Numenta HTM Visualizations, 10 Jun 2015

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Comportex Notebook [DEMO #7] (2015 Spring NuPIC Hackathon)

Description

Marcus Lewis shows off a tool much like the iPython Notebook, but for Comportex.

http://nupic2015spring.challengepost.com/submissions/37804-comportex-notebook

A

Is this our guy? Thank you. Yeah I, yeah I stole Felix's work, I'm combining it and putting it together into a notebook format. The goal is to put the code closer to the visualization of it, so like the execution of it being aware of.

A

What's going on, is right next to the code, and I think that this is a good way of bringing them together, I really like notebooks, this format of it too, like in case anyone doesn't know like it's sort of like um ipython notebook you can evaluate is, but it's in the document form other concepts. Just real quick, comport x is HTM, implemented and closure about live Felix and comport X vis.

A

Similarly, he made a bunch of really cool visualizations, which I steal profusely, so I'm just going to go ahead and jump in so I just want to get something on the screen you see like this is basically come. I am NOT command line output, but you see this model. The model is just like. This is a spec up here: it's creating a model, and rather than evaluating into some big data structure or variable name, you can see it, and so you can. You can like toy with it.

A

You can change the spec a little bit like if I want to have like 800 columns, or you can do multiple, that you choose the cells per column as well right now, it's to lower that and you can do a bunch of regions and series and and so on, so I will, okay and and now in some ways.

A

This is really boring, though, because I was where comport X vis is really cool is when it brings time into the picture, so I create a sequence which is the simulation and go ahead and visualize that so what you see across like left to right, you see time. This is completely felix, like that.

A

This is stole this from comport exposed and I inserted it into a notebook so and this this simulation is actually happening on the server and then it's talking to a browser which is so it's tossing these models across the wire like the the data and it's being visualized here and I'll go ahead and drop.

A

This was just like a toy example which so is this, but I decided to put hot gym data through it and I mean this was all today starting fresh with like a different implementation of Mei, new comp or text already, but starting hot gym today on it, I didn't I, it works, but I tuned. It a lot more if I had a bunch more time, but you can sort of see the investigative process of how you would start tuning it. So here's the model I started with I'll just go on run. A few things.

A

Helper functions will see the beginning of the model. um All of these are red because it's the beginning, they'll start to turn blue and purple as their predicted, and so it's the beginning, there's no predictive model. Yet another thing that that's going to be useful to know later is these inputs. The way encoders are often like this, where you can sort of read them like reading the matrix. This one is the time the hour of the day.

A

This one is whether it's the weekend or not- and this is the consumption oh and I- should show those numbers for people who haven't seen hot gem. I'll show a plot of so the corresponding numbers. To that were the red line. Is that actual data? Blue dots are predictions I can't predict yet because there's no model for it yeah, it hasn't from those distal synapses I'll jump ahead.

A

So the reason this is taking time is because closure is kind of cool right now, there's just a sequence: that's set up like as you read from it, it evaluates the it evaluates the simulation, so you could potentially have like an infinite sequence, which represents a simulation and it that's just a fun way to work with data you would set. So this is 500 time steps later and you see it starting to predict I'll just jump head to a thousand, and similarly it's about to do 500 type stumps.

A

In that time it shouldn't take too long and I mean so it's predicting, though a lot of times these nuts, like I, see this and I'm, not super experience, but I see it and it's like a lot of the predictions are, after the fact, it's kind of just saying what was before.

A

So it's not a it's, not that good, so I kind of like to drill down a little bit and see what the columns are right now and see if I could glean, maybe a little bit of insight, so what it would be a good time yeah. The first ten should be good, there's plenty of. Let's start it like 1004 and do like sure, 10 steps works and you can jump up and down and we're looking at the early ones. One thing that you're seeing here you there are different visualizations.

A

These ones are very quick to generate. You can do a more in-depth 1 I'll. Do that as well. I creatively called them steps and steps plus plus 104.

A

So the reason I brought up the encoders is when you're starting to look at columns and wondering why there why? Which ones are predicted? Are it's not very useful to look at ones that are connected to the upper end coders only and because it's only consumption that we care about? That's the one that we want to predict correctly and so I. Let- and I think I saw that towards the bottom. It's sort of a it's, it's sort of topographical, and so you could.

A

You know to look at those and you start like looking at looking around seeing which ones are bursting, the red ones are bursting, and so you see this one was somewhat predicted. I think, but not totally um so you start getting an idea of. Is this being over predicted? Am I seeing a lot of blue blue means that was predicted, but not active red means active, but not predicted, and it just gives you I mean you, you kind of learn how it all works. A lot better, just through seeing this, but I think it guides.

A

What further questions you ask and I think that's like a really useful thing, and this is about as far as I got on on hot gym, so I'll just go ahead and leave it there, but I think you can see how that, like, I just said how it would guide your next questions. So another feature we will I'm going to and save this save your file in c oj.

A

So, as you see right now, it's running on localhost, which means it's useless to most the world right now and I will drop this just on a file share basically I'm to a static website. Okay, that's going, and so what I'm doing right now is I'm posting this to the web, and it's going to be in the different form where it sits a little bit. uh What less powerful in that you can't execute new code, but has it done yet 30 seconds?

A

Okay, I got to talk a little bit, so it was pastor or later sorry.

A

So basically, what I've just done is I mean I was writing code, but I was also writing a blog post or an essay, or you could slides slides for a presentation. In fact, once it's there and 10 seconds, why is it so? Okay, it's a little bit big yeah that a lot of the work was making it were these weren't enormous. You learn a hundred ways to hang a browser when you're dumping new pic models into it HTM models into it, so it should be there now.

A

It's downloading that oh gosh, hopefully that so and I will discuss a little bit more about um like I, said it's running on the the HTM in this case is running on a server and visualizations and the browser which is new for its. That is basically the model you would use with new pic with comport, X and visualizations of it, because it conveniently can compile down to closure script.

A

All the simulations up to this point ran in the browser and which is like often good, but they're often reasons you would want to do it this way as well, and that was a big part of the work here was making that possible. So it loaded as you can see, it's no longer like this text in foot area, but I'll go way down to this cool one.

A

This one we were just yet, and it's still interactive, so you can post a link to these and like it, you just saw me create it in front of you and I think it's just opens up new scenarios. It's good to get away from localhost so and, like I said that was that was the trivia a little bit different from comfort xmas. So, as mostly all I've got I think seeing is often really good at making you notice questions you should have been asking or that you didn't know to ask and notebooks are really good.

A

Dev environment, like what you just saw, was a dev environment that was like I wrote code in that constantly this this weekend and what these powers combined it one way that this is I would say more convenient than comport X vis is in. We we were both struck by this is like we're confused by something those going on there and that we're like wait. We're in a coating place right now. We can probe in right now and like look at it.

A

It was a mind shift, so so bringing the two together, like closer I, think makes cool stuff at them. So that's what I got.

A

B

So you have a really big fan of notebooks and I just curious that technology was behind. This, like are using other open source technologies.

A

Here's not good call um yeah, it's a comport. Sorry, it's a gorilla repple, which is basically someone saw ipython. Notebook he'll, probably watch this at some point, because I'll use this to make them accept my pull requests but, but anyway and he'll see that this is getting recursive so but yeah, it's called guerilla repple. It's the similar one for closure.