@rmounir presents the implementation of burst sampling, a way for Monty to expand its hypothesis space in response to prediction error; this enables Monty to organically update its hypotheses in continually changing environments, including those with multiple or compositional objects.
Summary Video
Main Video
0:00 Introduction
1:33 Monty Relies on Hypotheses for Inference
5:26 Types of Hypotheses
23:11 When to Sample New Hypotheses
34:08 Interactive Visualization of Burst Sampling
43:31 Burst Sampling on Compositional Objects
52:08 Benchmark Improvements
1:01:14 Burst Sampling Parameters
1:03:50 Parameters: Burst Trigger Slope
1:07:06 Parameters: Deletion Trigger Slope
1:10:15 Relationship to HTM Bursts
