Jeff presents on the problem of efficient learning, which centers around how the thousands of columns in our cortex can learn if only a few of them are getting input at any point in time. Previously, we also called this the “shared learning” problem with the example of learning about an object with our right finger tip and then inferring it with our left finger tip or our toe (which are connected to different columns). We talk about related ideas and the role of the hippocampus. @hlee presents some thoughts around learning hierarchical representations using the example of Korean characters.
Summary Video
Main Video
0:00 Introduction
0:26 The Problem Setup
2:42 Related Issues
14:03 Thought Experiment Around Attention and Knowledge Sharing
19:01 Learning Must Occur in Many Columns Simultaneously
36:40 Restating the Problem.
44:32 Visual Depiction of the Problem.
51:26 Hippocampus Processing the Highest Level of Objects
56:03 Class IDs and Object IDs
1:03:39 Class IDs and Object IDs Hangul Example
1:24:49 Recap
1:25:40 What is the Speed of Learning in the Hippocampus
1:31:02 Learning a Mug with No Prior Knowledge Walkthrough
1:44:49 Associative Learning in Columns in Separate Modalities
1:50:43 SDR Sampling vs Dense ID Representations in Monty
