Mapping CLS Theory to AI Systems

Introduction to CLS Theory

When I first delved into Complementary Learning Systems (CLS) theory, I was struck by its potential to inform the design of AI architectures. The theory, proposed by McClelland, McNaughton, and O'Reilly in 1995, suggests that the brain relies on two learning systems: fast hippocampal encoding and slow neocortical consolidation. This dichotomy resonated with my work on Samson, a unified personal AI brain that I designed to mimic certain aspects of human cognition.

In Samson, I implemented episodic memory to capture individual conversation turns, which can be seen as a form of fast hippocampal encoding. This allows the system to quickly store and retrieve information from recent interactions. On the other hand, the sleep consolidation process in Samson, which runs nightly, merges patterns and strengthens connections in the entity graph, mirroring the slow neocortical consolidation process. This two-stage approach enables Samson to balance the need for rapid learning with the need for long-term knowledge consolidation.

Mapping CLS to Samson's Architecture

The two-layer curriculum model in Samson, comprising a canonical backbone and dynamic emergent modules, reflects the hippocampal-neocortical division. Both layers undergo the same verification trust gates, ensuring that knowledge is thoroughly vetted before being incorporated into the system. I designed this architecture to align with CLS theory, which predicts that the hippocampus and neocortex work together to facilitate learning and memory formation.

In Samson, dynamic curriculum modules start as hippocampal-like fast captures from conversation, and only get promoted to core knowledge after repeated verification, which can be seen as a form of consolidation. This process allows the system to refine its understanding of the world and develop a more nuanced knowledge graph. The entity graph, which is central to Samson's architecture, utilizes Hebbian co-occurrence learning and spreading activation to strengthen connections between entities, further reinforcing the idea that learning is an iterative process.

Design Decisions and Rationale

When designing Samson's entity graph, I initially used a flat structure where every mentioned entity received a node. However, this approach quickly became unusable due to the sheer volume of single-mention entities competing with core concepts for attention. To address this issue, I implemented a two-tier promotion system, where new entities start as candidates with a salience floor and only get promoted to active after repeated co-occurrence. This design decision was informed by the hippocampal handling of new information, which involves fast initial encoding but consolidation only for patterns that recur.

The Governor component in Samson, which provides centralized adaptive configuration and safety gates for actions, plays a crucial role in ensuring that the system operates within predetermined boundaries. This is particularly important in a system that learns and adapts over time, as it helps prevent unintended consequences. The Lenses, which provide specialized interfaces for various aspects of the system, such as work, curriculum, and personal interactions, allow Samson to focus on specific contexts and develop a deeper understanding of the user's needs.

Open Questions and Future Directions

As I continue to develop and refine Samson, I am acutely aware of the limitations and challenges that lie ahead. One of the key open questions is how to balance the trade-off between exploration and exploitation in the system's learning process. While the dynamic emergent modules allow for rapid adaptation to new information, there is a risk of overfitting or drifting away from the core knowledge base. To address this, I plan to explore the use of more sophisticated verification mechanisms and trust gates to ensure that new knowledge is thoroughly vetted before being incorporated into the system.

Another area of interest is the potential for cortical specialization in Samson's architecture. By allowing the system to develop specialized interfaces and knowledge graphs for specific domains or tasks, I hope to enable more efficient and effective learning and adaptation. This could involve the use of more advanced Hebbian learning mechanisms or the incorporation of additional neuroscience-inspired patterns, such as synaptic plasticity or neural oscillations.

Future Research Directions

In the context of my upcoming PhD research at the University of Washington, I plan to delve deeper into the application of CLS theory to AI architectures, with a focus on developing more sophisticated and adaptive learning systems. My work on Ajax Studio, a multimodal creative AI platform, has already highlighted the importance of persistent memory systems in maintaining consistent creative identity across sessions. I believe that the insights gained from this research will be highly relevant to the development of more advanced AI systems that can learn, consolidate, and adapt in a human-like manner.

Conclusion

In conclusion, my work on Samson has demonstrated the potential for mapping CLS theory to a production AI system. By incorporating elements of hippocampal and neocortical learning, such as episodic memory and sleep consolidation, I have developed a system that can learn and adapt in a more human-like manner. While there are still many open questions and challenges to be addressed, I believe that this approach holds great promise for the development of more advanced AI architectures that can truly learn, reason, and interact with humans in a meaningful way.