Home > Uncategorized > Weekly QuEST Discussion Topics, 17 July

Weekly QuEST Discussion Topics, 17 July

QuEST 17 July 2015

There will be a QuEST meeting this week with the focus on AC^3 (AFRL Conscious Content Curation) – specifically I want to provide a forum for a discussion on the tradespace of ideas/articles/concepts that will tie what we seek in QuEST and what we seek in this ‘incubator’ effort – specifically I want to discuss several articles that the AC3 team has been looking at – like the spectacular article by LeCun, Bengio, Hinto –“Deep Learning”,  the “Ask me anything: Dynamic Memory Network” article by Kumar et al, some slides / work from Socher …:

Incubator:  “AC3”


What:  AFRL Conscious Content Curation C3 – AFRL C3 startup – AC3:

Generate a capability to describe video content (what is the meaning in this data?)  The idea is using state of the art deep learning (possibly provided by the commercial partners / or AFRL generated with our own SMEs) with the AFRL unique artificially conscious wrapper to facilitate the generation of the meaning of the content of a video snippet that a human can better understand and a machine can use to search/locate other potentially stored related data.  The AC3 solution will initially work with images but the vision is to move towards a multi-temporal-scale description for snippets and whole videos.  AFRL prior interaction with commercial entities has convinced us that there is a market requiring such technology (example ESPN, HBO, YouTube…) and AFRL knowledge of the state of the art in describing the content of images/video along with our need to reduce the human capital required to generated analysis of streaming data and our unique approach to artificial consciousness makes this a prime contender for the first incubator.

Why:  3rd offset strategy:  The key to the 3rd offset strategy is autonomy.  What is first required from autonomy is (from Autonomous Horizons) a low impedance human-machine teaming solution.  In a recent chapter on autonomy (situation consciousness) AFRL authors point out to achieve autonomy we need to be able to generate responses for the Unexpected Query (UQ).  The unanticipated stimuli from the perspective of the designing engineer (we need systems to be able to work in operating conditions that will not be completely understood by the designing engineers.  AFRL unique approach to this problem is artificial consciousness. Qualia Exploitation of Sensing Technology (QuEST) is a unique to AFRL and an approach to responding to an unexpected query / autonomy.   This is more generally related to the problem of BIG DATA in that meaning is the more general application but this ADIUx effort is starting with streaming multi-domain imagery data.

There are current limitations of big data approaches in that they look for correlations and most function forensically.  The AC3 effort will focus on making sense of the data in real time (streaming analytics) and will attack the very difficult problem of extracting meaning.  AC3 will do that by the AFRL unique approach to text analysis, cognitive modelling / human state sensing and machine learning that includes artificial consciousness.  The initial goal is to generate a prototype for autonomous generation of meta-data useful for describing the meaning of the streaming data.  It is the QuEST view that meaning is agent centric – it is not intrinsic to the data, so human-machine teaming is a key attribute of AC3.

Computer labeling of an image (auto generation of metadata) for later use in retrieval for content curation is all about meaning.  What is the meaning of this video snippet?

Human cognition is based on a dual set of processes.  Meaning is for a human a combination of the two representations.  Behavior based learning systems should use this insight to generate results that humans can understand.  There are subconscious and conscious processes.  If you train a computer based model based on behavior only the model gets conflicting information as it has not done the estimation of what is driving the current behavior (conscious or subconscious processing).  AC3 suggest that a content curation system that estimates models of the human’s conscious / subconscious states and even goes so far as to blend the two models can better predict behavior / better provide more valuable content

Imagine a system that generates not just a single sentence description but a set of sentences that is then submitted to a blender that attempts to link pieces of alternate description into the winning set of relationships (links).  That winning set is then used to seed a simulation.  The simulation is the artificial consciousness that is used to generate a narrative that we will use as part of the meaning of the stimuli.

Why us:  Technical approach:

AFRL unique approach is artificial consciousness (QuEST unique to AFRL).  The innovation of generating a dual process model of the human user that combines the best of breed in deep learning for the pattern matching big-data subconscious piece BUT also adds a ‘link-based’ relationship (situated) representation to dynamically change the model of the user’s conscious state and then blend the two is where we will bring value.  We have a representation of the meaning of content that is unique.

Categories: Uncategorized
  1. No comments yet.
  1. No trackbacks yet.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google+ photo

You are commenting using your Google+ account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )


Connecting to %s

%d bloggers like this: