What is required for ANI or artificial narrow intelligence, the basic building block of AI, evolve into AGI?

The word Artificial Intelligence paints a picture of robots taking over humanity or a Roomba sweeping the floors. It depends on your AI IQ. It's been 15 years since Ben Goertzel coined the word artificial general intelligence but we are not even close to it. Every application from chatbots, and machine translation to expert AI game-players, remains in the domain of narrow AI. Why narrow AI is resistant to evolution? What are the challenges? What best practices can help artificial narrow intelligence evolve into a general-purpose AI?

What is Narrow AI?

ANI or artificial narrow intelligence performs a specific subset of tasks by taking information from a particular data set. It is essentially programmed for performing single tasks such as playing chess or crawling web pages for raw data. They are capable of performing tasks in real-time in spite of not having programs that are part of their programming. ANI is integral to many AI applications such as Google Translate and Siri. Though limited in functionality, they are of immense value because they are goal-oriented and focused but considered weak because they are no match to human intelligence.

Advantages of Narrow AI:

As they have good data processing abilities and agility in competing tasks, they facilitate quicker completion of tasks. For eg., IBM Watson was successful in helping doctors in making quick data-driven decisions by taming the power of AI intelligence. On the one hand while the developments in narrow AI have, for the most part, freed humans from routine tasks, on the other hand, it plays a crucial role in building foundations for Artificial General Intelligence. Affective AI, is one such example, which is designed for sensing nuanced human emotions.

Challenges of ANI:

The black-box approach every AI model is infused with makes it difficult to understand the underlying processes hindering explainability, a crucial parameter for dependency. Particularly, for people who apply AI in high stake businesses making huge investments, it will prove to be a hindrance. As neural networks are exploited intensively there is a grave need for including impenetrable security because even a single violation can result in serious implications. Imagine a self-driving car mistaking a road sign!!  That apart, since learning from small data that they will have to apply to larger problems, the AI stumbles when it comes to making complex connections. They are prone to biases in the data they are fed with.

Best practices in AI

• Human-centric design approach: The end-user experience is as important as successful task completion. To imbibe a human-centric approach designers can train with augmentation and feedback traits. The AI system, instead of delivering a generic reply to every user, it will be able to offer a number of options.
• Examination of raw data: Data forms a critical part of AI or ML models. Examining for errors such as missing values, and incorrect labels ensures the quality of output.

• Having a thorough understanding of limitations and loopholes in AI models: Using a correlation AI model to make inferences will lead to wrong conclusions. By setting the limits for the scope and coverage of the data, the model can make correct predictions.
• Monitoring and testing post-deployment: It will ensure the model will comply to real-world scenarios by including user feedback in the systems.