Advances in machine learning has taken over almost every aspect of today's technological solutions at a breathtaking speed. Various deep learning solutions have made a significant real-world impact across industries - robotics, healthcare, virtual assistants, gaming and entertainment and many more. A simple approach powers this complex technology - give the system many examples and the system figures out what is needed to follow these examples.

This has indeed been great for many applications as the system does lot of the heavy lifting necessary to follow the examples. What about applications where it is not easy to come up with these examples - may be because we don't know (yet) how to come up with these examples or the problem domain is so expansive that you cannot cover the gamut of cases with an example. Moreover, what happens when an example that is ideal now no longer remains ideal as the problem domain undergoes changes. Moreover, we typically possess vast pre-existing knowledge in each domain and it may be pretty difficult to enumerate such knowledge as concrete examples that completely represents this pre-existing knowledge. And we need all these examples before we can use the system for our first prediction! So the aforesaid simple approach, although great for many applications, comes up short for many others. We got a great start in deep learning but much more needs to be done to perfect it!

The novel autonomous learning approach demonstrated in this project is an earnest effort in that direction. The simple approach undergoes some much needed makeover - Start with the aforesaid simple approach with what is available and build a rudimentary model and automatically improvise upon this model by selectively overlaying it with better targeted models.

Autonomous learning approach demonstrated by this project was reviewed and is also published in the following peer reviewed publications

1. IEEE SSCI 2021 "Autonomous Learning with Automatically Created Models and a Novel Model Selection," in IEEE Symposium Series on Computational Intelligence (IEEE SSCI 2021), 2021
2. CAEPIA-21 Autonomous learning: Tackling expansive mutating domains,” in Proceedings of 19th Conference of the Spanish Association for Artificial Intelligence, General Session, CAEPIA 2020/21, 2021

As an application of this autonomous learning project, my project DocAide: A Collaborative AI Medical Assistant was also awarded at the 2021 Regeneron International Science and Engineering Fair (ISEF-21).

The autonomous learning approach of this project involves creating a state-of-art machine learning base model trained for the entire problem domain. Since the input feature space of the expansive problem domain is quite large, the base model is not uniformly effective. Portions of input feature space where base model is non-performant are automatically identified and small better performing models are automatically built to overshadow weak portions of the base model. Each small model excels as it targets only a small slice of the huge input feature space. New better performing targeted models are automatically created based on a result-driven relevancy analysis which continuously finds exceptional results, leverages past knowledge, identifies new trends, and detects mutations in problem domain. Furthermore, these additional tiny models may be lazily added as needed, making this technique highly versatile for fast changing domains.

A reinforcement learning based model selector identifies which model is optimal for a query using a novel approach with automatically created hierarchical states. This alleviates need for big complex monolithic models that require extensive training, automatically adapts to changes in the domain without destabilizing the models and offers better control over query performance.

In doing so, the system performs knowledge refinement automatically by determining what can be improved and trains itself without explicit guidance to do so. As a result, smart solutions that go beyond initial training become possible even when domain is expansive and high quality training may not be feasible for entire expanse of the domain, or when spontaneous mutations alter various aspects of the domain.

Results for mutiple sample applications using this approach shows that the system learns quickly, adapts to changes very rapidly and performs quite well against prevalent learning methods.

1. Doc-Aide: DocAide is a general purpose AI-based medical guidance system to guide doctors for surgeries and new diseases like Covid-19. A medical expert initially trains the system to provide this guidance, and then the system uses autonomous learning to refine its knowledge automatically far beyond the human expert. Refer my ISEF-21 project DocAide: A Collaborative AI Medical Assistant .
2. American Football coach-sidekick: The coach-sidekick predicts play-calls to maximize points for an offensive drive. The system uses autonomous learning to refine its knowledge of the expansive domain comprising of game strategies based on capabilities of players and numerous intricate rules. It also automatically adapts to spontaneous mutations due to change in playing conditions or new rules. Refer the published articles that describe use of this approach for this application.

Please visit here for some other cool projects and technologies.