BLOG ILLUSTRATION / EXECUTION STORY

Different roles for observation, action and prediction

The action/observation trace and prediction/memory trace show different roles. They do not give a new implementation or a schedule for all implementations.

Different roles for observation, action and predictionPerception gives state to the actor and world model, and the environment returns observations after an action. The world model uses state, action and memory to predict results for cost evaluation. These traces show selected roles, not a necessary global schedule or a cost-to-actor edge. Perception Visual state World model Predict transitions Cost Evaluate state Memory Short-term state Actor Select action Environment Real / simulated Different roles State Prediction Action New percept
Author-redrawn explanation of selected connections in the modular architecture in Xun Huang’s blog and Yann LeCun’s slides. The diagram does not show the configurator or auxiliary connections. This is a role diagram. It is not a new implementation or evidence that modularity makes scaling better. Source: Xun Huang's blog and Yann LeCun's slides. The playback is an example. Animation time is not measured runtime. Open full-size SVG