view thesis/org/first-chapter.org @ 410:e6a7e80f885a

refactor, fix null pointer bug.
author Robert McIntyre <rlm@mit.edu>
date Tue, 18 Mar 2014 22:29:03 -0400
parents 7ee735a836da
children
line wrap: on
line source
1 #+title: =CORTEX=
2 #+author: Robert McIntyre
3 #+email: rlm@mit.edu
4 #+description: Using embodied AI to facilitate Artificial Imagination.
5 #+keywords: AI, clojure, embodiment
6 #+SETUPFILE: ../../aurellem/org/setup.org
7 #+INCLUDE: ../../aurellem/org/level-0.org
8 #+babel: :mkdirp yes :noweb yes :exports both
9 #+OPTIONS: toc:nil, num:nil
11 * Artificial Imagination
13 Imagine watching a video of someone skateboarding. When you watch
14 the video, you can imagine yourself skateboarding, and your
15 knowledge of the human body and its dynamics guides your
16 interpretation of the scene. For example, even if the skateboarder
17 is partially occluded, you can infer the positions of his arms and
18 body from your own knowledge of how your body would be positioned if
19 you were skateboarding. If the skateboarder suffers an accident, you
20 wince in sympathy, imagining the pain your own body would experience
21 if it were in the same situation. This empathy with other people
22 guides our understanding of whatever they are doing because it is a
23 powerful constraint on what is probable and possible. In order to
24 make use of this powerful empathy constraint, I need a system that
25 can generate and make sense of sensory data from the many different
26 senses that humans possess. The two key proprieties of such a system
27 are /embodiment/ and /imagination/.
29 ** What is imagination?
31 One kind of imagination is /sympathetic/ imagination: you imagine
32 yourself in the position of something/someone you are
33 observing. This type of imagination comes into play when you follow
34 along visually when watching someone perform actions, or when you
35 sympathetically grimace when someone hurts themselves. This type of
36 imagination uses the constraints you have learned about your own
37 body to highly constrain the possibilities in whatever you are
38 seeing. It uses all your senses to including your senses of touch,
39 proprioception, etc. Humans are flexible when it comes to "putting
40 themselves in another's shoes," and can sympathetically understand
41 not only other humans, but entities ranging from animals to cartoon
42 characters to [[http://www.youtube.com/watch?v=0jz4HcwTQmU][single dots]] on a screen!
44 Another kind of imagination is /predictive/ imagination: you
45 construct scenes in your mind that are not entirely related to
46 whatever you are observing, but instead are predictions of the
47 future or simply flights of fancy. You use this type of imagination
48 to plan out multi-step actions, or play out dangerous situations in
49 your mind so as to avoid messing them up in reality.
51 Of course, sympathetic and predictive imagination blend into each
52 other and are not completely separate concepts. One dimension along
53 which you can distinguish types of imagination is dependence on raw
54 sense data. Sympathetic imagination is highly constrained by your
55 senses, while predictive imagination can be more or less dependent
56 on your senses depending on how far ahead you imagine. Daydreaming
57 is an extreme form of predictive imagination that wanders through
58 different possibilities without concern for whether they are
59 related to whatever is happening in reality.
61 For this thesis, I will mostly focus on sympathetic imagination and
62 the constraint it provides for understanding sensory data.
64 ** What problems can imagination solve?
66 Consider a video of a cat drinking some water.
68 #+caption: A cat drinking some water. Identifying this action is beyond the state of the art for computers.
69 #+ATTR_LaTeX: width=5cm
70 [[../images/cat-drinking.jpg]]
72 It is currently impossible for any computer program to reliably
73 label such an video as "drinking". I think humans are able to label
74 such video as "drinking" because they imagine /themselves/ as the
75 cat, and imagine putting their face up against a stream of water
76 and sticking out their tongue. In that imagined world, they can
77 feel the cool water hitting their tongue, and feel the water
78 entering their body, and are able to recognize that /feeling/ as
79 drinking. So, the label of the action is not really in the pixels
80 of the image, but is found clearly in a simulation inspired by
81 those pixels. An imaginative system, having been trained on
82 drinking and non-drinking examples and learning that the most
83 important component of drinking is the feeling of water sliding
84 down one's throat, would analyze a video of a cat drinking in the
85 following manner:
87 - Create a physical model of the video by putting a "fuzzy" model
88 of its own body in place of the cat. Also, create a simulation of
89 the stream of water.
91 - Play out this simulated scene and generate imagined sensory
92 experience. This will include relevant muscle contractions, a
93 close up view of the stream from the cat's perspective, and most
94 importantly, the imagined feeling of water entering the mouth.
96 - The action is now easily identified as drinking by the sense of
97 taste alone. The other senses (such as the tongue moving in and
98 out) help to give plausibility to the simulated action. Note that
99 the sense of vision, while critical in creating the simulation,
100 is not critical for identifying the action from the simulation.
102 More generally, I expect imaginative systems to be particularly
103 good at identifying embodied actions in videos.
105 * Cortex
107 The previous example involves liquids, the sense of taste, and
108 imagining oneself as a cat. For this thesis I constrain myself to
109 simpler, more easily digitizable senses and situations.
111 My system, =CORTEX= performs imagination in two different simplified
112 worlds: /worm world/ and /stick-figure world/. In each of these
113 worlds, entities capable of imagination recognize actions by
114 simulating the experience from their own perspective, and then
115 recognizing the action from a database of examples.
117 In order to serve as a framework for experiments in imagination,
118 =CORTEX= requires simulated bodies, worlds, and senses like vision,
119 hearing, touch, proprioception, etc.
121 ** A Video Game Engine takes care of some of the groundwork
123 When it comes to simulation environments, the engines used to
124 create the worlds in video games offer top-notch physics and
125 graphics support. These engines also have limited support for
126 creating cameras and rendering 3D sound, which can be repurposed
127 for vision and hearing respectively. Physics collision detection
128 can be expanded to create a sense of touch.
130 jMonkeyEngine3 is one such engine for creating video games in
131 Java. It uses OpenGL to render to the screen and uses screengraphs
132 to avoid drawing things that do not appear on the screen. It has an
133 active community and several games in the pipeline. The engine was
134 not built to serve any particular game but is instead meant to be
135 used for any 3D game. I chose jMonkeyEngine3 it because it had the
136 most features out of all the open projects I looked at, and because
137 I could then write my code in Clojure, an implementation of LISP
138 that runs on the JVM.
140 ** =CORTEX= Extends jMonkeyEngine3 to implement rich senses
142 Using the game-making primitives provided by jMonkeyEngine3, I have
143 constructed every major human sense except for smell and
144 taste. =CORTEX= also provides an interface for creating creatures
145 in Blender, a 3D modeling environment, and then "rigging" the
146 creatures with senses using 3D annotations in Blender. A creature
147 can have any number of senses, and there can be any number of
148 creatures in a simulation.
150 The senses available in =CORTEX= are:
152 - [[../../cortex/html/vision.html][Vision]]
153 - [[../../cortex/html/hearing.html][Hearing]]
154 - [[../../cortex/html/touch.html][Touch]]
155 - [[../../cortex/html/proprioception.html][Proprioception]]
156 - [[../../cortex/html/movement.html][Muscle Tension]]
158 * A roadmap for =CORTEX= experiments
160 ** Worm World
162 Worms in =CORTEX= are segmented creatures which vary in length and
163 number of segments, and have the senses of vision, proprioception,
164 touch, and muscle tension.
166 #+attr_html: width=755
167 #+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).
168 [[../images/finger-UV.png]]
171 #+begin_html
172 <div class="figure">
173 <center>
174 <video controls="controls" width="550">
175 <source src="../video/worm-touch.ogg" type="video/ogg"
176 preload="none" />
177 </video>
178 <br> <a href="http://youtu.be/RHx2wqzNVcU"> YouTube </a>
179 </center>
180 <p>The worm responds to touch.</p>
181 </div>
182 #+end_html
184 #+begin_html
185 <div class="figure">
186 <center>
187 <video controls="controls" width="550">
188 <source src="../video/test-proprioception.ogg" type="video/ogg"
189 preload="none" />
190 </video>
191 <br> <a href="http://youtu.be/JjdDmyM8b0w"> YouTube </a>
192 </center>
193 <p>Proprioception in a worm. The proprioceptive readout is
194 in the upper left corner of the screen.</p>
195 </div>
196 #+end_html
198 A worm is trained in various actions such as sinusoidal movement,
199 curling, flailing, and spinning by directly playing motor
200 contractions while the worm "feels" the experience. These actions
201 are recorded both as vectors of muscle tension, touch, and
202 proprioceptive data, but also in higher level forms such as
203 frequencies of the various contractions and a symbolic name for the
204 action.
206 Then, the worm watches a video of another worm performing one of
207 the actions, and must judge which action was performed. Normally
208 this would be an extremely difficult problem, but the worm is able
209 to greatly diminish the search space through sympathetic
210 imagination. First, it creates an imagined copy of its body which
211 it observes from a third person point of view. Then for each frame
212 of the video, it maneuvers its simulated body to be in registration
213 with the worm depicted in the video. The physical constraints
214 imposed by the physics simulation greatly decrease the number of
215 poses that have to be tried, making the search feasible. As the
216 imaginary worm moves, it generates imaginary muscle tension and
217 proprioceptive sensations. The worm determines the action not by
218 vision, but by matching the imagined proprioceptive data with
219 previous examples.
221 By using non-visual sensory data such as touch, the worms can also
222 answer body related questions such as "did your head touch your
223 tail?" and "did worm A touch worm B?"
225 The proprioceptive information used for action identification is
226 body-centric, so only the registration step is dependent on point
227 of view, not the identification step. Registration is not specific
228 to any particular action. Thus, action identification can be
229 divided into a point-of-view dependent generic registration step,
230 and a action-specific step that is body-centered and invariant to
231 point of view.
233 ** Stick Figure World
235 This environment is similar to Worm World, except the creatures are
236 more complicated and the actions and questions more varied. It is
237 an experiment to see how far imagination can go in interpreting
238 actions.