Mercurial > cortex

annotate thesis/cortex.org @ 447:284316604be0

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
minor changes from Dylan.
author Robert McIntyre <rlm@mit.edu>
date Tue, 25 Mar 2014 11:30:15 -0400
parents 3e91585b2a1c
children af13fc73e851
rev   line source
rlm@425 1 #+title: =CORTEX=
rlm@425 2 #+author: Robert McIntyre
rlm@425 3 #+email: rlm@mit.edu
rlm@425 4 #+description: Using embodied AI to facilitate Artificial Imagination.
rlm@425 5 #+keywords: AI, clojure, embodiment
rlm@422 6
rlm@437 7
rlm@439 8 * Empathy and Embodiment as problem solving strategies
rlm@437 9
rlm@437 10 By the end of this thesis, you will have seen a novel approach to
rlm@437 11 interpreting video using embodiment and empathy. You will have also
rlm@437 12 seen one way to efficiently implement empathy for embodied
rlm@447 13 creatures. Finally, you will become familiar with =CORTEX=, a system
rlm@447 14 for designing and simulating creatures with rich senses, which you
rlm@447 15 may choose to use in your own research.
rlm@437 16
rlm@441 17 This is the core vision of my thesis: That one of the important ways
rlm@441 18 in which we understand others is by imagining ourselves in their
rlm@441 19 position and emphatically feeling experiences relative to our own
rlm@441 20 bodies. By understanding events in terms of our own previous
rlm@441 21 corporeal experience, we greatly constrain the possibilities of what
rlm@441 22 would otherwise be an unwieldy exponential search. This extra
rlm@441 23 constraint can be the difference between easily understanding what
rlm@441 24 is happening in a video and being completely lost in a sea of
rlm@441 25 incomprehensible color and movement.
rlm@435 26
rlm@436 27 ** Recognizing actions in video is extremely difficult
rlm@437 28
rlm@447 29 Consider for example the problem of determining what is happening
rlm@447 30 in a video of which this is one frame:
rlm@437 31
rlm@441 32 #+caption: A cat drinking some water. Identifying this action is
rlm@441 33 #+caption: beyond the state of the art for computers.
rlm@441 34 #+ATTR_LaTeX: :width 7cm
rlm@441 35 [[./images/cat-drinking.jpg]]
rlm@441 36
rlm@441 37 It is currently impossible for any computer program to reliably
rlm@447 38 label such a video as ``drinking''. And rightly so -- it is a very
rlm@441 39 hard problem! What features can you describe in terms of low level
rlm@441 40 functions of pixels that can even begin to describe at a high level
rlm@441 41 what is happening here?
rlm@437 42
rlm@447 43 Or suppose that you are building a program that recognizes chairs.
rlm@447 44 How could you ``see'' the chair in figure \ref{invisible-chair} and
rlm@447 45 figure \ref{hidden-chair}?
rlm@441 46
rlm@441 47 #+caption: When you look at this, do you think ``chair''? I certainly do.
rlm@441 48 #+name: invisible-chair
rlm@441 49 #+ATTR_LaTeX: :width 10cm
rlm@441 50 [[./images/invisible-chair.png]]
rlm@441 51
rlm@441 52 #+caption: The chair in this image is quite obvious to humans, but I
rlm@441 53 #+caption: doubt that any computer program can find it.
rlm@441 54 #+name: hidden-chair
rlm@441 55 #+ATTR_LaTeX: :width 10cm
rlm@441 56 [[./images/fat-person-sitting-at-desk.jpg]]
rlm@441 57
rlm@441 58 Finally, how is it that you can easily tell the difference between
rlm@441 59 how the girls /muscles/ are working in figure \ref{girl}?
rlm@441 60
rlm@441 61 #+caption: The mysterious ``common sense'' appears here as you are able
rlm@441 62 #+caption: to discern the difference in how the girl's arm muscles
rlm@441 63 #+caption: are activated between the two images.
rlm@441 64 #+name: girl
rlm@441 65 #+ATTR_LaTeX: :width 10cm
rlm@441 66 [[./images/wall-push.png]]
rlm@437 67
rlm@441 68 Each of these examples tells us something about what might be going
rlm@441 69 on in our minds as we easily solve these recognition problems.
rlm@441 70
rlm@441 71 The hidden chairs show us that we are strongly triggered by cues
rlm@447 72 relating to the position of human bodies, and that we can determine
rlm@447 73 the overall physical configuration of a human body even if much of
rlm@447 74 that body is occluded.
rlm@437 75
rlm@441 76 The picture of the girl pushing against the wall tells us that we
rlm@441 77 have common sense knowledge about the kinetics of our own bodies.
rlm@441 78 We know well how our muscles would have to work to maintain us in
rlm@441 79 most positions, and we can easily project this self-knowledge to
rlm@441 80 imagined positions triggered by images of the human body.
rlm@441 81
rlm@441 82 ** =EMPATH= neatly solves recognition problems
rlm@441 83
rlm@441 84 I propose a system that can express the types of recognition
rlm@441 85 problems above in a form amenable to computation. It is split into
rlm@441 86 four parts:
rlm@441 87
rlm@447 88 - Free/Guided Play (Training) :: The creature moves around and
rlm@447 89 experiences the world through its unique perspective. Many
rlm@447 90 otherwise complicated actions are easily described in the
rlm@447 91 language of a full suite of body-centered, rich senses. For
rlm@447 92 example, drinking is the feeling of water sliding down your
rlm@447 93 throat, and cooling your insides. It's often accompanied by
rlm@447 94 bringing your hand close to your face, or bringing your face
rlm@447 95 close to water. Sitting down is the feeling of bending your
rlm@447 96 knees, activating your quadriceps, then feeling a surface with
rlm@447 97 your bottom and relaxing your legs. These body-centered action
rlm@441 98 descriptions can be either learned or hard coded.
rlm@447 99 - Alignment (Posture imitation) :: When trying to interpret a video
rlm@447 100 or image, the creature takes a model of itself and aligns it
rlm@447 101 with whatever it sees. This alignment can even cross species,
rlm@447 102 as when humans try to align themselves with things like
rlm@447 103 ponies, dogs, or other humans with a different body type.
rlm@447 104 - Empathy (Sensory extrapolation) :: The alignment triggers
rlm@447 105 associations with sensory data from prior experiences. For
rlm@447 106 example, the alignment itself easily maps to proprioceptive
rlm@447 107 data. Any sounds or obvious skin contact in the video can to a
rlm@447 108 lesser extent trigger previous experience. Segments of
rlm@447 109 previous experiences are stitched together to form a coherent
rlm@447 110 and complete sensory portrait of the scene.
rlm@447 111 - Recognition (Classification) :: With the scene described in terms
rlm@447 112 of first person sensory events, the creature can now run its
rlm@447 113 action-identification programs on this synthesized sensory
rlm@447 114 data, just as it would if it were actually experiencing the
rlm@447 115 scene first-hand. If previous experience has been accurately
rlm@447 116 retrieved, and if it is analogous enough to the scene, then
rlm@447 117 the creature will correctly identify the action in the scene.
rlm@447 118
rlm@441 119 For example, I think humans are able to label the cat video as
rlm@447 120 ``drinking'' because they imagine /themselves/ as the cat, and
rlm@441 121 imagine putting their face up against a stream of water and
rlm@441 122 sticking out their tongue. In that imagined world, they can feel
rlm@441 123 the cool water hitting their tongue, and feel the water entering
rlm@447 124 their body, and are able to recognize that /feeling/ as drinking.
rlm@447 125 So, the label of the action is not really in the pixels of the
rlm@447 126 image, but is found clearly in a simulation inspired by those
rlm@447 127 pixels. An imaginative system, having been trained on drinking and
rlm@447 128 non-drinking examples and learning that the most important
rlm@447 129 component of drinking is the feeling of water sliding down one's
rlm@447 130 throat, would analyze a video of a cat drinking in the following
rlm@447 131 manner:
rlm@441 132
rlm@447 133 1. Create a physical model of the video by putting a ``fuzzy''
rlm@447 134 model of its own body in place of the cat. Possibly also create
rlm@447 135 a simulation of the stream of water.
rlm@441 136
rlm@441 137 2. Play out this simulated scene and generate imagined sensory
rlm@441 138 experience. This will include relevant muscle contractions, a
rlm@441 139 close up view of the stream from the cat's perspective, and most
rlm@441 140 importantly, the imagined feeling of water entering the
rlm@443 141 mouth. The imagined sensory experience can come from a
rlm@441 142 simulation of the event, but can also be pattern-matched from
rlm@441 143 previous, similar embodied experience.
rlm@441 144
rlm@441 145 3. The action is now easily identified as drinking by the sense of
rlm@441 146 taste alone. The other senses (such as the tongue moving in and
rlm@441 147 out) help to give plausibility to the simulated action. Note that
rlm@441 148 the sense of vision, while critical in creating the simulation,
rlm@441 149 is not critical for identifying the action from the simulation.
rlm@441 150
rlm@441 151 For the chair examples, the process is even easier:
rlm@441 152
rlm@441 153 1. Align a model of your body to the person in the image.
rlm@441 154
rlm@441 155 2. Generate proprioceptive sensory data from this alignment.
rlm@437 156
rlm@441 157 3. Use the imagined proprioceptive data as a key to lookup related
rlm@441 158 sensory experience associated with that particular proproceptive
rlm@441 159 feeling.
rlm@437 160
rlm@443 161 4. Retrieve the feeling of your bottom resting on a surface, your
rlm@443 162 knees bent, and your leg muscles relaxed.
rlm@437 163
rlm@441 164 5. This sensory information is consistent with the =sitting?=
rlm@441 165 sensory predicate, so you (and the entity in the image) must be
rlm@441 166 sitting.
rlm@440 167
rlm@441 168 6. There must be a chair-like object since you are sitting.
rlm@440 169
rlm@441 170 Empathy offers yet another alternative to the age-old AI
rlm@441 171 representation question: ``What is a chair?'' --- A chair is the
rlm@441 172 feeling of sitting.
rlm@441 173
rlm@441 174 My program, =EMPATH= uses this empathic problem solving technique
rlm@441 175 to interpret the actions of a simple, worm-like creature.
rlm@437 176
rlm@441 177 #+caption: The worm performs many actions during free play such as
rlm@441 178 #+caption: curling, wiggling, and resting.
rlm@441 179 #+name: worm-intro
rlm@446 180 #+ATTR_LaTeX: :width 15cm
rlm@445 181 [[./images/worm-intro-white.png]]
rlm@437 182
rlm@447 183 #+caption: =EMPATH= recognized and classified each of these poses by
rlm@447 184 #+caption: inferring the complete sensory experience from
rlm@447 185 #+caption: proprioceptive data.
rlm@441 186 #+name: worm-recognition-intro
rlm@446 187 #+ATTR_LaTeX: :width 15cm
rlm@445 188 [[./images/worm-poses.png]]
rlm@441 189
rlm@441 190 One powerful advantage of empathic problem solving is that it
rlm@441 191 factors the action recognition problem into two easier problems. To
rlm@441 192 use empathy, you need an /aligner/, which takes the video and a
rlm@441 193 model of your body, and aligns the model with the video. Then, you
rlm@441 194 need a /recognizer/, which uses the aligned model to interpret the
rlm@441 195 action. The power in this method lies in the fact that you describe
rlm@447 196 all actions form a body-centered, viewpoint You are less tied to
rlm@447 197 the particulars of any visual representation of the actions. If you
rlm@441 198 teach the system what ``running'' is, and you have a good enough
rlm@441 199 aligner, the system will from then on be able to recognize running
rlm@441 200 from any point of view, even strange points of view like above or
rlm@441 201 underneath the runner. This is in contrast to action recognition
rlm@441 202 schemes that try to identify actions using a non-embodied approach
rlm@447 203 such as TODO:REFERENCE. If these systems learn about running as
rlm@447 204 viewed from the side, they will not automatically be able to
rlm@447 205 recognize running from any other viewpoint.
rlm@441 206
rlm@441 207 Another powerful advantage is that using the language of multiple
rlm@441 208 body-centered rich senses to describe body-centerd actions offers a
rlm@441 209 massive boost in descriptive capability. Consider how difficult it
rlm@441 210 would be to compose a set of HOG filters to describe the action of
rlm@447 211 a simple worm-creature ``curling'' so that its head touches its
rlm@447 212 tail, and then behold the simplicity of describing thus action in a
rlm@441 213 language designed for the task (listing \ref{grand-circle-intro}):
rlm@441 214
rlm@446 215 #+caption: Body-centerd actions are best expressed in a body-centered
rlm@446 216 #+caption: language. This code detects when the worm has curled into a
rlm@446 217 #+caption: full circle. Imagine how you would replicate this functionality
rlm@446 218 #+caption: using low-level pixel features such as HOG filters!
rlm@446 219 #+name: grand-circle-intro
rlm@446 220 #+begin_listing clojure
rlm@446 221 #+begin_src clojure
rlm@446 222 (defn grand-circle?
rlm@446 223 "Does the worm form a majestic circle (one end touching the other)?"
rlm@446 224 [experiences]
rlm@446 225 (and (curled? experiences)
rlm@446 226 (let [worm-touch (:touch (peek experiences))
rlm@446 227 tail-touch (worm-touch 0)
rlm@446 228 head-touch (worm-touch 4)]
rlm@446 229 (and (< 0.55 (contact worm-segment-bottom-tip tail-touch))
rlm@446 230 (< 0.55 (contact worm-segment-top-tip head-touch))))))
rlm@446 231 #+end_src
rlm@446 232 #+end_listing
rlm@446 233
rlm@435 234
rlm@437 235 ** =CORTEX= is a toolkit for building sensate creatures
rlm@435 236
rlm@436 237 Hand integration demo
rlm@435 238
rlm@437 239 ** Contributions
rlm@435 240
rlm@436 241 * Building =CORTEX=
rlm@435 242
rlm@436 243 ** To explore embodiment, we need a world, body, and senses
rlm@435 244
rlm@436 245 ** Because of Time, simulation is perferable to reality
rlm@435 246
rlm@436 247 ** Video game engines are a great starting point
rlm@435 248
rlm@436 249 ** Bodies are composed of segments connected by joints
rlm@435 250
rlm@436 251 ** Eyes reuse standard video game components
rlm@436 252
rlm@436 253 ** Hearing is hard; =CORTEX= does it right
rlm@436 254
rlm@436 255 ** Touch uses hundreds of hair-like elements
rlm@436 256
rlm@440 257 ** Proprioception is the sense that makes everything ``real''
rlm@436 258
rlm@436 259 ** Muscles are both effectors and sensors
rlm@436 260
rlm@436 261 ** =CORTEX= brings complex creatures to life!
rlm@436 262
rlm@436 263 ** =CORTEX= enables many possiblities for further research
rlm@435 264
rlm@435 265 * Empathy in a simulated worm
rlm@435 266
rlm@436 267 ** Embodiment factors action recognition into managable parts
rlm@435 268
rlm@436 269 ** Action recognition is easy with a full gamut of senses
rlm@435 270
rlm@437 271 ** Digression: bootstrapping touch using free exploration
rlm@435 272
rlm@436 273 ** \Phi-space describes the worm's experiences
rlm@435 274
rlm@436 275 ** Empathy is the process of tracing though \Phi-space
rlm@435 276
rlm@441 277 ** Efficient action recognition with =EMPATH=
rlm@425 278
rlm@432 279 * Contributions
rlm@432 280 - Built =CORTEX=, a comprehensive platform for embodied AI
rlm@432 281 experiments. Has many new features lacking in other systems, such
rlm@432 282 as sound. Easy to model/create new creatures.
rlm@432 283 - created a novel concept for action recognition by using artificial
rlm@432 284 imagination.
rlm@426 285
rlm@436 286 In the second half of the thesis I develop a computational model of
rlm@436 287 empathy, using =CORTEX= as a base. Empathy in this context is the
rlm@436 288 ability to observe another creature and infer what sorts of sensations
rlm@436 289 that creature is feeling. My empathy algorithm involves multiple
rlm@436 290 phases. First is free-play, where the creature moves around and gains
rlm@436 291 sensory experience. From this experience I construct a representation
rlm@447 292 of the creature's sensory state space, which I call \Phi-space. Using
rlm@447 293 \Phi-space, I construct an efficient function for enriching the
rlm@436 294 limited data that comes from observing another creature with a full
rlm@436 295 compliment of imagined sensory data based on previous experience. I
rlm@436 296 can then use the imagined sensory data to recognize what the observed
rlm@436 297 creature is doing and feeling, using straightforward embodied action
rlm@436 298 predicates. This is all demonstrated with using a simple worm-like
rlm@436 299 creature, and recognizing worm-actions based on limited data.
rlm@432 300
rlm@436 301 Embodied representation using multiple senses such as touch,
rlm@436 302 proprioception, and muscle tension turns out be be exceedingly
rlm@436 303 efficient at describing body-centered actions. It is the ``right
rlm@436 304 language for the job''. For example, it takes only around 5 lines of
rlm@436 305 LISP code to describe the action of ``curling'' using embodied
rlm@436 306 primitives. It takes about 8 lines to describe the seemingly
rlm@436 307 complicated action of wiggling.
rlm@432 308
rlm@437 309
rlm@437 310
rlm@437 311 * COMMENT names for cortex
rlm@447 312 - bioland
rlm@447 313
rlm@447 314
rlm@447 315
rlm@447 316
rlm@447 317 # An anatomical joke:
rlm@447 318 # - Training
rlm@447 319 # - Skeletal imitation
rlm@447 320 # - Sensory fleshing-out
rlm@447 321 # - Classification