Numenta HTM School, 4 Aug 2017

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Temporal Memory Part 2 (Episode 12)

Description

This episode contains the answer to last episodes puzzler regarding Single order vs High order memory. We also go over how bursting kicks off the learning of new sequences by choosing winner cells to represent those transitions.

A

Do you remember at the end of last episode, when I challenged you all with a puzzler what you didn't watch the entire episode? Don't you know, I hide all the good stuff at the end of each episode of HTM school.

A

Okay, so the puzzler last episode was to think about what would happen if the temporal memory algorithm was restraint using only one cell per column, I asked how this would affect sequence learning inside temporal memory. To answer this question, we need to learn a bit about single order and high order: sequence memory, dementia research, engineer, Marcus Lewis, so.

B

First order or sequence memory is kind of a special case where it forms the next prediction based solely on what it seeing right now. Was it something right now the current feature causes predictions. Let's.

A

Take a look at an example sequence and how these different memory systems might understand it. In this simple sequence, you can see there are some repeating patterns.

C

A

Sign it's when the saints go marching in oh when the saints go marching in what are you doing.

C

You're too early you're ruining the scene, hey.

A

Man I'm just doing my thing. That's.

C

Fine just wait for your cue.

A

Let's bring out single order memory data, look at this sequence, hi, I'm, single order, memory and I'm going to learn a thing and tell you about it. Oh look! There's the thing: okay, I learned the thing: okay, single order memory. If you hear the note e, what note is most likely to come? Next, that's easy. I'll just go through the whole sequence and look for all the ease and keep count of what follows done.

A

How many times that way, I can tell you that chances are good and F is going to follow an e in a sequence. But what if you knew more than one note in the sequence? Let's say you didn't hear just an e, but you heard a GE right before the e. Can you refine your prediction? Wait? What do you mean before E there's nothing before e F comes after E. This looks like a job for a high order memory.

A

B

Looks like a job for a high.

A

B

High order, sequence memory it uses the currently plus and what is happened previously.

A

That's right: markets from the Mensa I can solve this issue, because I can.

C

Remember more about the notes that led.

A

Up to the current state, so I know that if the G is followed by an e I can more accurately predict what comes next and in this case, GE will definitely be followed by C. What you need led up to the current state, never mind a single order, remembering thank you guys.

A

You told me: I can have a fidgets pair if I did this for you later so having one cell per column restricts each column to represent a spatial feature in only one context, meaning it only has information from the current state to make predictions about the future. This severely limits the functionality is a learning algorithm, a good sequence. Memory needs to recognize spatial patterns with many different contexts, and that's why we need many cells for mini-com.

A

We talked a little about bursting in the last episode, but one key idea is still missing when a column bursts. How is a cell within the bursting column chosen to represent that new sequence in the future? Let's say that proximal input causes some columns to activate here's one of those columns. You can tell that this spatial input was not predicted because there are no predicted cells in this column, so the mini column bursts and every cell is activated. Now let me introduce the idea.

B

Of a winner cell.

A

Every time a mini column becomes active, a winner cell is chosen to represent the latest step in the sequence. If there is a correctly predicted cell in the mini column, that cell.

C

Is automatically the.

A

Winner, because it correctly predicted it would be active. However, if the column is bursting, we still need to choose a winner cell to represent this new pattern, I'm going to explain how the selection is made in a moment, but for now let's just choose one. Moving on to the next time step we get a new spatial input.

B

And another set.

A

Of columns are activated, here's.

B

One of those columns.

A

Once again, this mini column, among others, was activated because it recognized a certain spatial feature in the input space through its feed-forward proximal dendritic segments. Again, this spatial feature was not predicted, so the column bursts. Now, let's talk about how we choose a winter cell to represent this new transition in our sequence.

A

First, we will look through these potential winners in the column to see if any of them almost predicted the last input. This means that they have distal segments that match the previously active cells, but they're permanent values are not high enough to form a connection. If they have been connected, the cell would have been in a predictive state and the column would not have burst. Remember that we look at all the segments of all the bursting cells which can lead to any other cells in the structure.

A

So this graphic is a bit misleading, because it's focusing on just a few columns. There are probably thousands of columns involved, dozens of which are activated by this for input simultaneously. So, given that this was the winter cell last time, if this top cell happened to have a segment to that, so it might become the winter.

C

Even though it's not.

A

Yet fully connected, but what if there were no matching segments at all, let's investigate that with the next time step. The setup for this is exactly the same as our previous steps. Spatial input activates some number of columns. Here's.

C

One of them, which.

A

Is active because it recognized some spatial feature of the input, all the cell's activates, because there were no cells predicting this input. Here's the cell in the last example column, that was the winter cell, but perhaps none of the active cells in the current column have any segments that match any previous winter cells.

A

So we basically have no clue what's going on and we need to represent this as a brand new sequence in this case we're going to inspect all the bursting cells to find a cell with the fewest number of segments and make that bloom winter cell. This ensures that we are utilizing as many cells as possible and not overloading cells with meaning unnecessarily. It's also important that we randomly break any ties that occur when finding this winter cell. Now that the winter self selection is done, we need to either create segments or grow.

C

A

From all the active cells in this bursting column, two cells representing the previous state. We do this by increasing the distal permanence values between these active cells and the winter cells in the last time, step. Remember that we're applying this learning to every winter cell in every active column. In this time step and a couple of notes before we close up this episode, every active column will have a winter cell, even if it's not bursting, in which case the winner is the active cell in the column.

A

Also, not all columns must burst at the same time, it's quite typical for input noise to introduce bursting, because this noise creates many different sequences that are still very similar. So, while completely new spatial input will burst all the columns many times subtle, deviations in spatial features only burst some of the columns. This brings our discussion about HTM sequence memory. To a close, however, there's a lot more to HTM than just sequence. Memory, HTM sensorimotor theory uses essentially the same learning algorithms I've gone over in HTML so far to do.

A

3D object, recognition in the next few episodes I'll be talking about bigger structures like cortical columns, multiple layers of cortex and how interactions between cortical columns and layers allow your brain to do sensory motor integration to recognize objects in 3-dimensional space based upon motor commands and sensory input. If this sounds exciting to you, please, like this video and subscribe to the HTML channel, where we will continue to provide educational materials about hierarchical temporal networks, thanks again for watching ACM.