Conference Presentation, Keynote, Product Demonstration
Mark Ring
- Mark Ring, co-founder of the startup Kojitai, posits that current AI (e.g., Siri, Alexa) remains a "computationally sophisticated party trick" lacking genuine world understanding.
- Ring argues that nearly any question requiring the visualization of physical world mechanics exceeds the capabilities of existing AI technologies.
- He frames AI as the final step in human automation, following the industrial revolution (physical labor), the information revolution (mental labor), and supervised/reinforcement learning (tedious development/trial-and-error).
- Ring identifies "continual learning" as the ultimate goal for AI, defined as machines that independently add to their knowledge and skills without specific task constraints.
- The proposed solution is an "iso-laminar learning mechanism," a single, reusable mechanism for learning at every level of complexity, analogous to the human brain.
- Forecast, a key mechanism developed by Ring and collaborator Rich Sutton, encodes knowledge as behavior-dependent, subjective predictions (e.g., "I will see blue if I look up").
- Forecasts function hierarchically, allowing one forecast to predict the output of another, enabling the construction of complex abstract knowledge from simple sensory inputs.
- A thought experiment demonstrated a robot agent with only a camera, a touch sensor, and five basic motor actions acquiring world understanding through continual learning.
- The agent generated a "touch forecast" predicting sensor activation upon extending its finger when facing an obstacle.
- Subsequent "touch left" and "touch right" forecasts were built to predict the values of the touch forecast based on rotation, creating an awareness of immediate surroundings.
- A "touch adjacent" forecast allowed the agent to identify positions from which a touch event could be triggered by rotation.
- By chaining these predictions, the agent developed a "distance forecast," estimating steps required to reach an obstacle based on visual cues and action outcomes.
- The system eventually constructed high-level abstractions, such as the concept of a "wall," by predicting whether the agent could maintain an obstacle on its side while rolling forward.
- This hierarchical forecasting enabled the robot to distinguish rooms, map doorways, and understand relationships between houses and rooms based on geometric configurations.
- The demonstration confirms that an iso-laminar mechanism can transform raw sensory-motor streams into verifiable, grounded abstract knowledge without programmer-assigned labels.
- Significant engineering challenges remain before forecasts are viable in real-world applications, including determining when to create or discard forecasts and managing computational load at scale.
- Ring estimates these technical hurdles are solvable within the next few years given sufficient focus and effort.
- Ring predicts that successful continual learning agents will impact nearly every aspect of human life, including smart robotics, autonomous vehicles, search engines, and Q&A systems.