A concept and demonstration for a framework that allows artificial intelligence to repair itself using self-supervised learning.
The next generation of robotics will have an emphasis on collaborative problem solving. As humans, we each have our fair share of handicaps that limit us in some form and differentiate us from others, yet we work collectively in a team toward a common goal every day. Even upon injury, humans still carry out day-to-day duties. Artificial intelligence (AI) and robots should be no different. An additional layer of resilience is necessary, in order to expect a robotic agent to collaborate robustly in adverse environments, even upon damage. One way to solve this is by leveraging principles of biological neuroplasticity. Neuroplasticity refers to the brain’s ability to adapt to change by modifying its structure and functionality. Many efforts have begun to define the landscape of neuroplasticity in artificial intelligence [6] [7], and others [22] have shown great success in achieving state-of-the-art performance with a single model across multiple data modalities, but it is to my knowledge that little published work exists on attributing explicit effort toward the real-time repair of damage. In the following framework, two key related concepts are demonstrated: (1) damaged hardware does not entirely compromise software and firmware associated with that hardware—the software can be repurposed for other tasks so long as it is dynamically programmed to do so; and (2) trained machine learning models unaffected by damage can provide supervised learning over models associated with damaged hardware in order to perform this repair. I demonstrate the framework for one full system and propose a design for two more systems in order to replicate results and show scalability, and I use the 6D-framework to do so.