When Alan Turing first thought of coming up with machines that could think like humans, he was probably thinking about machines that could one day make the life of human beings easier. Fast forward 70 years, and AI has been able to perform tasks that have undoubtedly made life more comfortable. Conversational AI, flying drones, bots, language translation, facial recognition, etc., are some of the most promising AI applications we have today. But these fall under Narrow AI rather than the Artificial General Intelligence, which is something different.
What Is Narrow AI?
As the definition goes, narrow AI is a specific type of artificial intelligence in which technology outperforms humans in a narrowly defined task. It focuses on a single subset of cognitive abilities and advances in that spectrum.
Over the years, narrow AI has outperformed humans at certain tasks. These include calculations and quantification that have been performed more efficiently with this technology. Today, it has also outperformed human beings in complex games like Go and chess, along with helping make intelligent business decisions, and more.
After narrow AI trumped human performance, the next step came in the form of general AI.
General AI vs Narrow AI
When AI was first explored, researchers had one thing on their minds – to create a system that can learn tasks and solve problems without explicitly being instructed on every single detail. This system should also be able to perform these tasks with reasoning, abstraction, and should also be able to transfer knowledge from one domain to another.
But with time, scientists have struggled to create an AI that can satisfy all these requirements. As the years went on, the original idea of AI, where the system is required to imitate the human brain and its thinking process, found itself in a new category altogether – a different type of AI called General AI or Artificial General Intelligence (AGI).
Researchers still believe the idea of AGI becoming practical is decades away. Much of the truth to that statement comes from the fact that today’s AI systems are not even able to perform tasks that a human child can.
So, on the road to creating an AI system that can imitate humans, scientists and researchers have created many useful technologies. Each time such a technology is created, it is touted as a breakthrough, and the next time something more useful and practical is created, creating a benchmark for coming technologies. Narrow AI is something that encompasses all these useful technologies.
As the definition goes, narrow AI is good at performing a single task – or a limited range of tasks – and at times, can outperform human beings. But the problem with narrow AI is that as soon as it is put under a different setting, or is made to perform a task that is different from what it excels at, it fails. They are not able to transfer their learning from one field to another. For example, if DeepMind’s AlphaStar is made to play a different game, then it might not give the same kind of performance that it did in StarCraft 2 (Grand Master).
Some Barriers Faced
- Narrow AI applications have a lot of hard coded logic or parameters with pre-trained data sets, which are not that effective when it comes to real-time adaptive learning
- The narrow AI systems come with a wide range of architecture, algorithms, and data representations that are incompatible and impossible to combine
- Typical AI designs require certain selected competencies, while missing others that are crucial to AGI
- A lot of environments do not meet the data requirements and accuracy that is needed by narrow AI
- Sometimes, customers are not receptive to narrow AI technologies. This is particularly true for industries like hospitality and healthcare.