Skip to content Skip to sidebar Skip to footer

Part 5: Large Language Models have Roadblocks to Discovery

Leveraging Neuroscience for Real-world Reasoning in Biomimetic Digital Twin Ecosystems

From the birth of AI in the 1950s, biomimicry was the intent. Neural networks were designed based on speculations about intelligence in the human brain because there was no scientific evidence at the time. In recent years evidence has emerged thanks to advances in brain scanning and research into language.

What can we learn from neuroscience? Research by Hawkins et al provides fascinating insights. https://www.frontiersin.org/journals/neural-circuits/

The human neocortex learns an incredibly complex and detailed model of the world. Each of us can recognize 1000s of objects. We know how these objects appear through vision, touch, and audition, we know how these objects behave and change when we interact with them, and we know their location in the world. The human neocortex also learns MODELS of ABSTRACT OBJECTS, structures that don’t physically exist or that we cannot directly sense. The circuitry of the neocortex is also complex. Understanding how the complex circuitry of the neocortex learns complex models of the world is one of the primary goals of neuroscience.

The similarity of circuitry observed in all cortical regions is strong evidence that even high-level cognitive tasks are learned and represented in a location-based framework.

The diagram on the left side of the image below (from the paper) shows the difference between the premises that influenced LLM architecture and what the scans revealed:

Premise – HIERARCHY – granular inputs progressively assembled until an object is represented. This is analogous to how LLMs create sentences or images from text.

Reality – ECOSYSTEM – inputs are complete objects and integrated by location and relationships. How can we leverage this learning? By modeling abstract ecosystems (right side of the image). Each ecosystem is a digital twin of the relevant components as well as their relationships and interactions.

-Data uploaded to the pond in its native schema
-No integration, normalization, or cleansing
-Output delivered to data pond: New knowledge and evidence details
-Operations define implemented analysis problem classes
-Context Ontology defines concept domain, concepts, variants, and attributes of each

For example, a genomics research ecosystem would include at minimum twins of:
-Patients
-Phenotypes
-Genes

This becomes a small element of a real-world model that can be expanded and interconnected.

But how does our brain discover and learn its world model? Not by language, and understanding aspects of human language is key to grasping the limitations of LLMs.

About the author: Joe Glick, Co-Founder, Chief Innovation Officer, RYLTI

Leave a comment