Phil 4.25.17

7:00 – 8:30 Research

  • Wikipedia founder Jimmy Wales launches Wikitribune, a large-scale attempt to combat fake news
  • Listening to the BBC Business Daily on Machine Learning. They had an interview with Joanna J Bryson (Scholar). She has an approach for explaining the behavior of AI that seems to involve simulation? Here are some papers that look interesting:
    • Behavior Oriented Design (MIT Dissertation: Intelligence by Design: Principles of Modularity and Coordination for Engineering Complex Adaptive Agents)
    • Learning from Play: Facilitating character design through genetic programming and human mimicry
      • Mimicry and play are fundamental learning processes by which individuals can acquire behaviours, skills and norms. In this paper we utilise these two processes to create new game characters by mimicking and learning from actual human players. We present our approach towards aiding the design process of game characters through the use of genetic programming. The current state of the art in game character design relies heavily on human designers to manually create and edit scripts and rules for game characters. Computational creativity approaches this issue with fully autonomous character generators, replacing most of the design process using black box solutions such as neural networks. Our GP approach to this problem not only mimics actual human play but creates character controllers which can be further authored and developed by a designer. This keeps the designer in the loop while reducing repetitive labour. Our system also provides insights into how players express themselves in games and into deriving appropriate models for representing those insights. We present our framework and preliminary results supporting our claim.
    • Replicators, Lineages and Interactors: One page note on cultural evolution
      • If we adopt the other option and refer to culture itself is the lineage, then the culture itself can evolve since the replicators are the ideas and practices that exist within that culture. However, if it is the culture that is the lineage, we cannot say that it evolves when it takes more territory, in the same way that a species does not evolve with more individuals. Adaptation is presently understood to be about changes in the frequency of replicators, not about absolute numbers of interactors. In sum, cultural evolution (changes of practices within a group) is necessarily a separate process from cultural group selection (changes of the frequency of group-types at a specific location).
    • The behavior-oriented design of modular agent intelligence
    • Should probably cite some of these and a reference to Behavior-Oriented Design in the conclusions section of the paper
  • Continuing Examining the Alternative Media Ecosystem through the Production of Alternative Narratives of Mass Shooting Events on Twitter
    • We collected data using the Twitter Streaming API, tracking on the following terms (shooter, shooting, gunman, gunmen, gunshot, gunshots, shooters, gun shot, gun shots, shootings) for a ten-month period between January 1 and October 5, 2016. This collection resulted in 58M total tweets. We then scoped that data to include only tweets related to alternative narratives of the event—false flag, falseflag, crisis actor, crisisactor, staged, hoax and “1488”.
      • These keywords specify a ‘primary information space’. Bag-of-words of text correlated with each term could make this a linear axis
    • Of 15,150 users who sent at least one tweet with a link, only 1372 sent (over the course to the collection period) tweets citing more than one domain.
      • This is the difference between implicit behaviors (clicking, reading, navigating) and explicit actions. Twitter monitors what people are willing to write
    • Interestingly, the two most influential Domains in Alternative Narrative Tweets Interesting, the two most highly tweeted domains were both associated with significant automated account or “bot” activity. The Real Strategy, an alternative news site with a conspiracy theory orientation, is the most tweeted domain in our dataset (by far). The temporal signature of tweets citing this domain reveals a consistent pattern of coordinated bursts of activity at regular intervals generated by 200 accounts that appear to be connected to each other (via following relationships) and coordinated through an external tool.
      • There is clearly a desire to have a greater effect through the use of bots. Two questions: 1) How does this work? 2) How did this emerge?
    • The InfoWars domain, an alternative news website that focuses on Alt-Right and conspiracy theory themes, was the second-most tweeted domain, but as (Figure 1) shows it was only tenuously connected to one other node.
      • Why? Is InforWars more polarized? Is it using something other than Twitter?
      • Infowars Inbound links
        Domain score Domain trust score Domain Backlinks IP Address Country First seen Last seen
        0 0 1857029 us 2015-09-28 2017-03-26
        4 4 1335835 us 2014-01-19 2017-03-21
        33 39 648958 us 2013-06-07 2017-03-25
        1 0 346153 us 2014-01-19 2017-03-21
        13 31 182060 cz 2013-06-07 2017-03-26
        12 30 151778 us 2016-06-27 2017-03-22
        1 0 92766 us 2014-11-14 2017-03-23
        4 29 49288 us 2015-02-04 2017-03-26
        14 30 47195 us 2014-10-02 2017-03-20
        1 0 43748 us 2016-06-08 2017-03-24

9:00 – 5:30 BRC

  • John is having trouble getting Linux running on the laptop
    • No luck. Re-submitting for an Alienware deskside
  • Back to getting the temporal coherence. last try to finish up, then switching to fitness landscape optimization, which I dreamed about last night
  • Finished coherence! Had to include a state check for a timeline to see if a DIRTY state had been touched with an update. If not, then the timeline is set to CLOSED. If a new cluster appears that would have had some overlap, a new timeline is created anyway. This could be an optional behavior.
    • Still need to test rigorously across multiple data sets
  • Long scrum, then ML meeting.
    • Hard tasks
      • TF server set up to work in our environment
      • Pre-calculated models to speed up training from research browser
      • T-SNE or other mapping of returned CSE text to support exploration
      • Fast, on-the-fly classification and entity extraction within the research browser framework. Plus interactive training
      • NMF (or other) topic extraction tied to human labeling and curation, plus cross-user validation of topics
  • Poster with Aaron later? Yep. Couple of hours. Done?
  • Oh, just why? Spent an hour on this before going brute force:
    def get_last_cluster(self) -> ClusterSample:
        # return self._cluster_dict[self._cluster_dict.keys()[-1] TODO: This should work
        toReturn = None
        for key in self._cluster_dict:
            toReturn = self._cluster_dict[key]
        return toReturn
  • Walked through some gradient descent regression code with Bob. More tomorrow?
  • Got the new sort working with Aaron. Much faster progress as a pair