Numenta 2015 HTM Challenge, 4 Dec 2015

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Cloud Sonar - Takenori Sato

Description

2015 HTM Challenge Application submission

A

Next up is cloud sonar. Unfortunately, takanori is not available for for for us right now. So we're going to watch the video and comment without him a new.

B

Pic this is like an era stato a principal engineer: claudine thanks for watching my demo, video of class honor for a stem challenge. Here's my contents. So let me first explain what is closer.

B

Clans honor is a tool that allows you to monitor your sava hell's view simple peeing, rest versus close 10 provides service status, five areas and none a lot here. Our problems I'm going to tackle so for several health monitoring. All from monitoring items are too many or too specific when they're. On the other hand, even if you have a bunch of Motoring items, you still feel they are still insufficient. A heavy big data analytics sometimes regas areas, for example, spark and Hadoop.

B

They have a recess manager, butter which is based on fixed memory or CPU capacity, not on bit not based on dynamics of a lot and the fixed timeout complicates your server configurations. Suppose you have a multiple servers across datacenters, then a server on the remote data center should have a longer lettuces than an old non local data centers. So that's how you have to fix various timeouts based on these locations and here's an ideal solution, simple few motoring items clog only, but only when necessary, provide a health status by number.

B

So, let's see my closet could be an ideal solution.

B

So this is my demo on an easy to environment.

B

This is a class or directly, it contains ja, file and log4j and run script. So, let's run like this.

B

So it produces three logs FD HTM, an asana. If f demons, radio detector, it produces five areas, our health services and HTML CSS very produces our predictions and anomalies. So, let's see HTML CSV.

B

As you can see, it could be, it produces action, volumes, observer use and predictions and anna-marie score for each service.

B

So how does it work? This is a crop graph I made based on my sample.

B

Graber means a normally score. This. The range is from 0 point 99 to 1, and blue line means actual berries and orange lines means predictions.

B

So, as you can see, for example, here extra time went up and HTM show a bunch of melanoma race and for each spike also HTM shows no worries, and here akshay virus went down, but also a stim produce detected such a change and.

B

Show some anomalies here and you went off and we when we went down getting and animal score again, even for this small change: small spike, HD em, gee, Dex, nah, I, normally.

B

So here's conclusions confirmed HDM is very good at detecting a slight change, peeing residence time parents. Then these more tests, so I would make more tests like daily parents, weekly patterns and amounts the seasonal early parents, which is more realistic in practice. So thanks for watching my video.

A

Alright, thanks talking to Ari, so one thing I noted from that. Looking at that chart, you know there was the initial jump from from one amplitude to another, and then it jumped back down then jump back up again. That second jump up was no anomaly because it had seen that jump from that level to the other level before so. I just wanted to point that out. Well.

C

I asked him about that. Actually, yeah.

A

C

Said there was an anomaly, but his score was point 975, so it wasn't visible in the chart. So he has a threshold higher than that to highlight you.

A

Know that threshold is something that you can tune to whatever domain you're trying to apply it to you know we we use the threshold a very high level and groc for IT analytics, but I've seen other applications where it should be tuned down further to prevent those false positives.

A

Any other comments or questions on this side. I think this is a interesting application, just using just ping.

D

A

Mean you really don't need much integration with another system to ping it I.

D

Thought of the clever data source- yes, I, wasn't convinced the utility yet I wasn't certain I couldn't tell, but it was clever in the sense of really simple thing to do. I should.

A

Mention that Android data there's another member of our community who has worked with a service called Pingdom, which does which pings things and it's a it's a it just tells you when your services are up or down, and it also has a history of all of those response times, and that was an application that he wrote. You know a year and a half ago with HTM.

A

So this isn't a new idea, but.

D

Also, they did that with a champ too. Yes,.

E

Wonder if you serve a performance better, if you, instead of just a ping, you actually loaded a small jpg so there's at least a normal data load that would be representative of how users would be using the server. So you can see if users are getting good response or.

A

You could do it because the disks.

E

Were really slow right, yeah you'd know that or.

A

Just to test a network responsiveness from one node to another, doing something like that: pushing a small piece of data from one place to another. Yeah.

E

D

Think it's a clever idea to make data as opposed to find data. Yeah true, and so you know the idea. We could extend that to other things like you know, you just suggested bill, but other ways. You could say they doesn't exist here, but we will make it somehow I interesting idea. Well,.

C

It's an interesting future application, perhaps for motor feedback, where your pinging different, pin different ways: pinging different parts, the network out you're moving your eye around yeah. You.

E

C

To get creative from there, he.

D

Writes interesting.

A

C

Thank you, takanori.

A