#+title: =CORTEX=

 #+author: Robert McIntyre

 #+email: rlm@mit.edu

 #+description: Using embodied AI to facilitate Artificial Imagination.

 #+keywords: AI, clojure, embodiment

 #+SETUPFILE: ../../aurellem/org/setup.org

 #+INCLUDE: ../../aurellem/org/level-0.org

 #+babel: :mkdirp yes :noweb yes :exports both

 #+OPTIONS: toc:nil, num:nil

 

 * Artificial Imagination

   Imagine watching a video of someone skateboarding. When you watch

   the video, you can imagine yourself skateboarding, and your

   knowledge of the human body and its dynamics guides your

   interpretation of the scene. For example, even if the skateboarder

   is partially occluded, you can infer the positions of his arms and

   body from your own knowledge of how your body would be positioned if

   you were skateboarding. If the skateboarder suffers an accident, you

   wince in sympathy, imagining the pain your own body would experience

   if it were in the same situation. This empathy with other people

   guides our understanding of whatever they are doing because it is a

   powerful constraint on what is probable and possible. In order to

   make use of this powerful empathy constraint, I need a system that

   can generate and make sense of sensory data from the many different

   senses that humans possess. The two key proprieties of such a system

   are /embodiment/ and /imagination/.

 

 ** What is imagination?

 

    One kind of imagination is /sympathetic/ imagination: you imagine

    yourself in the position of something/someone you are

    observing. This type of imagination comes into play when you follow

    along visually when watching someone perform actions, or when you

    sympathetically grimace when someone hurts themselves. This type of

    imagination uses the constraints you have learned about your own

    body to highly constrain the possibilities in whatever you are

    seeing. It uses all your senses to including your senses of touch,

    proprioception, etc. Humans are flexible when it comes to "putting

    themselves in another's shoes," and can sympathetically understand

    not only other humans, but entities ranging from animals to cartoon

    characters to [[http://www.youtube.com/watch?v=0jz4HcwTQmU][single dots]] on a screen!

 

 # and can infer intention from the actions of not only other humans,

 # but also animals, cartoon characters, and even abstract moving dots

 # on a screen!

 

    Another kind of imagination is /predictive/ imagination: you

    construct scenes in your mind that are not entirely related to

    whatever you are observing, but instead are predictions of the

    future or simply flights of fancy. You use this type of imagination

    to plan out multi-step actions, or play out dangerous situations in

    your mind so as to avoid messing them up in reality.

 

    Of course, sympathetic and predictive imagination blend into each

    other and are not completely separate concepts. One dimension along

    which you can distinguish types of imagination is dependence on raw

    sense data. Sympathetic imagination is highly constrained by your

    senses, while predictive imagination can be more or less dependent

    on your senses depending on how far ahead you imagine. Daydreaming

    is an extreme form of predictive imagination that wanders through

    different possibilities without concern for whether they are

    related to whatever is happening in reality.

 

    For this thesis, I will mostly focus on sympathetic imagination and

    the constraint it provides for understanding sensory data.

    

 ** What problems can imagination solve?

 

    Consider a video of a cat drinking some water.

 

    #+caption: A cat drinking some water. Identifying this action is beyond the state of the art for computers.

    #+ATTR_LaTeX: width=5cm

    [[../images/cat-drinking.jpg]]

 

    It is currently impossible for any computer program to reliably

    label such an video as "drinking". I think humans are able to label

    such video as "drinking" because they imagine /themselves/ as the

    cat, and imagine putting their face up against a stream of water

    and sticking out their tongue. In that imagined world, they can

    feel the cool water hitting their tongue, and feel the water

    entering their body, and are able to recognize that /feeling/ as

    drinking. So, the label of the action is not really in the pixels

    of the image, but is found clearly in a simulation inspired by

    those pixels. An imaginative system, having been trained on

    drinking and non-drinking examples and learning that the most

    important component of drinking is the feeling of water sliding

    down one's throat, would analyze a video of a cat drinking in the

    following manner:

    

    - Create a physical model of the video by putting a "fuzzy" model

      of its own body in place of the cat. Also, create a simulation of

      the stream of water.

 

    - Play out this simulated scene and generate imagined sensory

      experience. This will include relevant muscle contractions, a

      close up view of the stream from the cat's perspective, and most

      importantly, the imagined feeling of water entering the mouth.

 

    - The action is now easily identified as drinking by the sense of

      taste alone. The other senses (such as the tongue moving in and

      out) help to give plausibility to the simulated action. Note that

      the sense of vision, while critical in creating the simulation,

      is not critical for identifying the action from the simulation.

 

    More generally, I expect imaginative systems to be particularly

    good at identifying embodied actions in videos.

 

 * Cortex

 

   The previous example involves liquids, the sense of taste, and

   imagining oneself as a cat. For this thesis I constrain myself to

   simpler, more easily digitizable senses and situations.

 

   My system, =CORTEX= performs imagination in two different simplified

   worlds: /worm world/ and /stick-figure world/. In each of these

   worlds, entities capable of imagination recognize actions by

   simulating the experience from their own perspective, and then

   recognizing the action from a database of examples.

 

   In order to serve as a framework for experiments in imagination,

   =CORTEX= requires simulated bodies, worlds, and senses like vision,

   hearing, touch, proprioception, etc.

 

 ** A Video Game Engine takes care of some of the groundwork

 

    When it comes to simulation environments, the engines used to

    create the worlds in video games offer top-notch physics and

    graphics support. These engines also have limited support for

    creating cameras and rendering 3D sound, which can be repurposed

    for vision and hearing respectively. Physics collision detection

    can be expanded to create a sense of touch.

    

    jMonkeyEngine3 is one such engine for creating video games in

    Java. It uses OpenGL to render to the screen and uses screengraphs

    to avoid drawing things that do not appear on the screen. It has an

    active community and several games in the pipeline. The engine was

    not built to serve any particular game but is instead meant to be

    used for any 3D game. I chose jMonkeyEngine3 it because it had the

    most features out of all the open projects I looked at, and because

    I could then write my code in Clojure, an implementation of LISP

    that runs on the JVM.

 

 ** =CORTEX= Extends jMonkeyEngine3 to implement rich senses

 

    Using the game-making primitives provided by jMonkeyEngine3, I have

    constructed every major human sense except for smell and

    taste. =CORTEX= also provides an interface for creating creatures

    in Blender, a 3D modeling environment, and then "rigging" the

    creatures with senses using 3D annotations in Blender. A creature

    can have any number of senses, and there can be any number of

    creatures in a simulation.

    

    The senses available in =CORTEX= are:

 

    - [[../../cortex/html/vision.html][Vision]]

    - [[../../cortex/html/hearing.html][Hearing]]

    - [[../../cortex/html/touch.html][Touch]]

    - [[../../cortex/html/proprioception.html][Proprioception]]

    - [[../../cortex/html/movement.html][Muscle Tension]]

 

 * A roadmap for =CORTEX= experiments

 

 ** Worm World

 

    Worms in =CORTEX= are segmented creatures which vary in length and

    number of segments, and have the senses of vision, proprioception,

    touch, and muscle tension.

 

 #+attr_html: width=755

 #+caption: This is the tactile-sensor-profile for the upper segment of a worm. It defines regions of high touch sensitivity (where there are many white pixels) and regions of low sensitivity (where white pixels are sparse).

 [[../images/finger-UV.png]]

 

 

 #+begin_html

 <div class="figure">

   <center>

     <video controls="controls" width="550">

       <source src="../video/worm-touch.ogg" type="video/ogg"

 	      preload="none" />

     </video>

     <br> <a href="http://youtu.be/RHx2wqzNVcU"> YouTube </a>

   </center>

   <p>The worm responds to touch.</p>

 </div>

 #+end_html

 

 #+begin_html

 <div class="figure">

   <center>

     <video controls="controls" width="550">

       <source src="../video/test-proprioception.ogg" type="video/ogg"

 	      preload="none" />

     </video>

     <br> <a href="http://youtu.be/JjdDmyM8b0w"> YouTube </a>

   </center>

   <p>Proprioception in a worm. The proprioceptive readout is

     in the upper left corner of the screen.</p>

 </div>

 #+end_html

 

    A worm is trained in various actions such as sinusoidal movement,

    curling, flailing, and spinning by directly playing motor

    contractions while the worm "feels" the experience. These actions

    are recorded both as vectors of muscle tension, touch, and

    proprioceptive data, but also in higher level forms such as

    frequencies of the various contractions and a symbolic name for the

    action.

 

    Then, the worm watches a video of another worm performing one of

    the actions, and must judge which action was performed. Normally

    this would be an extremely difficult problem, but the worm is able

    to greatly diminish the search space through sympathetic

    imagination. First, it creates an imagined copy of its body which

    it observes from a third person point of view. Then for each frame

    of the video, it maneuvers its simulated body to be in registration

    with the worm depicted in the video. The physical constraints

    imposed by the physics simulation greatly decrease the number of

    poses that have to be tried, making the search feasible. As the

    imaginary worm moves, it generates imaginary muscle tension and

    proprioceptive sensations. The worm determines the action not by

    vision, but by matching the imagined proprioceptive data with

    previous examples.

 

    By using non-visual sensory data such as touch, the worms can also

    answer body related questions such as "did your head touch your

    tail?" and "did worm A touch worm B?"

 

    The proprioceptive information used for action identification is

    body-centric, so only the registration step is dependent on point

    of view, not the identification step. Registration is not specific

    to any particular action. Thus, action identification can be

    divided into a point-of-view dependent generic registration step,

    and a action-specific step that is body-centered and invariant to

    point of view.

 

 ** Stick Figure World

 

    This environment is similar to Worm World, except the creatures are

    more complicated and the actions and questions more varied. It is

    an experiment to see how far imagination can go in interpreting

    actions.