How intelligent is your AI?

Though the concept of Artificial Intelligence came up in the 1950s with scientists, mathematicians, philosophers, and thinkers discussing its possibility it was not until the 1990s and 2000s that the idea began taking shape. We did not have the relevant computing power, the processing speed, the storage capacity, the hardware technology, large scale device connectivity, or the variety of databases.  Today we have the internet with over 4.5 billion users, 4.2 billion unique mobile phone users, 3.7 billion active social media users[1], and we create more data every passing year in various sizes, shapes, and forms. Systems can now decipher human languages (Natural Language Processing), interact back (bots), recognize multi-media files like pictures, audio, video (Computer Vision), collect data from various devices (Internet of Things), create data blocks with distribution ledger capability (blockchain),  store data in flexible infrastructure with huge storage capacity (data lakes, cloud), and many other capabilities that are all testimony to advancing data sciences. The role of AI comes in to decipher value from all the data. To achieve this objective, we began with algorithm led analytics that went from simpler descriptive Management Information Systems to advanced predictive and prescriptive output systems. Fast forward, we have Machine Learning and Deep Learning that can do much more when it comes to utilizing the data sciences skillset and providing a seemingly intelligent output. These two data to decision methods are the closest we have got to be declared AI. But are they intelligent?

To answer the question let us travel back to the foundation of AI as an idea. It was to replicate some level of human-like intelligence in computers. Cambridge Dictionary describes ‘intelligence’ is “the ability to learn, understand, and make judgments or have opinions that are based on reason.” (Point to be noted – the focus here is on reason alone and no other aspects like emotions, society, creativity, or other factors of human decision-making behavior. That is another topic to be dealt with another time, perhaps to understand the apocalyptic topic of whether AI will take over us or will remain benign. We are still contesting if AI is intelligent enough on logical reasoning.)

Dissecting the meaning we have four major capabilities that AI should have. The ability to ‘learn’, ‘understand’, ‘make judgments or opinions’, and make them based on ‘reason.’

Learn. Yes. Machine Learning, Deep Learning is about that. But learning is not the way we humans do. We go from basics to intermediate to advanced learning. Neither do we ingest the dictionary or encyclopedia on day 1 nor do we collect pictures on a specific subject and classify them under one category. Our learning is structured and sequential leading us to ‘understand’ and create a context. We begin from basics. Then, we absorb further learning on top of it or apply new learning incrementally or do both. In that sense, our learning is continuous and expands over time. Machine Learning or Deep Learning are largely library creation activities. There is no scope of context creation. It is just data storage as per a pre-decided model that can be retrieved as and when required. That is one reason why Machine Learning is process is highly iterative, where a new set of information needs to be included (coded) in the library classifications repeatedly. And it not just unintelligent, it is also expensive to do it because you require incremental time of data scientists and developers. Further, making judgment gets distorted if the process of learning and understanding have been distorted as well (as they say garbage-in-garbage-out.) The same goes for a presumption of judgment being based on reason because the reason has also been coded for the system.

In summary, the classic human approach to most forms of intelligence begins with learning that creates a context, which allows us to ‘understand’ the backdrop before we can logically process any new information on the subject. This context is developed with basic learning, which then become autonomous in processing new information accepting, rejecting, and editing existing learning along the way. In contrast, the so-called AI engines have been based on the input-output algorithm by identifying, classifying, and arranging data sets that the algorithms can crawl but give a pre-determined (or some level regression or linear algebra backed probabilistic model) to deliver an ‘intelligent looking’ output. This is not intelligence; it is an advanced form of robotics. While Deep Learning is a better approach to Machine Learning, it too lacks the basic building block – in terms of building the ‘understanding’ elements as it requires ‘end-to-end’ learning or some form of encyclopedia ingestion. Both approaches do help boost productivity and they have their uses. That is good, but not yet intelligent. Experts call these narrow AI or weak AI.

We will attempt to reimagine an intelligent AI or strong AI in our next blog. Keep watching this space.

 

 

 

 

[1] Statista, Numbers as of April 2020