annotate thesis/org/first-chapter.org @ 445:47cfbe84f00e

complete images in first third of first chapter.
author Robert McIntyre <rlm@mit.edu>
date Tue, 25 Mar 2014 03:18:04 -0400
parents 5205535237fb
children
rev   line source
rlm@401 1 #+title: =CORTEX=
rlm@401 2 #+author: Robert McIntyre
rlm@401 3 #+email: rlm@mit.edu
rlm@401 4 #+description: Using embodied AI to facilitate Artificial Imagination.
rlm@401 5 #+keywords: AI, clojure, embodiment
rlm@401 6 #+SETUPFILE: ../../aurellem/org/setup.org
rlm@401 7 #+INCLUDE: ../../aurellem/org/level-0.org
rlm@401 8 #+babel: :mkdirp yes :noweb yes :exports both
rlm@401 9 #+OPTIONS: toc:nil, num:nil
rlm@401 10
rlm@401 11 * Artificial Imagination
rlm@401 12 Imagine watching a video of someone skateboarding. When you watch
rlm@401 13 the video, you can imagine yourself skateboarding, and your
rlm@401 14 knowledge of the human body and its dynamics guides your
rlm@401 15 interpretation of the scene. For example, even if the skateboarder
rlm@401 16 is partially occluded, you can infer the positions of his arms and
rlm@401 17 body from your own knowledge of how your body would be positioned if
rlm@401 18 you were skateboarding. If the skateboarder suffers an accident, you
rlm@401 19 wince in sympathy, imagining the pain your own body would experience
rlm@401 20 if it were in the same situation. This empathy with other people
rlm@401 21 guides our understanding of whatever they are doing because it is a
rlm@401 22 powerful constraint on what is probable and possible. In order to
rlm@401 23 make use of this powerful empathy constraint, I need a system that
rlm@401 24 can generate and make sense of sensory data from the many different
rlm@401 25 senses that humans possess. The two key proprieties of such a system
rlm@401 26 are /embodiment/ and /imagination/.
rlm@401 27
rlm@401 28 ** What is imagination?
rlm@401 29
rlm@401 30 One kind of imagination is /sympathetic/ imagination: you imagine
rlm@401 31 yourself in the position of something/someone you are
rlm@401 32 observing. This type of imagination comes into play when you follow
rlm@401 33 along visually when watching someone perform actions, or when you
rlm@401 34 sympathetically grimace when someone hurts themselves. This type of
rlm@401 35 imagination uses the constraints you have learned about your own
rlm@401 36 body to highly constrain the possibilities in whatever you are
rlm@401 37 seeing. It uses all your senses to including your senses of touch,
rlm@401 38 proprioception, etc. Humans are flexible when it comes to "putting
rlm@401 39 themselves in another's shoes," and can sympathetically understand
rlm@401 40 not only other humans, but entities ranging from animals to cartoon
rlm@401 41 characters to [[http://www.youtube.com/watch?v=0jz4HcwTQmU][single dots]] on a screen!
rlm@401 42
rlm@429 43 # and can infer intention from the actions of not only other humans,
rlm@429 44 # but also animals, cartoon characters, and even abstract moving dots
rlm@429 45 # on a screen!
rlm@429 46
rlm@401 47 Another kind of imagination is /predictive/ imagination: you
rlm@401 48 construct scenes in your mind that are not entirely related to
rlm@401 49 whatever you are observing, but instead are predictions of the
rlm@401 50 future or simply flights of fancy. You use this type of imagination
rlm@401 51 to plan out multi-step actions, or play out dangerous situations in
rlm@401 52 your mind so as to avoid messing them up in reality.
rlm@401 53
rlm@401 54 Of course, sympathetic and predictive imagination blend into each
rlm@401 55 other and are not completely separate concepts. One dimension along
rlm@401 56 which you can distinguish types of imagination is dependence on raw
rlm@401 57 sense data. Sympathetic imagination is highly constrained by your
rlm@401 58 senses, while predictive imagination can be more or less dependent
rlm@401 59 on your senses depending on how far ahead you imagine. Daydreaming
rlm@401 60 is an extreme form of predictive imagination that wanders through
rlm@401 61 different possibilities without concern for whether they are
rlm@401 62 related to whatever is happening in reality.
rlm@401 63
rlm@401 64 For this thesis, I will mostly focus on sympathetic imagination and
rlm@401 65 the constraint it provides for understanding sensory data.
rlm@401 66
rlm@401 67 ** What problems can imagination solve?
rlm@401 68
rlm@401 69 Consider a video of a cat drinking some water.
rlm@401 70
rlm@401 71 #+caption: A cat drinking some water. Identifying this action is beyond the state of the art for computers.
rlm@401 72 #+ATTR_LaTeX: width=5cm
rlm@401 73 [[../images/cat-drinking.jpg]]
rlm@401 74
rlm@401 75 It is currently impossible for any computer program to reliably
rlm@401 76 label such an video as "drinking". I think humans are able to label
rlm@401 77 such video as "drinking" because they imagine /themselves/ as the
rlm@401 78 cat, and imagine putting their face up against a stream of water
rlm@401 79 and sticking out their tongue. In that imagined world, they can
rlm@401 80 feel the cool water hitting their tongue, and feel the water
rlm@401 81 entering their body, and are able to recognize that /feeling/ as
rlm@401 82 drinking. So, the label of the action is not really in the pixels
rlm@401 83 of the image, but is found clearly in a simulation inspired by
rlm@401 84 those pixels. An imaginative system, having been trained on
rlm@401 85 drinking and non-drinking examples and learning that the most
rlm@401 86 important component of drinking is the feeling of water sliding
rlm@401 87 down one's throat, would analyze a video of a cat drinking in the
rlm@401 88 following manner:
rlm@401 89
rlm@401 90 - Create a physical model of the video by putting a "fuzzy" model
rlm@401 91 of its own body in place of the cat. Also, create a simulation of
rlm@401 92 the stream of water.
rlm@401 93
rlm@401 94 - Play out this simulated scene and generate imagined sensory
rlm@401 95 experience. This will include relevant muscle contractions, a
rlm@401 96 close up view of the stream from the cat's perspective, and most
rlm@401 97 importantly, the imagined feeling of water entering the mouth.
rlm@401 98
rlm@401 99 - The action is now easily identified as drinking by the sense of
rlm@401 100 taste alone. The other senses (such as the tongue moving in and
rlm@401 101 out) help to give plausibility to the simulated action. Note that
rlm@401 102 the sense of vision, while critical in creating the simulation,
rlm@401 103 is not critical for identifying the action from the simulation.
rlm@401 104
rlm@401 105 More generally, I expect imaginative systems to be particularly
rlm@401 106 good at identifying embodied actions in videos.
rlm@401 107
rlm@401 108 * Cortex
rlm@401 109
rlm@401 110 The previous example involves liquids, the sense of taste, and
rlm@401 111 imagining oneself as a cat. For this thesis I constrain myself to
rlm@401 112 simpler, more easily digitizable senses and situations.
rlm@401 113
rlm@401 114 My system, =CORTEX= performs imagination in two different simplified
rlm@401 115 worlds: /worm world/ and /stick-figure world/. In each of these
rlm@401 116 worlds, entities capable of imagination recognize actions by
rlm@401 117 simulating the experience from their own perspective, and then
rlm@401 118 recognizing the action from a database of examples.
rlm@401 119
rlm@401 120 In order to serve as a framework for experiments in imagination,
rlm@401 121 =CORTEX= requires simulated bodies, worlds, and senses like vision,
rlm@401 122 hearing, touch, proprioception, etc.
rlm@401 123
rlm@401 124 ** A Video Game Engine takes care of some of the groundwork
rlm@401 125
rlm@401 126 When it comes to simulation environments, the engines used to
rlm@401 127 create the worlds in video games offer top-notch physics and
rlm@401 128 graphics support. These engines also have limited support for
rlm@401 129 creating cameras and rendering 3D sound, which can be repurposed
rlm@401 130 for vision and hearing respectively. Physics collision detection
rlm@401 131 can be expanded to create a sense of touch.
rlm@401 132
rlm@401 133 jMonkeyEngine3 is one such engine for creating video games in
rlm@401 134 Java. It uses OpenGL to render to the screen and uses screengraphs
rlm@401 135 to avoid drawing things that do not appear on the screen. It has an
rlm@401 136 active community and several games in the pipeline. The engine was
rlm@401 137 not built to serve any particular game but is instead meant to be
rlm@401 138 used for any 3D game. I chose jMonkeyEngine3 it because it had the
rlm@401 139 most features out of all the open projects I looked at, and because
rlm@401 140 I could then write my code in Clojure, an implementation of LISP
rlm@401 141 that runs on the JVM.
rlm@401 142
rlm@401 143 ** =CORTEX= Extends jMonkeyEngine3 to implement rich senses
rlm@401 144
rlm@401 145 Using the game-making primitives provided by jMonkeyEngine3, I have
rlm@401 146 constructed every major human sense except for smell and
rlm@401 147 taste. =CORTEX= also provides an interface for creating creatures
rlm@401 148 in Blender, a 3D modeling environment, and then "rigging" the
rlm@401 149 creatures with senses using 3D annotations in Blender. A creature
rlm@401 150 can have any number of senses, and there can be any number of
rlm@401 151 creatures in a simulation.
rlm@401 152
rlm@401 153 The senses available in =CORTEX= are:
rlm@401 154
rlm@401 155 - [[../../cortex/html/vision.html][Vision]]
rlm@401 156 - [[../../cortex/html/hearing.html][Hearing]]
rlm@401 157 - [[../../cortex/html/touch.html][Touch]]
rlm@401 158 - [[../../cortex/html/proprioception.html][Proprioception]]
rlm@401 159 - [[../../cortex/html/movement.html][Muscle Tension]]
rlm@401 160
rlm@401 161 * A roadmap for =CORTEX= experiments
rlm@401 162
rlm@401 163 ** Worm World
rlm@401 164
rlm@401 165 Worms in =CORTEX= are segmented creatures which vary in length and
rlm@401 166 number of segments, and have the senses of vision, proprioception,
rlm@401 167 touch, and muscle tension.
rlm@401 168
rlm@401 169 #+attr_html: width=755
rlm@401 170 #+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).
rlm@401 171 [[../images/finger-UV.png]]
rlm@401 172
rlm@401 173
rlm@401 174 #+begin_html
rlm@401 175 <div class="figure">
rlm@401 176 <center>
rlm@401 177 <video controls="controls" width="550">
rlm@401 178 <source src="../video/worm-touch.ogg" type="video/ogg"
rlm@401 179 preload="none" />
rlm@401 180 </video>
rlm@401 181 <br> <a href="http://youtu.be/RHx2wqzNVcU"> YouTube </a>
rlm@401 182 </center>
rlm@401 183 <p>The worm responds to touch.</p>
rlm@401 184 </div>
rlm@401 185 #+end_html
rlm@401 186
rlm@401 187 #+begin_html
rlm@401 188 <div class="figure">
rlm@401 189 <center>
rlm@401 190 <video controls="controls" width="550">
rlm@401 191 <source src="../video/test-proprioception.ogg" type="video/ogg"
rlm@401 192 preload="none" />
rlm@401 193 </video>
rlm@401 194 <br> <a href="http://youtu.be/JjdDmyM8b0w"> YouTube </a>
rlm@401 195 </center>
rlm@401 196 <p>Proprioception in a worm. The proprioceptive readout is
rlm@401 197 in the upper left corner of the screen.</p>
rlm@401 198 </div>
rlm@401 199 #+end_html
rlm@401 200
rlm@401 201 A worm is trained in various actions such as sinusoidal movement,
rlm@401 202 curling, flailing, and spinning by directly playing motor
rlm@401 203 contractions while the worm "feels" the experience. These actions
rlm@401 204 are recorded both as vectors of muscle tension, touch, and
rlm@401 205 proprioceptive data, but also in higher level forms such as
rlm@401 206 frequencies of the various contractions and a symbolic name for the
rlm@401 207 action.
rlm@401 208
rlm@401 209 Then, the worm watches a video of another worm performing one of
rlm@401 210 the actions, and must judge which action was performed. Normally
rlm@401 211 this would be an extremely difficult problem, but the worm is able
rlm@401 212 to greatly diminish the search space through sympathetic
rlm@401 213 imagination. First, it creates an imagined copy of its body which
rlm@401 214 it observes from a third person point of view. Then for each frame
rlm@401 215 of the video, it maneuvers its simulated body to be in registration
rlm@401 216 with the worm depicted in the video. The physical constraints
rlm@401 217 imposed by the physics simulation greatly decrease the number of
rlm@401 218 poses that have to be tried, making the search feasible. As the
rlm@401 219 imaginary worm moves, it generates imaginary muscle tension and
rlm@401 220 proprioceptive sensations. The worm determines the action not by
rlm@401 221 vision, but by matching the imagined proprioceptive data with
rlm@401 222 previous examples.
rlm@401 223
rlm@401 224 By using non-visual sensory data such as touch, the worms can also
rlm@401 225 answer body related questions such as "did your head touch your
rlm@401 226 tail?" and "did worm A touch worm B?"
rlm@401 227
rlm@401 228 The proprioceptive information used for action identification is
rlm@401 229 body-centric, so only the registration step is dependent on point
rlm@401 230 of view, not the identification step. Registration is not specific
rlm@401 231 to any particular action. Thus, action identification can be
rlm@401 232 divided into a point-of-view dependent generic registration step,
rlm@401 233 and a action-specific step that is body-centered and invariant to
rlm@401 234 point of view.
rlm@401 235
rlm@401 236 ** Stick Figure World
rlm@401 237
rlm@401 238 This environment is similar to Worm World, except the creatures are
rlm@401 239 more complicated and the actions and questions more varied. It is
rlm@401 240 an experiment to see how far imagination can go in interpreting
rlm@401 241 actions.