unintended consequences and assess the likelihood and severity of these consequences.
• Develop methods such as spam filtering to counter adversarial machine learning. Additionally, test the system’s performance in an adversarial environment using tools like CleverHans.

4. Closing thoughts

Today, most industries adopt Narrow AI as it achieves exceptional accuracy and performance when completing specific tasks. Factors such as robust IoT connectivity, the prevalence of connected devices, and faster computing have propelled the development of AI systems.

While current AI surpasses humans in specific tasks, the current challenge is how Narrow AI can evolve into a more general AI and broader superintelligent AI. Only time will tell if AI can master cross-domain tasks by building entirely new neural networks from scratch while transitioning from one domain to another.

What about your thoughts? If you want to know further about it, don’t hesitate to share it with us! 😀

This post is for educational purposes only. All materials I used were the different reference sources. Hope you like and follow us and feel free to reach out to us if there is an exchange of information. Cheers! 🍻

#leclevn #leclevietnam #NarrowAI #ANI #ArtificialNarrowIntelligence