BLOG ILLUSTRATION / EXECUTION STORY

Memory read before update

Vision V encodes the observation as z_t. Controller C reads z_t and recurrent memory h_t to select a_t. The environment uses a_t for execution. Memory M uses a_t, z_t and h_t to calculate h_(t+1). These two operations can occur in parallel. Controller C does not read the predicted next visual code directly.

Memory read before updateVision V encodes the observation as z_t. Controller C reads z_t and recurrent memory h_t to select a_t. The environment uses a_t for execution. Memory M uses a_t, z_t and h_t to calculate h_(t+1). These two operations can occur in parallel. Controller C does not read the predicted next visual code directly. Memory read before update Observation Input o_t Vision V Encoded z_t Controller C Read [z_t, h_t] Environment Action a_t Memory now h_t Memory M Read z_t, a_t, h_t Memory next h_(t+1)
In the classic V–M–C design, the controller reads current visual code z_t and memory h_t to select a_t. The recurrent memory then updates from z_t and a_t. The controller does not directly receive the memory model's next-code prediction. Author-created explanatory diagram. Playback time is illustrative and is not measured runtime. Open full-size SVG