In this recent brainstorming meeting, we focused on how the cortex could learn causal relationships between actions and observed changes in the world. A significant new idea was discussed: in addition to columns having outputs for voting and hierarchical composition, they also have outputs to learn temporal causality - when we observe a series of changes and a series of actions, we can learn the causal relationships. This mechanism is proposed to occur via Layer 1 synapses on apical dendrites. Then the team discussed how the neocortex could utilize attention, the challenge of temporal delays, and details around learning forward and reverse causal relationships between actions and outcomes.
Summary Video
Main Video
00:00 Discussions Around Learning Causality
04:27 Associating a Behavior With Another Behavior
08:57 Shifting Attention to Learn Causality
21:00 Broadcasting Object IDs and Behavior IDs Across the Cortex
37:16 Associating a Behavioral Model With a Morphology Model
42:56 Trace Synapses and Temporal Delays
54:48 Can We Also Associate Actions Instead of Observed Behaviors?
01:02:22 Reverse Causality
01:07:45 How Does a Column Initially Figure Out How to Achieve a Goal?
01:15:22 Causing Change Requires a Forward Model and an Inverse Model
