Narrow AI: Boundaries of modern AI – By Aditya Abeysinghe

Narrow AI

Narrow Artificial Intelligence (AI) is a type of AI which can handle only tasks that are assigned. These tasks need to be defined before they can be processed and narrow AI models cannot operate beyond the level of training. This constraint makes this type “weak” or narrow compared to strong types such as Artificial General Intelligence and Artificial Super Intelligence. Narrow AI is the mostly used AI type in commercial systems.

Why it is narrow in AI?

Narrow AI is designed to perform a task based on pre-defined models and inputs. Artificial General Intelligence-based models on the other hand can “think” and perform tasks on their own. However, the latter type requires AI which goes beyond how humans “think”. This type is still being researched and is not used in most commercial AI systems. Models which rely on training and testing is the standard that is used in commercial systems and these models are been used in many types of AI applications without the need for the latter type.

Another reason why most AI models used are narrow is due to the need of large storage and computation to train and use strong models. Strong AI models cannot often be trained on low cost hardware as high compute power and storage is required to train and use them. Therefore, most applications and systems which are based on AI cannot use them or train them due to high costs and unavailability of software and knowledge. Also, most small-scale apps which rely on AI have less storage and processing speed and cannot use strong AI models within those hardware limits.

Advantages of narrow AI

Narrow AI is often considered the limit of intelligence in machines. While Artificial General Intelligence and Artificial Super Intelligence are being used for many research-based AI models, they are seen to be unusable in most applications which are used today. The main drawback of these two types is the inability to predict the outcome of a model. As these types can act on their own processes, it could be difficult to handle intelligent tasks in most systems due to risks of inconsistency in their output.

Another issue of strong AI is the lack of explainable output. Explainable models are required to ensure that AI algorithms are trustworthy to users. Unbiased outputs in these algorithms cannot be ensured when there is lack of knowledge of the output based on an algorithm’s inner processes. Narrow AI models are also not fully explainable when hidden layers such as those in neural networks are used. However, they can often be predicted since the model is trained on datasets with known parameters and the model can only learn based on what it has been initially trained.

Limits of narrow AI

Narrow AI improved many systems and tasks people use. It also enhanced many domains such as self-driving vehicles, object recognition and predictive systems. However, AI is still being used only in models which are trained using existing data and is limited to how an algorithm is trained. AI is not only processing inputs based on what has been trained or learned but also on what a model can perform based on what it has thought. Strong AI is being developed mainly to improve current limits of intelligence in AI. With rapid use of AI, the limits of “narrow” AI is realized in many use cases which would reach plateau of technology if current limits are not surpassed.

Image Courtesy: https://semiconductor.samsung.com/