Conference Presentation, Keynote, Product Demonstration
Mark Ring
- AI is projected as the final stage of human automation, with Mark Ring arguing that visualizing world mechanics is achievable and that continual learning represents the only viable path to real AI.
- A central goal is the development of a "holy grail" agent that possesses no specific task but continuously improves its knowledge and interaction capabilities over time, mirroring the human 20-year accumulation of skills.
- Current forecasts can span from raw sensory-motor streams to abstract world knowledge, potentially turning raw experience into organized abstractions using isolaminar mechanisms for the first time.
- Demonstrations involve agents with no built-in knowledge, utilizing neural networks to recognize images and predict sensor values such as touch (left/right) and distance, enabling the robot to distinguish rooms and houses based on predictable structural characteristics.
- A chain of forecasts allows agents to track necessary actions to reach high-value states, while agents equipped with richer sensory apparatuses are expected to achieve deeper understanding.
- Computational challenges regarding the timing for building new forecasts, removing old ones, and managing massive forecast volumes are deemed relatively soluble problems expected to be resolved within the next few years given sufficient focus.
- Upon solving these technical hurdles, machines are predicted to understand the world similarly to humans, leading to transformative impacts across nearly all areas of human life including smart robots, cars, question-answering systems, and search.
- Supervised machine learning currently reduces software development burdens but requires manual data compilation, whereas the envisioned continual learning approach aims to eliminate specific task constraints in favor of universal skill acquisition.