Numenta HTM Challenge Submissions, 9 Nov 2015

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Semantic Anomaly Detection with the Cortical.io Retina API and the HTM - Numenta HTM Challenge 2015

Description

This is the Cortical.io submission for the Numenta HTM Challenge 2015. It's an intelligent Twitter monitoring service that detects semantic anomalies - meaning that it uncovers unusual changes in topics in an individual's Twitter feeds. It has been configured to monitor the Twitter accounts of several top US Presidential candidates so users can investigate what they post are about and use the power of the HTM and Cortical's Retina API to learn more about their elected officials.

More information at:
http://devpost.com/software/news-reader-0wkjli
http://www.cortical.io/
http://numenta.com/

A

Hi I'm taylor from cortical, I o in vienna, austria- and this is our team submission for the numenta hdm challenge. We wanted to show how to combine the hdm with our retina api, which works on text data, so we built a tool that analyzes the semantics of twitter posts made by several us presidential candidates. So the data source is the text content of their tweets, which is publicly available, and then we process this text, data and encode it using our written api into sdrs that the htm can understand and then using the hdm.

A

We detect semantic anomalies or unexpected changes in the text. Content of the candidate's tweets. So first some quick background information about what we do at cortical and how this relates to the htm.

A

Basically, we provide an api that encodes text into a numeric representation, and this encoding process is similar to the way information is distributed throughout different areas of the brain.

A

So on the left here, you can see a graphical representation of how we store semantic information in a 128x128 matrix with individual bits of the matrix, representing a specific meaning and with related pieces of information. Being stored close to each other, uh just like in the brain, and so we refer to these representations as semantic fingerprints and they're kind of a sparse distributed representation or word str, as some people call them, and we can compute one of these sdrs for any kind of text in a variety of languages.

A

So once you have one of these semantic fingerprint representations of a piece of text, there are a lot of things you can do with it, and one of the cool things about our sdis is that they're compatible with the hdm and can be fed directly into the temporal pool.

A

So in a way, our api acts as a text encoder and spatial puller in one, and once you start creating sdrs for text, you can use them to let the htm learn patterns in human language, detect anomalies and also make predictions.

A

So that's what we did with this application.

A

First, we extracted the text from the twitter feeds of six presidential candidates shown here in no particular order, and then we group the tweets per candidate by day and create a semantic fingerprint for that group of tweets.

A

Then we input those fingerprints into the htm and graph the anomaly scores that it outputs by day and because we use semantic fingerprints as input for the htm, we're not graphing the anomaly scores based on the volume of tweets, but by the actual semantic content of them. So what the candidates are actually talking about, so the higher the anomaly score, the more unexpected the content of the twitter post was for that day.

A

So when you see peaks in the graph like here or here, the hdm determined that whatever the candidates post about on those days was unusual for that candidate and for reference. We also plotted a few real world events on the graph as vertical red lines.

A

So you can see how detected anomalies correspond with events like the candidates, making official announcements, holding campaign, rallies and taking part in debates, and so the graphs are interactive and you can move your mouse over data points to see the keywords and exact anomaly scores for those days and then you can also click on a data point to see the full text of the tweets. For that day.

A

And so for most of the candidates, you can see that at the beginning of the graphs, the anomaly scores were initially quite high.

A

This is because the html was still learning a pattern of topics that they post about and then, after learning a pattern, the anomaly scores tend to drop off quite a bit, especially, you can see that with bernie sanders here.

A

But, for example, you can see with hillary clinton the top graph as soon as she officially announced her candidacy in mid-april. The htm immediately detected a change in what she was posting about on her twitter account.

A

And then it quickly adjusted to this new pattern of topics interfeed with only minor anomalies, popping up after that, like this one here that seems to correspond with a rally that you held on labor day.

A

So as an additional feature, we also added the ability to filter the twitter feed semantically by social and economic issues.

A

This is done by working on the fingerprint level of the tweets to determine what the candidates are talking about and not just simple keyword matching. So when you click these buttons, it reduces the twitter feeds to only posts, have a high similarity to these topics, and then we train separate htms for each candidate on these feeds.

A

So the anomaly scores are based only on the filtered data, so you can see that certain candidates tend to post more about certain issues and for some candidates it's actually an anomaly when they do talk about certain.

A

A

And the entire application is available live right now at this url I will put the link in the video description, and so we encourage people to take a look at it, draw their own conclusions about the anomalies detected and hopefully use it as a way to get a clearer picture about how politicians speak in the media.

A

So that's it semantic anomaly. Detection with the corticlio retina api and the hdm, and we at cortical are big fans of the hdm and we're very much inspired by the work that numenta does. So, if you have any questions about how to integrate our software with the hdm, then please feel free to contact us thanks.