We always think of a computer as having only as much “intelligence” as its human creators bestow upon it. In this context, self-learning computers come as a refreshing surprise, besides suggesting that artificial intelligence (AI) is all set to graduate to a new level.

New sights and sounds, concepts and events unfailingly come your way every day. Invariably, you absorb a lot from this exposure. But have you ever paused in this process of taking it all in to determine just how much of your learning is involuntary and unsupervised?

Taking this a step further, imagine a group of teens trekking through the countryside. They are bound to explore every seemingly interesting nook and cranny, and pick up knowledge through their exploration. Exploration and experimentation are two perfect examples of self-supervised learning. You decide how far you want to go, limited only by your capacity to assimilate —that is, your ability to take in, store and process, or make sense of, information.

AI evolves to new heights

Now imagine a computer doing the same—learning on its own, completely unsupervised by a human, restricted only by its processing and storage capacity. If that sounds bizarre, you have evidently never heard of self-learning software—a kind of AI also known as seed AI.

What differentiates seed AI from traditional AI is its focus on developing an AI software system that learns—or increases its knowledge and skills, and enhances its capacity—by itself, without the explicit teaching and programming that usually goes into developing AI systems. Muralidar Chakravarthi, CTO, Espion International Inc, describes the process of seeding AI as, “Nothing more than bootstrapping AI with limited knowledge to accelerate learning.” Of course, this mandates meticulously designed basic programming that allows the system to take over, so to speak, having understood its own concept and design.

At the core of seed AI is an accurate identification of the fundamentals of intelligence. It strips intelligence down to its basics, leaving aside complexities that are often not required in dedicated, domain-oriented AI systems. This focus comes from more intricate algorithms that may be inter-linked so as to ‘train’ each other, while working in powerful computers.

“Allow me to classify your documents”

For instance, the IBM Classification Module for OmniFind Discovery Edition is sometimes categorised as self-learning software. OmniFind actually classifies large volumes of content in the form of unstructured (or semi-structured) text in documents and e-mails, in real-time, by automatically learning and interpreting its meaning. While the software understands words, linguistics, semantics, and the context of the language as well as associated metadata, it simultaneously improves its own accuracy over time, without requiring any human intervention. This ability renders the OmniFind a useful platform in managing the classification of content archives to render content easy to find, access and use.

As OmniFind is enabled for Services-Oriented Architecture (SOA) environments, it may be usefully applied to deliver an automated content classification service to any content-centric application. For instance, the software can review the content of incoming e-mails of a customer care centre, and determine which ones should be marked as ‘priority’ cases needing the urgent attention of service personnel. Interestingly, OmniFind’s classification function is not limited to its in-built categories. The system functions independent of manual intervention, and can create new taxonomies to classify the content presented to it.

Explaining how IBM’s classification module does not use versatile AI, but rather uses traditional learning techniques that make the learning process as transparent as possible to the administrator deploying it, Dr Daniel Dias, director, IBM India Research Laboratory, describes the OmniFind as a good example of how IBM has incorporated some level of self-learning capabilities into a wide variety of its products.