Courtesy of Joe Glick, Chief Data Scientist, RYLTI
This year, the MIT EmTech session about knowledge and learning concluded that in the past we measured success as learning to predict the future; now we need to measure success as learning to respond to the present. Very relevant to managing AI risk.
Uncertainty risk refers to inherent limitations of the knowledge we use to address problems and goals, including assumptions, premises, and theories for which we have insufficient evidence or cannot test experimentally. As discussed in Part 1 of this series of posts, with AI we must add the inherent uncertainty of statistical conclusions which are core to AI/ML/NLP algorithms. Einstein said, “So far as the laws of mathematics refer to reality, they are not certain. And so far as they are certain, they do not refer to reality”. AI trained on rapidly proliferating junk websites scales the problem.
The conclusion of Part 3 stated that a biomimetic digital twins ecosystem of independent agents that incorporate expert knowledge graphs can be leveraged to address VUCA risk drivers, including our ability to respond to uncertainty. How?
- Exposing dark data in the problem domain and including it in the analytics
- Stochastic discovery of scenarios
- More objective exploration of response options and trade-offs
- More precise and comprehensive modeling of the problem domain
The value of the first three benefits is straightforward and clear. The fourth item is intuitively correct but may be more challenging to visualize. Consider an example: what makes food taste good?
If you are a chef, the problem domain includes ingredients, preparation, and some knowledge about your guests, including preferences and dietary constraints.
If you are a healthcare researcher working on conditions that prevent people from enjoying food, the domain could include biological, neuronal, psychological systems that produce and govern the senses of taste and smell, the communication of the senses with the brain, the interpretation of those communications by the brain, the genomics responsible for the enjoyment of food, and other domain elements that an expert in the field would define, which I am not.
The point is that when dealing with uncertainty risk, in addition to trying to minimize what we do not know and cannot measure, we need to model our problem domain as precisely and as comprehensively as we can. A biomimetic digital twins ecosystem where each agent is designed independently and incorporates multidisciplinary expert knowledge graphs can significantly improve problem analysis and dynamic adaptability to uncertainty.