Beyond Reinforcement Learning: How OpenCog Hyperon Experiences the World

Join us this Tuesday, May 7th, 2024, at 5pm UTC for the first session of a special two-part SingularityNET's Technical Tuesdays mini-series dedicated to the latest advancements in the development of a unified experiential learning component for OpenCog Hyperon.
Session 1
- Demonstrating how NACE deals with uncertainty and revision of the rules in real-time by applying Non-Axiomatic Logic (NAL) aspects;
- Integrating the AIRIS (Autonomous Intelligent Reinforcement Interpreted Symbolism) causality-based learning AI into Hyperon.
---
NACE: github.com/patham9/NACE
ONA-style NARS implementation in MeTTa: github.com/patham9/metta-nars
AIRIS: github.com/berickcook/AIRIS_P...
AIRIS 2023 Demo - • AIRIS Project 2023 Dem...
AIRIS Experiential Minecraft Demo - • AIRIS Experiential Min...
---
These advancements are part of our ongoing initiative to consolidate the strengths of several systems -ROCCA, NARS, OpenPsi, and AIRIS- to create a unified experiential learning component for Hyperon. This approach will allow AI models to:
- Develop a goal-independent understanding of their environment through causal knowledge gained from planned and spontaneous interactions;
- Explore their environment with increased efficiency using a curiosity model that prioritizes situations with high uncertainty, challenging their existing causal knowledge.
Our preliminary findings indicate that this approach surpasses common Reinforcement Learning techniques in terms of data efficiency by orders of magnitude.
#AGI #OpenCogHyperon #AI

[@williamjmccartan8879]

Thank you everyone for sharing your time and work, Peter, Barick, Patrick, and Hailey, just getting started, Peter speaking about the garbage collector tool being created sounds ingenious yet a very important part of our basic infrastructure, enjoyed Berick's AIRIS presentation, I was wondering if you could have the option to follow along with the chat if you are watching the podcast at a later time, I've seen it used in the past, where you can slide back and forth in the comment section, Peter had a great presentation as well showcasing the MeTTa - NARS, NACE relationship, very cool, discovered YOLO, you only look once, neat.

[@youssefbusalham8449]

👏👍

[@I-Dophler]

How do you envision the role of experiential learning mechanisms, such as those employed by OpenCog Hyperon, in shaping the future of Artificial General Intelligence (AGI)? We're eager to hear your insights and predictions!

[@BerickCook]

Good question! I believe that experiential learning will be one of the cornerstones of AGI. It will give AI agents the flexibility, adaptability, and interpretability that current RL cannot.

[@I-Dophler]

@@BerickCook You make an excellent point about the importance of experiential learning mechanisms for achieving true Artificial General Intelligence (AGI). Current reinforcement learning approaches, while powerful, are fundamentally limited by their narrow reward functions and inability to generalize learned knowledge to vastly different scenarios. Experiential learning that allows AI systems to accumulate knowledge through interaction with rich environments and multi-modal data streams could provide the flexibility, contextual understanding, and robust common sense reasoning lacking in today's AI. Approaches inspired by human learning, incorporating curiosity, play, abstraction, and continual adaptation, may prove essential for developing AI with the general intelligence and interpretability you highlight. However, realizing effective experiential learning at scale presents immense challenges around exploration, credit assignment, knowledge integration, and maintaining performance. OpenCog's Hyperon architecture is an intriguing step, but surmounting obstacles like the exploration-exploitation tradeoff and combinatorial complexity will likely require key conceptual and technical breakthroughs. Ultimately, I agree experiential learning has transformative potential, but significant open research questions remain in achieving the adaptability and interpretability you envision for credible AGI. What are your thoughts on specific paths toward overcoming these hurdles?​​​​​​​​​​​​​​​​

[@BerickCook]

@@I-Dophler Those are good points. There is a lot of open research questions for Experiential Learning systems as they are rather novel. There are a lot of techniques used in traditional GOFAI such as Goal Oriented Action Planning that will translate well into Experiential Learning. These techniques help alleviate some of those challenges. We are just beginning to probe the depths of the capabilities and limitations of these agents, so I am likewise certain that there are more hurdles like these to discover and overcome!

[@I-Dophler]

Absolutely, the integration of traditional GOFAI methodologies with modern Experiential Learning frameworks represents a pivotal shift in our approach to developing more generalized AI systems. The combination of Goal Oriented Action Planning (GOAP) with Experiential Learning not only aligns with cognitive architectures that seek to mimic human-like understanding but also enhances the system’s ability to adapt to novel situations without explicit pre-programming. It’s crucial to highlight that while these integrations offer a pathway towards more versatile and capable AI, they also introduce new complexities in system design and validation. For instance, ensuring that learning outcomes from these experiences lead to generalizable knowledge without accruing biases or undesirable behaviors is a significant challenge. Moreover, as we push the boundaries of what these AI systems can achieve, we must also rigorously evaluate the ethical implications and potential societal impacts of deploying systems with higher levels of autonomy and capability. This journey towards AGI is not just a technical quest but also a deeply philosophical and societal one, requiring us to be as proactive about the frameworks of governance and ethics as we are about the technology itself. We are indeed just beginning to scratch the surface of what is possible with these advanced cognitive models. The road ahead is as thrilling as it is daunting, but it's clear that the convergence of experiential learning with robust AI planning and execution mechanisms will be crucial to our success in creating truly intelligent systems.

[@I-Dophler]

@@BerickCook Indeed, the integration of traditional GOFAI techniques into Experiential Learning seems promising for addressing the unique challenges of this novel field. It's exciting to think about how these established methods can enhance the adaptability and decision-making capabilities of experiential agents. The journey ahead in probing the depths of these systems is undoubtedly filled with both challenges and opportunities for groundbreaking discoveries.