Courtesy of Joe Glick, Chief Data Scientist, RYLTI
As a reminder, the VUCA framework helps us to prepare for what we cannot predict, prevent, or control even when some elements of those risks can be mitigated. It does not address the standard software-related risks that can be mitigated by good management and engineering practices.
Let’s begin with an overview of the biomimetic principles and methods that apply to all four risk drivers in AI. I will address the methodology required for each one individually.
Principle 1 – Analysis of complex options and tradeoffs using real-world evidence (RWE) ecosystems. AI risks result from interactions that are invisible to us, part of what Gartner labeled dark data, because RWE is excluded/obscured by efforts to integrate databases and scale systems.
The RWE continuum: 1. Data – 2. Information – 3. Context – 4. Interpretation – 5. Learning – 6. Adaptability
AI developers focus on the first two because they are addressed by IT methods, and that’s the easy part. The remaining four require scientific expertise, collaboration, experimentation, objectivity, and adaptability. That’s the hard part, especially the last two. How is this addressed by RWE ecosystems?
Principle 2 – Decision analysis leveraging real-world reasoning (RWR) agents.
The RWR continuum: 1. Problem Definition – 2. Identification of Premises and Theories (including algorithms) – 3. Integration of Domain Expertise – 4. Relevance Computation Rules – 5. Dynamic Adaptability Process
The calculations from DARPA and Dr. Markram are RWE that we cannot achieve RWR by means of exhaustive computation of the possible interactions. A biomimetic approach is required that emulates human thinking.
Brain processes are SYSTEMIC and leverage what neuroscientists label PLASTICITY and SPARSITY.
- Plasticity is the ability to engage diverse combinations of neurons and synapses by relevance to the purpose of the analysis and to dynamically adapt internal functional architectures.
- Sparsity is the ability to identify the minimum data required. The brain can respond to situations that are simultaneously new on multiple dimensions and can even categorize one data point.
- The neuronal and synaptic architecture of the brain is an ecosystem.
Systemic architecture, plasticity, and sparsity are core to biological learning but are NOT similar to ML algorithms. The biomimetic technologies that enable elements of RWR are:
- EXPERTISE GRAPHS and
- NEURAL SYSTEM DYNAMICS DIGITAL TWINS
We can imitate principles of plasticity and sparsity by implementing qualitative expertise graphs and leveraging them for the contextual selection of data and methods from the in-memory model library.
System dynamics has been taught at MIT for some time, and digital twinning is not new. Combining the above methods delivers the evidence to enable VUCA analytics.