HTM School Episode 8: Spatial Pooling Learning

spatial-pooling

#1

Sorry it took so long!


#2

Worth the wait! Your idea of tracking a column across all of its history for a given input is AWESOME! Added dimensions to my understanding, thanks! :thumbsup:


#3

Hi,

Are the demonstration tools (the UI that is shown in the video) available anywhere to use for performing a demo? Thanks!


#4

Happy to share, but I created this codebase just for myself to do HTM School visualizations, so no promises about readability etc.

All communication between servers is HTTP.


#5

Hi Matt,

I have a question regarding the video:
If only the active columns participate in learning how do inactive columns become active?

Thank you for the brilliant work you do.


#6

This is driven entirely by proximal input. Each minicolumn has a receptive field, or viewport of the input space. If enough bits in that space are on at any given time step, the minicolumn will become active, always in response to changing spatial proximal input over time.

A minicolumn might go without becoming active if by chance there is just not much spatial input within its receptive field, but we try to encourage minicolumns to represent more patterns with boosting:


#7

Hi, I’m wondering some concepts between each input cell in input space and column in spatial pooler

Seeing this time of video,

Q1. Pattern of collection of cells in input space is different as Matt moves mouse from one column to other column.
Let’s say I check first column and second column in spatial pooler.

Then, pattern of collection of cells in input space which you can see as you move column in spatial pooler will be different, right?

What does pattern of collection of cell in input space mean?
Does one input space come from 6th day/weekend in “power value temporal data”?
And does another input space which you can see when you hover mouse onto second column mean, for example, 7th day/week in “power value temporal data”?

Q2. So, each of all columns in spatial pooler means, for example, collection of pointers for 365 days data (if number of columns in spatial pooler is 365) which point each corresponding encoding of each day data in input space?