Mercurial > cortex

annotate thesis/cortex.org @ 465:e4104ce9105c

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
working on body/joints.
author Robert McIntyre <rlm@mit.edu>
date Fri, 28 Mar 2014 11:08:32 -0400
parents 8bf4bb02ed05
children da311eefbb09
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@451 6 #+LaTeX_CLASS_OPTIONS: [nofloat]
rlm@422 7
rlm@465 8 * COMMENT templates
rlm@465 9 #+caption:
rlm@465 10 #+caption:
rlm@465 11 #+caption:
rlm@465 12 #+caption:
rlm@465 13 #+name: name
rlm@465 14 #+begin_listing clojure
rlm@465 15 #+begin_src clojure
rlm@465 16 #+end_src
rlm@465 17 #+end_listing
rlm@465 18
rlm@465 19 #+caption:
rlm@465 20 #+caption:
rlm@465 21 #+caption:
rlm@465 22 #+name: name
rlm@465 23 #+ATTR_LaTeX: :width 10cm
rlm@465 24 [[./images/Eve.jpg]]
rlm@465 25
rlm@465 26
rlm@465 27
rlm@465 28 * COMMENT Empathy and Embodiment as problem solving strategies
rlm@437 29
rlm@437 30 By the end of this thesis, you will have seen a novel approach to
rlm@437 31 interpreting video using embodiment and empathy. You will have also
rlm@437 32 seen one way to efficiently implement empathy for embodied
rlm@447 33 creatures. Finally, you will become familiar with =CORTEX=, a system
rlm@447 34 for designing and simulating creatures with rich senses, which you
rlm@447 35 may choose to use in your own research.
rlm@437 36
rlm@441 37 This is the core vision of my thesis: That one of the important ways
rlm@441 38 in which we understand others is by imagining ourselves in their
rlm@441 39 position and emphatically feeling experiences relative to our own
rlm@441 40 bodies. By understanding events in terms of our own previous
rlm@441 41 corporeal experience, we greatly constrain the possibilities of what
rlm@441 42 would otherwise be an unwieldy exponential search. This extra
rlm@441 43 constraint can be the difference between easily understanding what
rlm@441 44 is happening in a video and being completely lost in a sea of
rlm@441 45 incomprehensible color and movement.
rlm@435 46
rlm@436 47 ** Recognizing actions in video is extremely difficult
rlm@437 48
rlm@447 49 Consider for example the problem of determining what is happening
rlm@447 50 in a video of which this is one frame:
rlm@437 51
rlm@441 52 #+caption: A cat drinking some water. Identifying this action is
rlm@441 53 #+caption: beyond the state of the art for computers.
rlm@441 54 #+ATTR_LaTeX: :width 7cm
rlm@441 55 [[./images/cat-drinking.jpg]]
rlm@441 56
rlm@441 57 It is currently impossible for any computer program to reliably
rlm@447 58 label such a video as ``drinking''. And rightly so -- it is a very
rlm@441 59 hard problem! What features can you describe in terms of low level
rlm@441 60 functions of pixels that can even begin to describe at a high level
rlm@441 61 what is happening here?
rlm@437 62
rlm@447 63 Or suppose that you are building a program that recognizes chairs.
rlm@448 64 How could you ``see'' the chair in figure \ref{hidden-chair}?
rlm@441 65
rlm@441 66 #+caption: The chair in this image is quite obvious to humans, but I
rlm@448 67 #+caption: doubt that any modern computer vision program can find it.
rlm@441 68 #+name: hidden-chair
rlm@441 69 #+ATTR_LaTeX: :width 10cm
rlm@441 70 [[./images/fat-person-sitting-at-desk.jpg]]
rlm@441 71
rlm@441 72 Finally, how is it that you can easily tell the difference between
rlm@441 73 how the girls /muscles/ are working in figure \ref{girl}?
rlm@441 74
rlm@441 75 #+caption: The mysterious ``common sense'' appears here as you are able
rlm@441 76 #+caption: to discern the difference in how the girl's arm muscles
rlm@441 77 #+caption: are activated between the two images.
rlm@441 78 #+name: girl
rlm@448 79 #+ATTR_LaTeX: :width 7cm
rlm@441 80 [[./images/wall-push.png]]
rlm@437 81
rlm@441 82 Each of these examples tells us something about what might be going
rlm@441 83 on in our minds as we easily solve these recognition problems.
rlm@441 84
rlm@441 85 The hidden chairs show us that we are strongly triggered by cues
rlm@447 86 relating to the position of human bodies, and that we can determine
rlm@447 87 the overall physical configuration of a human body even if much of
rlm@447 88 that body is occluded.
rlm@437 89
rlm@441 90 The picture of the girl pushing against the wall tells us that we
rlm@441 91 have common sense knowledge about the kinetics of our own bodies.
rlm@441 92 We know well how our muscles would have to work to maintain us in
rlm@441 93 most positions, and we can easily project this self-knowledge to
rlm@441 94 imagined positions triggered by images of the human body.
rlm@441 95
rlm@441 96 ** =EMPATH= neatly solves recognition problems
rlm@441 97
rlm@441 98 I propose a system that can express the types of recognition
rlm@441 99 problems above in a form amenable to computation. It is split into
rlm@441 100 four parts:
rlm@441 101
rlm@448 102 - Free/Guided Play :: The creature moves around and experiences the
rlm@448 103 world through its unique perspective. Many otherwise
rlm@448 104 complicated actions are easily described in the language of a
rlm@448 105 full suite of body-centered, rich senses. For example,
rlm@448 106 drinking is the feeling of water sliding down your throat, and
rlm@448 107 cooling your insides. It's often accompanied by bringing your
rlm@448 108 hand close to your face, or bringing your face close to water.
rlm@448 109 Sitting down is the feeling of bending your knees, activating
rlm@448 110 your quadriceps, then feeling a surface with your bottom and
rlm@448 111 relaxing your legs. These body-centered action descriptions
rlm@448 112 can be either learned or hard coded.
rlm@448 113 - Posture Imitation :: When trying to interpret a video or image,
rlm@448 114 the creature takes a model of itself and aligns it with
rlm@448 115 whatever it sees. This alignment can even cross species, as
rlm@448 116 when humans try to align themselves with things like ponies,
rlm@448 117 dogs, or other humans with a different body type.
rlm@448 118 - Empathy :: The alignment triggers associations with
rlm@448 119 sensory data from prior experiences. For example, the
rlm@448 120 alignment itself easily maps to proprioceptive data. Any
rlm@448 121 sounds or obvious skin contact in the video can to a lesser
rlm@448 122 extent trigger previous experience. Segments of previous
rlm@448 123 experiences are stitched together to form a coherent and
rlm@448 124 complete sensory portrait of the scene.
rlm@448 125 - Recognition :: With the scene described in terms of first
rlm@448 126 person sensory events, the creature can now run its
rlm@447 127 action-identification programs on this synthesized sensory
rlm@447 128 data, just as it would if it were actually experiencing the
rlm@447 129 scene first-hand. If previous experience has been accurately
rlm@447 130 retrieved, and if it is analogous enough to the scene, then
rlm@447 131 the creature will correctly identify the action in the scene.
rlm@447 132
rlm@441 133 For example, I think humans are able to label the cat video as
rlm@447 134 ``drinking'' because they imagine /themselves/ as the cat, and
rlm@441 135 imagine putting their face up against a stream of water and
rlm@441 136 sticking out their tongue. In that imagined world, they can feel
rlm@441 137 the cool water hitting their tongue, and feel the water entering
rlm@447 138 their body, and are able to recognize that /feeling/ as drinking.
rlm@447 139 So, the label of the action is not really in the pixels of the
rlm@447 140 image, but is found clearly in a simulation inspired by those
rlm@447 141 pixels. An imaginative system, having been trained on drinking and
rlm@447 142 non-drinking examples and learning that the most important
rlm@447 143 component of drinking is the feeling of water sliding down one's
rlm@447 144 throat, would analyze a video of a cat drinking in the following
rlm@447 145 manner:
rlm@441 146
rlm@447 147 1. Create a physical model of the video by putting a ``fuzzy''
rlm@447 148 model of its own body in place of the cat. Possibly also create
rlm@447 149 a simulation of the stream of water.
rlm@441 150
rlm@441 151 2. Play out this simulated scene and generate imagined sensory
rlm@441 152 experience. This will include relevant muscle contractions, a
rlm@441 153 close up view of the stream from the cat's perspective, and most
rlm@441 154 importantly, the imagined feeling of water entering the
rlm@443 155 mouth. The imagined sensory experience can come from a
rlm@441 156 simulation of the event, but can also be pattern-matched from
rlm@441 157 previous, similar embodied experience.
rlm@441 158
rlm@441 159 3. The action is now easily identified as drinking by the sense of
rlm@441 160 taste alone. The other senses (such as the tongue moving in and
rlm@441 161 out) help to give plausibility to the simulated action. Note that
rlm@441 162 the sense of vision, while critical in creating the simulation,
rlm@441 163 is not critical for identifying the action from the simulation.
rlm@441 164
rlm@441 165 For the chair examples, the process is even easier:
rlm@441 166
rlm@441 167 1. Align a model of your body to the person in the image.
rlm@441 168
rlm@441 169 2. Generate proprioceptive sensory data from this alignment.
rlm@437 170
rlm@441 171 3. Use the imagined proprioceptive data as a key to lookup related
rlm@441 172 sensory experience associated with that particular proproceptive
rlm@441 173 feeling.
rlm@437 174
rlm@443 175 4. Retrieve the feeling of your bottom resting on a surface, your
rlm@443 176 knees bent, and your leg muscles relaxed.
rlm@437 177
rlm@441 178 5. This sensory information is consistent with the =sitting?=
rlm@441 179 sensory predicate, so you (and the entity in the image) must be
rlm@441 180 sitting.
rlm@440 181
rlm@441 182 6. There must be a chair-like object since you are sitting.
rlm@440 183
rlm@441 184 Empathy offers yet another alternative to the age-old AI
rlm@441 185 representation question: ``What is a chair?'' --- A chair is the
rlm@441 186 feeling of sitting.
rlm@441 187
rlm@441 188 My program, =EMPATH= uses this empathic problem solving technique
rlm@441 189 to interpret the actions of a simple, worm-like creature.
rlm@437 190
rlm@441 191 #+caption: The worm performs many actions during free play such as
rlm@441 192 #+caption: curling, wiggling, and resting.
rlm@441 193 #+name: worm-intro
rlm@446 194 #+ATTR_LaTeX: :width 15cm
rlm@445 195 [[./images/worm-intro-white.png]]
rlm@437 196
rlm@462 197 #+caption: =EMPATH= recognized and classified each of these
rlm@462 198 #+caption: poses by inferring the complete sensory experience
rlm@462 199 #+caption: from proprioceptive data.
rlm@441 200 #+name: worm-recognition-intro
rlm@446 201 #+ATTR_LaTeX: :width 15cm
rlm@445 202 [[./images/worm-poses.png]]
rlm@441 203
rlm@441 204 One powerful advantage of empathic problem solving is that it
rlm@441 205 factors the action recognition problem into two easier problems. To
rlm@441 206 use empathy, you need an /aligner/, which takes the video and a
rlm@441 207 model of your body, and aligns the model with the video. Then, you
rlm@441 208 need a /recognizer/, which uses the aligned model to interpret the
rlm@441 209 action. The power in this method lies in the fact that you describe
rlm@448 210 all actions form a body-centered viewpoint. You are less tied to
rlm@447 211 the particulars of any visual representation of the actions. If you
rlm@441 212 teach the system what ``running'' is, and you have a good enough
rlm@441 213 aligner, the system will from then on be able to recognize running
rlm@441 214 from any point of view, even strange points of view like above or
rlm@441 215 underneath the runner. This is in contrast to action recognition
rlm@448 216 schemes that try to identify actions using a non-embodied approach.
rlm@448 217 If these systems learn about running as viewed from the side, they
rlm@448 218 will not automatically be able to recognize running from any other
rlm@448 219 viewpoint.
rlm@441 220
rlm@441 221 Another powerful advantage is that using the language of multiple
rlm@441 222 body-centered rich senses to describe body-centerd actions offers a
rlm@441 223 massive boost in descriptive capability. Consider how difficult it
rlm@441 224 would be to compose a set of HOG filters to describe the action of
rlm@447 225 a simple worm-creature ``curling'' so that its head touches its
rlm@447 226 tail, and then behold the simplicity of describing thus action in a
rlm@441 227 language designed for the task (listing \ref{grand-circle-intro}):
rlm@441 228
rlm@446 229 #+caption: Body-centerd actions are best expressed in a body-centered
rlm@446 230 #+caption: language. This code detects when the worm has curled into a
rlm@446 231 #+caption: full circle. Imagine how you would replicate this functionality
rlm@446 232 #+caption: using low-level pixel features such as HOG filters!
rlm@446 233 #+name: grand-circle-intro
rlm@452 234 #+attr_latex: [htpb]
rlm@452 235 #+begin_listing clojure
rlm@446 236 #+begin_src clojure
rlm@446 237 (defn grand-circle?
rlm@446 238 "Does the worm form a majestic circle (one end touching the other)?"
rlm@446 239 [experiences]
rlm@446 240 (and (curled? experiences)
rlm@446 241 (let [worm-touch (:touch (peek experiences))
rlm@446 242 tail-touch (worm-touch 0)
rlm@446 243 head-touch (worm-touch 4)]
rlm@462 244 (and (< 0.2 (contact worm-segment-bottom-tip tail-touch))
rlm@462 245 (< 0.2 (contact worm-segment-top-tip head-touch))))))
rlm@446 246 #+end_src
rlm@446 247 #+end_listing
rlm@446 248
rlm@435 249
rlm@449 250 ** =CORTEX= is a toolkit for building sensate creatures
rlm@435 251
rlm@448 252 I built =CORTEX= to be a general AI research platform for doing
rlm@448 253 experiments involving multiple rich senses and a wide variety and
rlm@448 254 number of creatures. I intend it to be useful as a library for many
rlm@462 255 more projects than just this thesis. =CORTEX= was necessary to meet
rlm@462 256 a need among AI researchers at CSAIL and beyond, which is that
rlm@462 257 people often will invent neat ideas that are best expressed in the
rlm@448 258 language of creatures and senses, but in order to explore those
rlm@448 259 ideas they must first build a platform in which they can create
rlm@448 260 simulated creatures with rich senses! There are many ideas that
rlm@448 261 would be simple to execute (such as =EMPATH=), but attached to them
rlm@448 262 is the multi-month effort to make a good creature simulator. Often,
rlm@448 263 that initial investment of time proves to be too much, and the
rlm@448 264 project must make do with a lesser environment.
rlm@435 265
rlm@448 266 =CORTEX= is well suited as an environment for embodied AI research
rlm@448 267 for three reasons:
rlm@448 268
rlm@448 269 - You can create new creatures using Blender, a popular 3D modeling
rlm@448 270 program. Each sense can be specified using special blender nodes
rlm@448 271 with biologically inspired paramaters. You need not write any
rlm@448 272 code to create a creature, and can use a wide library of
rlm@448 273 pre-existing blender models as a base for your own creatures.
rlm@448 274
rlm@448 275 - =CORTEX= implements a wide variety of senses, including touch,
rlm@448 276 proprioception, vision, hearing, and muscle tension. Complicated
rlm@448 277 senses like touch, and vision involve multiple sensory elements
rlm@448 278 embedded in a 2D surface. You have complete control over the
rlm@448 279 distribution of these sensor elements through the use of simple
rlm@448 280 png image files. In particular, =CORTEX= implements more
rlm@448 281 comprehensive hearing than any other creature simulation system
rlm@448 282 available.
rlm@448 283
rlm@448 284 - =CORTEX= supports any number of creatures and any number of
rlm@448 285 senses. Time in =CORTEX= dialates so that the simulated creatures
rlm@448 286 always precieve a perfectly smooth flow of time, regardless of
rlm@448 287 the actual computational load.
rlm@448 288
rlm@448 289 =CORTEX= is built on top of =jMonkeyEngine3=, which is a video game
rlm@448 290 engine designed to create cross-platform 3D desktop games. =CORTEX=
rlm@448 291 is mainly written in clojure, a dialect of =LISP= that runs on the
rlm@448 292 java virtual machine (JVM). The API for creating and simulating
rlm@449 293 creatures and senses is entirely expressed in clojure, though many
rlm@449 294 senses are implemented at the layer of jMonkeyEngine or below. For
rlm@449 295 example, for the sense of hearing I use a layer of clojure code on
rlm@449 296 top of a layer of java JNI bindings that drive a layer of =C++=
rlm@449 297 code which implements a modified version of =OpenAL= to support
rlm@449 298 multiple listeners. =CORTEX= is the only simulation environment
rlm@449 299 that I know of that can support multiple entities that can each
rlm@449 300 hear the world from their own perspective. Other senses also
rlm@449 301 require a small layer of Java code. =CORTEX= also uses =bullet=, a
rlm@449 302 physics simulator written in =C=.
rlm@448 303
rlm@448 304 #+caption: Here is the worm from above modeled in Blender, a free
rlm@448 305 #+caption: 3D-modeling program. Senses and joints are described
rlm@448 306 #+caption: using special nodes in Blender.
rlm@448 307 #+name: worm-recognition-intro
rlm@448 308 #+ATTR_LaTeX: :width 12cm
rlm@448 309 [[./images/blender-worm.png]]
rlm@448 310
rlm@449 311 Here are some thing I anticipate that =CORTEX= might be used for:
rlm@449 312
rlm@449 313 - exploring new ideas about sensory integration
rlm@449 314 - distributed communication among swarm creatures
rlm@449 315 - self-learning using free exploration,
rlm@449 316 - evolutionary algorithms involving creature construction
rlm@449 317 - exploration of exoitic senses and effectors that are not possible
rlm@449 318 in the real world (such as telekenisis or a semantic sense)
rlm@449 319 - imagination using subworlds
rlm@449 320
rlm@451 321 During one test with =CORTEX=, I created 3,000 creatures each with
rlm@448 322 their own independent senses and ran them all at only 1/80 real
rlm@448 323 time. In another test, I created a detailed model of my own hand,
rlm@448 324 equipped with a realistic distribution of touch (more sensitive at
rlm@448 325 the fingertips), as well as eyes and ears, and it ran at around 1/4
rlm@451 326 real time.
rlm@448 327
rlm@451 328 #+BEGIN_LaTeX
rlm@449 329 \begin{sidewaysfigure}
rlm@449 330 \includegraphics[width=9.5in]{images/full-hand.png}
rlm@451 331 \caption{
rlm@451 332 I modeled my own right hand in Blender and rigged it with all the
rlm@451 333 senses that {\tt CORTEX} supports. My simulated hand has a
rlm@451 334 biologically inspired distribution of touch sensors. The senses are
rlm@451 335 displayed on the right, and the simulation is displayed on the
rlm@451 336 left. Notice that my hand is curling its fingers, that it can see
rlm@451 337 its own finger from the eye in its palm, and that it can feel its
rlm@451 338 own thumb touching its palm.}
rlm@449 339 \end{sidewaysfigure}
rlm@451 340 #+END_LaTeX
rlm@448 341
rlm@437 342 ** Contributions
rlm@435 343
rlm@451 344 - I built =CORTEX=, a comprehensive platform for embodied AI
rlm@451 345 experiments. =CORTEX= supports many features lacking in other
rlm@451 346 systems, such proper simulation of hearing. It is easy to create
rlm@451 347 new =CORTEX= creatures using Blender, a free 3D modeling program.
rlm@449 348
rlm@451 349 - I built =EMPATH=, which uses =CORTEX= to identify the actions of
rlm@451 350 a worm-like creature using a computational model of empathy.
rlm@449 351
rlm@436 352 * Building =CORTEX=
rlm@435 353
rlm@462 354 I intend for =CORTEX= to be used as a general purpose library for
rlm@462 355 building creatures and outfitting them with senses, so that it will
rlm@462 356 be useful for other researchers who want to test out ideas of their
rlm@462 357 own. To this end, wherver I have had to make archetictural choices
rlm@462 358 about =CORTEX=, I have chosen to give as much freedom to the user as
rlm@462 359 possible, so that =CORTEX= may be used for things I have not
rlm@462 360 forseen.
rlm@462 361
rlm@465 362 ** COMMENT Simulation or Reality?
rlm@462 363
rlm@462 364 The most important archetictural decision of all is the choice to
rlm@462 365 use a computer-simulated environemnt in the first place! The world
rlm@462 366 is a vast and rich place, and for now simulations are a very poor
rlm@462 367 reflection of its complexity. It may be that there is a significant
rlm@462 368 qualatative difference between dealing with senses in the real
rlm@462 369 world and dealing with pale facilimilies of them in a
rlm@462 370 simulation. What are the advantages and disadvantages of a
rlm@462 371 simulation vs. reality?
rlm@462 372
rlm@462 373 *** Simulation
rlm@462 374
rlm@462 375 The advantages of virtual reality are that when everything is a
rlm@462 376 simulation, experiments in that simulation are absolutely
rlm@462 377 reproducible. It's also easier to change the character and world
rlm@462 378 to explore new situations and different sensory combinations.
rlm@462 379
rlm@462 380 If the world is to be simulated on a computer, then not only do
rlm@462 381 you have to worry about whether the character's senses are rich
rlm@462 382 enough to learn from the world, but whether the world itself is
rlm@462 383 rendered with enough detail and realism to give enough working
rlm@462 384 material to the character's senses. To name just a few
rlm@462 385 difficulties facing modern physics simulators: destructibility of
rlm@462 386 the environment, simulation of water/other fluids, large areas,
rlm@462 387 nonrigid bodies, lots of objects, smoke. I don't know of any
rlm@462 388 computer simulation that would allow a character to take a rock
rlm@462 389 and grind it into fine dust, then use that dust to make a clay
rlm@462 390 sculpture, at least not without spending years calculating the
rlm@462 391 interactions of every single small grain of dust. Maybe a
rlm@462 392 simulated world with today's limitations doesn't provide enough
rlm@462 393 richness for real intelligence to evolve.
rlm@462 394
rlm@462 395 *** Reality
rlm@462 396
rlm@462 397 The other approach for playing with senses is to hook your
rlm@462 398 software up to real cameras, microphones, robots, etc., and let it
rlm@462 399 loose in the real world. This has the advantage of eliminating
rlm@462 400 concerns about simulating the world at the expense of increasing
rlm@462 401 the complexity of implementing the senses. Instead of just
rlm@462 402 grabbing the current rendered frame for processing, you have to
rlm@462 403 use an actual camera with real lenses and interact with photons to
rlm@462 404 get an image. It is much harder to change the character, which is
rlm@462 405 now partly a physical robot of some sort, since doing so involves
rlm@462 406 changing things around in the real world instead of modifying
rlm@462 407 lines of code. While the real world is very rich and definitely
rlm@462 408 provides enough stimulation for intelligence to develop as
rlm@462 409 evidenced by our own existence, it is also uncontrollable in the
rlm@462 410 sense that a particular situation cannot be recreated perfectly or
rlm@462 411 saved for later use. It is harder to conduct science because it is
rlm@462 412 harder to repeat an experiment. The worst thing about using the
rlm@462 413 real world instead of a simulation is the matter of time. Instead
rlm@462 414 of simulated time you get the constant and unstoppable flow of
rlm@462 415 real time. This severely limits the sorts of software you can use
rlm@462 416 to program the AI because all sense inputs must be handled in real
rlm@462 417 time. Complicated ideas may have to be implemented in hardware or
rlm@462 418 may simply be impossible given the current speed of our
rlm@462 419 processors. Contrast this with a simulation, in which the flow of
rlm@462 420 time in the simulated world can be slowed down to accommodate the
rlm@462 421 limitations of the character's programming. In terms of cost,
rlm@462 422 doing everything in software is far cheaper than building custom
rlm@462 423 real-time hardware. All you need is a laptop and some patience.
rlm@435 424
rlm@465 425 ** COMMENT Because of Time, simulation is perferable to reality
rlm@435 426
rlm@462 427 I envision =CORTEX= being used to support rapid prototyping and
rlm@462 428 iteration of ideas. Even if I could put together a well constructed
rlm@462 429 kit for creating robots, it would still not be enough because of
rlm@462 430 the scourge of real-time processing. Anyone who wants to test their
rlm@462 431 ideas in the real world must always worry about getting their
rlm@465 432 algorithms to run fast enough to process information in real time.
rlm@465 433 The need for real time processing only increases if multiple senses
rlm@465 434 are involved. In the extreme case, even simple algorithms will have
rlm@465 435 to be accelerated by ASIC chips or FPGAs, turning what would
rlm@465 436 otherwise be a few lines of code and a 10x speed penality into a
rlm@465 437 multi-month ordeal. For this reason, =CORTEX= supports
rlm@462 438 /time-dialiation/, which scales back the framerate of the
rlm@465 439 simulation in proportion to the amount of processing each frame.
rlm@465 440 From the perspective of the creatures inside the simulation, time
rlm@465 441 always appears to flow at a constant rate, regardless of how
rlm@462 442 complicated the envorimnent becomes or how many creatures are in
rlm@462 443 the simulation. The cost is that =CORTEX= can sometimes run slower
rlm@462 444 than real time. This can also be an advantage, however ---
rlm@462 445 simulations of very simple creatures in =CORTEX= generally run at
rlm@462 446 40x on my machine!
rlm@462 447
rlm@465 448 ** COMMENT Video game engines are a great starting point
rlm@462 449
rlm@462 450 I did not need to write my own physics simulation code or shader to
rlm@462 451 build =CORTEX=. Doing so would lead to a system that is impossible
rlm@462 452 for anyone but myself to use anyway. Instead, I use a video game
rlm@462 453 engine as a base and modify it to accomodate the additional needs
rlm@462 454 of =CORTEX=. Video game engines are an ideal starting point to
rlm@462 455 build =CORTEX=, because they are not far from being creature
rlm@463 456 building systems themselves.
rlm@462 457
rlm@462 458 First off, general purpose video game engines come with a physics
rlm@462 459 engine and lighting / sound system. The physics system provides
rlm@462 460 tools that can be co-opted to serve as touch, proprioception, and
rlm@462 461 muscles. Since some games support split screen views, a good video
rlm@462 462 game engine will allow you to efficiently create multiple cameras
rlm@463 463 in the simulated world that can be used as eyes. Video game systems
rlm@463 464 offer integrated asset management for things like textures and
rlm@463 465 creatures models, providing an avenue for defining creatures.
rlm@463 466 Finally, because video game engines support a large number of
rlm@463 467 users, if I don't stray too far from the base system, other
rlm@463 468 researchers can turn to this community for help when doing their
rlm@463 469 research.
rlm@463 470
rlm@465 471 ** COMMENT =CORTEX= is based on jMonkeyEngine3
rlm@463 472
rlm@463 473 While preparing to build =CORTEX= I studied several video game
rlm@463 474 engines to see which would best serve as a base. The top contenders
rlm@463 475 were:
rlm@463 476
rlm@463 477 - [[http://www.idsoftware.com][Quake II]]/[[http://www.bytonic.de/html/jake2.html][Jake2]] :: The Quake II engine was designed by ID
rlm@463 478 software in 1997. All the source code was released by ID
rlm@463 479 software into the Public Domain several years ago, and as a
rlm@463 480 result it has been ported to many different languages. This
rlm@463 481 engine was famous for its advanced use of realistic shading
rlm@463 482 and had decent and fast physics simulation. The main advantage
rlm@463 483 of the Quake II engine is its simplicity, but I ultimately
rlm@463 484 rejected it because the engine is too tied to the concept of a
rlm@463 485 first-person shooter game. One of the problems I had was that
rlm@463 486 there does not seem to be any easy way to attach multiple
rlm@463 487 cameras to a single character. There are also several physics
rlm@463 488 clipping issues that are corrected in a way that only applies
rlm@463 489 to the main character and do not apply to arbitrary objects.
rlm@463 490
rlm@463 491 - [[http://source.valvesoftware.com/][Source Engine]] :: The Source Engine evolved from the Quake II
rlm@463 492 and Quake I engines and is used by Valve in the Half-Life
rlm@463 493 series of games. The physics simulation in the Source Engine
rlm@463 494 is quite accurate and probably the best out of all the engines
rlm@463 495 I investigated. There is also an extensive community actively
rlm@463 496 working with the engine. However, applications that use the
rlm@463 497 Source Engine must be written in C++, the code is not open, it
rlm@463 498 only runs on Windows, and the tools that come with the SDK to
rlm@463 499 handle models and textures are complicated and awkward to use.
rlm@463 500
rlm@463 501 - [[http://jmonkeyengine.com/][jMonkeyEngine3]] :: jMonkeyEngine3 is a new library for creating
rlm@463 502 games in Java. It uses OpenGL to render to the screen and uses
rlm@463 503 screengraphs to avoid drawing things that do not appear on the
rlm@463 504 screen. It has an active community and several games in the
rlm@463 505 pipeline. The engine was not built to serve any particular
rlm@463 506 game but is instead meant to be used for any 3D game.
rlm@463 507
rlm@463 508 I chose jMonkeyEngine3 because it because it had the most features
rlm@464 509 out of all the free projects I looked at, and because I could then
rlm@463 510 write my code in clojure, an implementation of =LISP= that runs on
rlm@463 511 the JVM.
rlm@435 512
rlm@436 513 ** Bodies are composed of segments connected by joints
rlm@435 514
rlm@464 515 For the simple worm-like creatures I will use later on in this
rlm@464 516 thesis, I could define a simple API in =CORTEX= that would allow
rlm@464 517 one to create boxes, spheres, etc., and leave that API as the sole
rlm@464 518 way to create creatures. However, for =CORTEX= to truly be useful
rlm@464 519 for other projects, it needs to have a way to construct complicated
rlm@464 520 creatures. If possible, it would be nice to leverage work that has
rlm@464 521 already been done by the community of 3D modelers, or at least
rlm@464 522 enable people who are talented at moedling but not programming to
rlm@464 523 design =CORTEX= creatures.
rlm@464 524
rlm@464 525 Therefore, I use Blender, a free 3D modeling program, as the main
rlm@464 526 way to create creatures in =CORTEX=. However, the creatures modeled
rlm@464 527 in Blender must also be simple to simulate in jMonkeyEngine3's game
rlm@464 528 engine, and must also be easy to rig with =CORTEX='s senses.
rlm@464 529
rlm@464 530 While trying to find a good compromise for body-design, one option
rlm@464 531 I ultimately rejected is to use blender's [[http://wiki.blender.org/index.php/Doc:2.6/Manual/Rigging/Armatures][armature]] system. The idea
rlm@464 532 would have been to define a mesh which describes the creature's
rlm@464 533 entire body. To this you add an skeleton which deforms this
rlm@464 534 mesh. This technique is used extensively to model humans and create
rlm@464 535 realistic animations. It is hard to use for my purposes because it
rlm@464 536 is difficult to update the creature's Physics Collision Mesh in
rlm@464 537 tandem with its Geometric Mesh under the influence of the
rlm@464 538 armature. Without this the creature will not be able to grab things
rlm@464 539 in its environment, and it won't be able to tell where its physical
rlm@464 540 body is by using its eyes. Also, armatures do not specify any
rlm@464 541 rotational limits for a joint, making it hard to model elbows,
rlm@464 542 shoulders, etc.
rlm@464 543
rlm@464 544 Instead of using the human-like ``deformable bag of bones''
rlm@464 545 approach, I decided to base my body plans on multiple solid objects
rlm@464 546 that are connected by joints, inspired by the robot =EVE= from the
rlm@464 547 movie WALL-E.
rlm@464 548
rlm@464 549 #+caption: =EVE= from the movie WALL-E. This body plan turns
rlm@464 550 #+caption: out to be much better suited to my purposes than a more
rlm@464 551 #+caption: human-like one.
rlm@465 552 #+ATTR_LaTeX: :width 10cm
rlm@464 553 [[./images/Eve.jpg]]
rlm@464 554
rlm@464 555 =EVE='s body is composed of several rigid components that are held
rlm@464 556 together by invisible joint constraints. This is what I mean by
rlm@464 557 ``eve-like''. The main reason that I use eve-style bodies is for
rlm@464 558 efficiency, and so that there will be correspondence between the
rlm@464 559 AI's vision and the physical presence of its body. Each individual
rlm@464 560 section is simulated by a separate rigid body that corresponds
rlm@464 561 exactly with its visual representation and does not change.
rlm@464 562 Sections are connected by invisible joints that are well supported
rlm@464 563 in jMonkeyEngine3. Bullet, the physics backend for jMonkeyEngine3,
rlm@464 564 can efficiently simulate hundreds of rigid bodies connected by
rlm@464 565 joints. Sections do not have to stay as one piece forever; they can
rlm@464 566 be dynamically replaced with multiple sections to simulate
rlm@464 567 splitting in two. This could be used to simulate retractable claws
rlm@464 568 or =EVE='s hands, which are able to coalesce into one object in the
rlm@464 569 movie.
rlm@465 570
rlm@465 571 *** Solidifying/Connecting the body
rlm@465 572
rlm@465 573 Importing bodies from =CORTEX= into blender involves encoding
rlm@465 574 metadata into the blender file that specifies the mass of each
rlm@465 575 component and the joints by which those components are connected. I
rlm@465 576 do this in Blender in two ways. First is by using the ``metadata''
rlm@465 577 field of each solid object to specify the mass. Second is by using
rlm@465 578 Blender ``empty nodes'' to specify the position and type of each
rlm@465 579 joint. Empty nodes have no mass, physical presence, or appearance,
rlm@465 580 but they can hold metadata and have names. I use a tree structure
rlm@465 581 of empty nodes to specify joints. There is a parent node named
rlm@465 582 ``joints'', and a series of empty child nodes of the ``joints''
rlm@465 583 node that each represent a single joint.
rlm@465 584
rlm@465 585 #+caption: View of the hand model in Blender showing the main ``joints''
rlm@465 586 #+caption: node (highlighted in yellow) and its children which each
rlm@465 587 #+caption: represent a joint in the hand. Each joint node has metadata
rlm@465 588 #+caption: specifying what sort of joint it is.
rlm@465 589 #+ATTR_LaTeX: :width 10cm
rlm@465 590 [[./images/hand-screenshot1.png]]
rlm@465 591
rlm@465 592
rlm@465 593
rlm@465 594
rlm@464 595
rlm@464 596
rlm@464 597
rlm@436 598 ** Eyes reuse standard video game components
rlm@436 599
rlm@436 600 ** Hearing is hard; =CORTEX= does it right
rlm@436 601
rlm@436 602 ** Touch uses hundreds of hair-like elements
rlm@436 603
rlm@440 604 ** Proprioception is the sense that makes everything ``real''
rlm@436 605
rlm@436 606 ** Muscles are both effectors and sensors
rlm@436 607
rlm@436 608 ** =CORTEX= brings complex creatures to life!
rlm@436 609
rlm@436 610 ** =CORTEX= enables many possiblities for further research
rlm@435 611
rlm@465 612 * COMMENT Empathy in a simulated worm
rlm@435 613
rlm@449 614 Here I develop a computational model of empathy, using =CORTEX= as a
rlm@449 615 base. Empathy in this context is the ability to observe another
rlm@449 616 creature and infer what sorts of sensations that creature is
rlm@449 617 feeling. My empathy algorithm involves multiple phases. First is
rlm@449 618 free-play, where the creature moves around and gains sensory
rlm@449 619 experience. From this experience I construct a representation of the
rlm@449 620 creature's sensory state space, which I call \Phi-space. Using
rlm@449 621 \Phi-space, I construct an efficient function which takes the
rlm@449 622 limited data that comes from observing another creature and enriches
rlm@449 623 it full compliment of imagined sensory data. I can then use the
rlm@449 624 imagined sensory data to recognize what the observed creature is
rlm@449 625 doing and feeling, using straightforward embodied action predicates.
rlm@449 626 This is all demonstrated with using a simple worm-like creature, and
rlm@449 627 recognizing worm-actions based on limited data.
rlm@449 628
rlm@449 629 #+caption: Here is the worm with which we will be working.
rlm@449 630 #+caption: It is composed of 5 segments. Each segment has a
rlm@449 631 #+caption: pair of extensor and flexor muscles. Each of the
rlm@449 632 #+caption: worm's four joints is a hinge joint which allows
rlm@451 633 #+caption: about 30 degrees of rotation to either side. Each segment
rlm@449 634 #+caption: of the worm is touch-capable and has a uniform
rlm@449 635 #+caption: distribution of touch sensors on each of its faces.
rlm@449 636 #+caption: Each joint has a proprioceptive sense to detect
rlm@449 637 #+caption: relative positions. The worm segments are all the
rlm@449 638 #+caption: same except for the first one, which has a much
rlm@449 639 #+caption: higher weight than the others to allow for easy
rlm@449 640 #+caption: manual motor control.
rlm@449 641 #+name: basic-worm-view
rlm@449 642 #+ATTR_LaTeX: :width 10cm
rlm@449 643 [[./images/basic-worm-view.png]]
rlm@449 644
rlm@449 645 #+caption: Program for reading a worm from a blender file and
rlm@449 646 #+caption: outfitting it with the senses of proprioception,
rlm@449 647 #+caption: touch, and the ability to move, as specified in the
rlm@449 648 #+caption: blender file.
rlm@449 649 #+name: get-worm
rlm@449 650 #+begin_listing clojure
rlm@449 651 #+begin_src clojure
rlm@449 652 (defn worm []
rlm@449 653 (let [model (load-blender-model "Models/worm/worm.blend")]
rlm@449 654 {:body (doto model (body!))
rlm@449 655 :touch (touch! model)
rlm@449 656 :proprioception (proprioception! model)
rlm@449 657 :muscles (movement! model)}))
rlm@449 658 #+end_src
rlm@449 659 #+end_listing
rlm@452 660
rlm@436 661 ** Embodiment factors action recognition into managable parts
rlm@435 662
rlm@449 663 Using empathy, I divide the problem of action recognition into a
rlm@449 664 recognition process expressed in the language of a full compliment
rlm@449 665 of senses, and an imaganitive process that generates full sensory
rlm@449 666 data from partial sensory data. Splitting the action recognition
rlm@449 667 problem in this manner greatly reduces the total amount of work to
rlm@449 668 recognize actions: The imaganitive process is mostly just matching
rlm@449 669 previous experience, and the recognition process gets to use all
rlm@449 670 the senses to directly describe any action.
rlm@449 671
rlm@436 672 ** Action recognition is easy with a full gamut of senses
rlm@435 673
rlm@449 674 Embodied representations using multiple senses such as touch,
rlm@449 675 proprioception, and muscle tension turns out be be exceedingly
rlm@449 676 efficient at describing body-centered actions. It is the ``right
rlm@449 677 language for the job''. For example, it takes only around 5 lines
rlm@449 678 of LISP code to describe the action of ``curling'' using embodied
rlm@451 679 primitives. It takes about 10 lines to describe the seemingly
rlm@449 680 complicated action of wiggling.
rlm@449 681
rlm@449 682 The following action predicates each take a stream of sensory
rlm@449 683 experience, observe however much of it they desire, and decide
rlm@449 684 whether the worm is doing the action they describe. =curled?=
rlm@449 685 relies on proprioception, =resting?= relies on touch, =wiggling?=
rlm@449 686 relies on a fourier analysis of muscle contraction, and
rlm@449 687 =grand-circle?= relies on touch and reuses =curled?= as a gaurd.
rlm@449 688
rlm@449 689 #+caption: Program for detecting whether the worm is curled. This is the
rlm@449 690 #+caption: simplest action predicate, because it only uses the last frame
rlm@449 691 #+caption: of sensory experience, and only uses proprioceptive data. Even
rlm@449 692 #+caption: this simple predicate, however, is automatically frame
rlm@449 693 #+caption: independent and ignores vermopomorphic differences such as
rlm@449 694 #+caption: worm textures and colors.
rlm@449 695 #+name: curled
rlm@452 696 #+attr_latex: [htpb]
rlm@452 697 #+begin_listing clojure
rlm@449 698 #+begin_src clojure
rlm@449 699 (defn curled?
rlm@449 700 "Is the worm curled up?"
rlm@449 701 [experiences]
rlm@449 702 (every?
rlm@449 703 (fn [[_ _ bend]]
rlm@449 704 (> (Math/sin bend) 0.64))
rlm@449 705 (:proprioception (peek experiences))))
rlm@449 706 #+end_src
rlm@449 707 #+end_listing
rlm@449 708
rlm@449 709 #+caption: Program for summarizing the touch information in a patch
rlm@449 710 #+caption: of skin.
rlm@449 711 #+name: touch-summary
rlm@452 712 #+attr_latex: [htpb]
rlm@452 713
rlm@452 714 #+begin_listing clojure
rlm@449 715 #+begin_src clojure
rlm@449 716 (defn contact
rlm@449 717 "Determine how much contact a particular worm segment has with
rlm@449 718 other objects. Returns a value between 0 and 1, where 1 is full
rlm@449 719 contact and 0 is no contact."
rlm@449 720 [touch-region [coords contact :as touch]]
rlm@449 721 (-> (zipmap coords contact)
rlm@449 722 (select-keys touch-region)
rlm@449 723 (vals)
rlm@449 724 (#(map first %))
rlm@449 725 (average)
rlm@449 726 (* 10)
rlm@449 727 (- 1)
rlm@449 728 (Math/abs)))
rlm@449 729 #+end_src
rlm@449 730 #+end_listing
rlm@449 731
rlm@449 732
rlm@449 733 #+caption: Program for detecting whether the worm is at rest. This program
rlm@449 734 #+caption: uses a summary of the tactile information from the underbelly
rlm@449 735 #+caption: of the worm, and is only true if every segment is touching the
rlm@449 736 #+caption: floor. Note that this function contains no references to
rlm@449 737 #+caption: proprioction at all.
rlm@449 738 #+name: resting
rlm@452 739 #+attr_latex: [htpb]
rlm@452 740 #+begin_listing clojure
rlm@449 741 #+begin_src clojure
rlm@449 742 (def worm-segment-bottom (rect-region [8 15] [14 22]))
rlm@449 743
rlm@449 744 (defn resting?
rlm@449 745 "Is the worm resting on the ground?"
rlm@449 746 [experiences]
rlm@449 747 (every?
rlm@449 748 (fn [touch-data]
rlm@449 749 (< 0.9 (contact worm-segment-bottom touch-data)))
rlm@449 750 (:touch (peek experiences))))
rlm@449 751 #+end_src
rlm@449 752 #+end_listing
rlm@449 753
rlm@449 754 #+caption: Program for detecting whether the worm is curled up into a
rlm@449 755 #+caption: full circle. Here the embodied approach begins to shine, as
rlm@449 756 #+caption: I am able to both use a previous action predicate (=curled?=)
rlm@449 757 #+caption: as well as the direct tactile experience of the head and tail.
rlm@449 758 #+name: grand-circle
rlm@452 759 #+attr_latex: [htpb]
rlm@452 760 #+begin_listing clojure
rlm@449 761 #+begin_src clojure
rlm@449 762 (def worm-segment-bottom-tip (rect-region [15 15] [22 22]))
rlm@449 763
rlm@449 764 (def worm-segment-top-tip (rect-region [0 15] [7 22]))
rlm@449 765
rlm@449 766 (defn grand-circle?
rlm@449 767 "Does the worm form a majestic circle (one end touching the other)?"
rlm@449 768 [experiences]
rlm@449 769 (and (curled? experiences)
rlm@449 770 (let [worm-touch (:touch (peek experiences))
rlm@449 771 tail-touch (worm-touch 0)
rlm@449 772 head-touch (worm-touch 4)]
rlm@449 773 (and (< 0.55 (contact worm-segment-bottom-tip tail-touch))
rlm@449 774 (< 0.55 (contact worm-segment-top-tip head-touch))))))
rlm@449 775 #+end_src
rlm@449 776 #+end_listing
rlm@449 777
rlm@449 778
rlm@449 779 #+caption: Program for detecting whether the worm has been wiggling for
rlm@449 780 #+caption: the last few frames. It uses a fourier analysis of the muscle
rlm@449 781 #+caption: contractions of the worm's tail to determine wiggling. This is
rlm@449 782 #+caption: signigicant because there is no particular frame that clearly
rlm@449 783 #+caption: indicates that the worm is wiggling --- only when multiple frames
rlm@449 784 #+caption: are analyzed together is the wiggling revealed. Defining
rlm@449 785 #+caption: wiggling this way also gives the worm an opportunity to learn
rlm@449 786 #+caption: and recognize ``frustrated wiggling'', where the worm tries to
rlm@449 787 #+caption: wiggle but can't. Frustrated wiggling is very visually different
rlm@449 788 #+caption: from actual wiggling, but this definition gives it to us for free.
rlm@449 789 #+name: wiggling
rlm@452 790 #+attr_latex: [htpb]
rlm@452 791 #+begin_listing clojure
rlm@449 792 #+begin_src clojure
rlm@449 793 (defn fft [nums]
rlm@449 794 (map
rlm@449 795 #(.getReal %)
rlm@449 796 (.transform
rlm@449 797 (FastFourierTransformer. DftNormalization/STANDARD)
rlm@449 798 (double-array nums) TransformType/FORWARD)))
rlm@449 799
rlm@449 800 (def indexed (partial map-indexed vector))
rlm@449 801
rlm@449 802 (defn max-indexed [s]
rlm@449 803 (first (sort-by (comp - second) (indexed s))))
rlm@449 804
rlm@449 805 (defn wiggling?
rlm@449 806 "Is the worm wiggling?"
rlm@449 807 [experiences]
rlm@449 808 (let [analysis-interval 0x40]
rlm@449 809 (when (> (count experiences) analysis-interval)
rlm@449 810 (let [a-flex 3
rlm@449 811 a-ex 2
rlm@449 812 muscle-activity
rlm@449 813 (map :muscle (vector:last-n experiences analysis-interval))
rlm@449 814 base-activity
rlm@449 815 (map #(- (% a-flex) (% a-ex)) muscle-activity)]
rlm@449 816 (= 2
rlm@449 817 (first
rlm@449 818 (max-indexed
rlm@449 819 (map #(Math/abs %)
rlm@449 820 (take 20 (fft base-activity))))))))))
rlm@449 821 #+end_src
rlm@449 822 #+end_listing
rlm@449 823
rlm@449 824 With these action predicates, I can now recognize the actions of
rlm@449 825 the worm while it is moving under my control and I have access to
rlm@449 826 all the worm's senses.
rlm@449 827
rlm@449 828 #+caption: Use the action predicates defined earlier to report on
rlm@449 829 #+caption: what the worm is doing while in simulation.
rlm@449 830 #+name: report-worm-activity
rlm@452 831 #+attr_latex: [htpb]
rlm@452 832 #+begin_listing clojure
rlm@449 833 #+begin_src clojure
rlm@449 834 (defn debug-experience
rlm@449 835 [experiences text]
rlm@449 836 (cond
rlm@449 837 (grand-circle? experiences) (.setText text "Grand Circle")
rlm@449 838 (curled? experiences) (.setText text "Curled")
rlm@449 839 (wiggling? experiences) (.setText text "Wiggling")
rlm@449 840 (resting? experiences) (.setText text "Resting")))
rlm@449 841 #+end_src
rlm@449 842 #+end_listing
rlm@449 843
rlm@449 844 #+caption: Using =debug-experience=, the body-centered predicates
rlm@449 845 #+caption: work together to classify the behaviour of the worm.
rlm@451 846 #+caption: the predicates are operating with access to the worm's
rlm@451 847 #+caption: full sensory data.
rlm@449 848 #+name: basic-worm-view
rlm@449 849 #+ATTR_LaTeX: :width 10cm
rlm@449 850 [[./images/worm-identify-init.png]]
rlm@449 851
rlm@449 852 These action predicates satisfy the recognition requirement of an
rlm@451 853 empathic recognition system. There is power in the simplicity of
rlm@451 854 the action predicates. They describe their actions without getting
rlm@451 855 confused in visual details of the worm. Each one is frame
rlm@451 856 independent, but more than that, they are each indepent of
rlm@449 857 irrelevant visual details of the worm and the environment. They
rlm@449 858 will work regardless of whether the worm is a different color or
rlm@451 859 hevaily textured, or if the environment has strange lighting.
rlm@449 860
rlm@449 861 The trick now is to make the action predicates work even when the
rlm@449 862 sensory data on which they depend is absent. If I can do that, then
rlm@449 863 I will have gained much,
rlm@435 864
rlm@436 865 ** \Phi-space describes the worm's experiences
rlm@449 866
rlm@449 867 As a first step towards building empathy, I need to gather all of
rlm@449 868 the worm's experiences during free play. I use a simple vector to
rlm@449 869 store all the experiences.
rlm@449 870
rlm@449 871 Each element of the experience vector exists in the vast space of
rlm@449 872 all possible worm-experiences. Most of this vast space is actually
rlm@449 873 unreachable due to physical constraints of the worm's body. For
rlm@449 874 example, the worm's segments are connected by hinge joints that put
rlm@451 875 a practical limit on the worm's range of motions without limiting
rlm@451 876 its degrees of freedom. Some groupings of senses are impossible;
rlm@451 877 the worm can not be bent into a circle so that its ends are
rlm@451 878 touching and at the same time not also experience the sensation of
rlm@451 879 touching itself.
rlm@449 880
rlm@451 881 As the worm moves around during free play and its experience vector
rlm@451 882 grows larger, the vector begins to define a subspace which is all
rlm@451 883 the sensations the worm can practicaly experience during normal
rlm@451 884 operation. I call this subspace \Phi-space, short for
rlm@451 885 physical-space. The experience vector defines a path through
rlm@451 886 \Phi-space. This path has interesting properties that all derive
rlm@451 887 from physical embodiment. The proprioceptive components are
rlm@451 888 completely smooth, because in order for the worm to move from one
rlm@451 889 position to another, it must pass through the intermediate
rlm@451 890 positions. The path invariably forms loops as actions are repeated.
rlm@451 891 Finally and most importantly, proprioception actually gives very
rlm@451 892 strong inference about the other senses. For example, when the worm
rlm@451 893 is flat, you can infer that it is touching the ground and that its
rlm@451 894 muscles are not active, because if the muscles were active, the
rlm@451 895 worm would be moving and would not be perfectly flat. In order to
rlm@451 896 stay flat, the worm has to be touching the ground, or it would
rlm@451 897 again be moving out of the flat position due to gravity. If the
rlm@451 898 worm is positioned in such a way that it interacts with itself,
rlm@451 899 then it is very likely to be feeling the same tactile feelings as
rlm@451 900 the last time it was in that position, because it has the same body
rlm@451 901 as then. If you observe multiple frames of proprioceptive data,
rlm@451 902 then you can become increasingly confident about the exact
rlm@451 903 activations of the worm's muscles, because it generally takes a
rlm@451 904 unique combination of muscle contractions to transform the worm's
rlm@451 905 body along a specific path through \Phi-space.
rlm@449 906
rlm@449 907 There is a simple way of taking \Phi-space and the total ordering
rlm@449 908 provided by an experience vector and reliably infering the rest of
rlm@449 909 the senses.
rlm@435 910
rlm@436 911 ** Empathy is the process of tracing though \Phi-space
rlm@449 912
rlm@450 913 Here is the core of a basic empathy algorithm, starting with an
rlm@451 914 experience vector:
rlm@451 915
rlm@451 916 First, group the experiences into tiered proprioceptive bins. I use
rlm@451 917 powers of 10 and 3 bins, and the smallest bin has an approximate
rlm@451 918 size of 0.001 radians in all proprioceptive dimensions.
rlm@450 919
rlm@450 920 Then, given a sequence of proprioceptive input, generate a set of
rlm@451 921 matching experience records for each input, using the tiered
rlm@451 922 proprioceptive bins.
rlm@449 923
rlm@450 924 Finally, to infer sensory data, select the longest consective chain
rlm@451 925 of experiences. Conecutive experience means that the experiences
rlm@451 926 appear next to each other in the experience vector.
rlm@449 927
rlm@450 928 This algorithm has three advantages:
rlm@450 929
rlm@450 930 1. It's simple
rlm@450 931
rlm@451 932 3. It's very fast -- retrieving possible interpretations takes
rlm@451 933 constant time. Tracing through chains of interpretations takes
rlm@451 934 time proportional to the average number of experiences in a
rlm@451 935 proprioceptive bin. Redundant experiences in \Phi-space can be
rlm@451 936 merged to save computation.
rlm@450 937
rlm@450 938 2. It protects from wrong interpretations of transient ambiguous
rlm@451 939 proprioceptive data. For example, if the worm is flat for just
rlm@450 940 an instant, this flattness will not be interpreted as implying
rlm@450 941 that the worm has its muscles relaxed, since the flattness is
rlm@450 942 part of a longer chain which includes a distinct pattern of
rlm@451 943 muscle activation. Markov chains or other memoryless statistical
rlm@451 944 models that operate on individual frames may very well make this
rlm@451 945 mistake.
rlm@450 946
rlm@450 947 #+caption: Program to convert an experience vector into a
rlm@450 948 #+caption: proprioceptively binned lookup function.
rlm@450 949 #+name: bin
rlm@452 950 #+attr_latex: [htpb]
rlm@452 951 #+begin_listing clojure
rlm@450 952 #+begin_src clojure
rlm@449 953 (defn bin [digits]
rlm@449 954 (fn [angles]
rlm@449 955 (->> angles
rlm@449 956 (flatten)
rlm@449 957 (map (juxt #(Math/sin %) #(Math/cos %)))
rlm@449 958 (flatten)
rlm@449 959 (mapv #(Math/round (* % (Math/pow 10 (dec digits))))))))
rlm@449 960
rlm@449 961 (defn gen-phi-scan
rlm@450 962 "Nearest-neighbors with binning. Only returns a result if
rlm@450 963 the propriceptive data is within 10% of a previously recorded
rlm@450 964 result in all dimensions."
rlm@450 965 [phi-space]
rlm@449 966 (let [bin-keys (map bin [3 2 1])
rlm@449 967 bin-maps
rlm@449 968 (map (fn [bin-key]
rlm@449 969 (group-by
rlm@449 970 (comp bin-key :proprioception phi-space)
rlm@449 971 (range (count phi-space)))) bin-keys)
rlm@449 972 lookups (map (fn [bin-key bin-map]
rlm@450 973 (fn [proprio] (bin-map (bin-key proprio))))
rlm@450 974 bin-keys bin-maps)]
rlm@449 975 (fn lookup [proprio-data]
rlm@449 976 (set (some #(% proprio-data) lookups)))))
rlm@450 977 #+end_src
rlm@450 978 #+end_listing
rlm@449 979
rlm@451 980 #+caption: =longest-thread= finds the longest path of consecutive
rlm@451 981 #+caption: experiences to explain proprioceptive worm data.
rlm@451 982 #+name: phi-space-history-scan
rlm@451 983 #+ATTR_LaTeX: :width 10cm
rlm@451 984 [[./images/aurellem-gray.png]]
rlm@451 985
rlm@451 986 =longest-thread= infers sensory data by stitching together pieces
rlm@451 987 from previous experience. It prefers longer chains of previous
rlm@451 988 experience to shorter ones. For example, during training the worm
rlm@451 989 might rest on the ground for one second before it performs its
rlm@451 990 excercises. If during recognition the worm rests on the ground for
rlm@451 991 five seconds, =longest-thread= will accomodate this five second
rlm@451 992 rest period by looping the one second rest chain five times.
rlm@451 993
rlm@451 994 =longest-thread= takes time proportinal to the average number of
rlm@451 995 entries in a proprioceptive bin, because for each element in the
rlm@451 996 starting bin it performes a series of set lookups in the preceeding
rlm@451 997 bins. If the total history is limited, then this is only a constant
rlm@451 998 multiple times the number of entries in the starting bin. This
rlm@451 999 analysis also applies even if the action requires multiple longest
rlm@451 1000 chains -- it's still the average number of entries in a
rlm@451 1001 proprioceptive bin times the desired chain length. Because
rlm@451 1002 =longest-thread= is so efficient and simple, I can interpret
rlm@451 1003 worm-actions in real time.
rlm@449 1004
rlm@450 1005 #+caption: Program to calculate empathy by tracing though \Phi-space
rlm@450 1006 #+caption: and finding the longest (ie. most coherent) interpretation
rlm@450 1007 #+caption: of the data.
rlm@450 1008 #+name: longest-thread
rlm@452 1009 #+attr_latex: [htpb]
rlm@452 1010 #+begin_listing clojure
rlm@450 1011 #+begin_src clojure
rlm@449 1012 (defn longest-thread
rlm@449 1013 "Find the longest thread from phi-index-sets. The index sets should
rlm@449 1014 be ordered from most recent to least recent."
rlm@449 1015 [phi-index-sets]
rlm@449 1016 (loop [result '()
rlm@449 1017 [thread-bases & remaining :as phi-index-sets] phi-index-sets]
rlm@449 1018 (if (empty? phi-index-sets)
rlm@449 1019 (vec result)
rlm@449 1020 (let [threads
rlm@449 1021 (for [thread-base thread-bases]
rlm@449 1022 (loop [thread (list thread-base)
rlm@449 1023 remaining remaining]
rlm@449 1024 (let [next-index (dec (first thread))]
rlm@449 1025 (cond (empty? remaining) thread
rlm@449 1026 (contains? (first remaining) next-index)
rlm@449 1027 (recur
rlm@449 1028 (cons next-index thread) (rest remaining))
rlm@449 1029 :else thread))))
rlm@449 1030 longest-thread
rlm@449 1031 (reduce (fn [thread-a thread-b]
rlm@449 1032 (if (> (count thread-a) (count thread-b))
rlm@449 1033 thread-a thread-b))
rlm@449 1034 '(nil)
rlm@449 1035 threads)]
rlm@449 1036 (recur (concat longest-thread result)
rlm@449 1037 (drop (count longest-thread) phi-index-sets))))))
rlm@450 1038 #+end_src
rlm@450 1039 #+end_listing
rlm@450 1040
rlm@451 1041 There is one final piece, which is to replace missing sensory data
rlm@451 1042 with a best-guess estimate. While I could fill in missing data by
rlm@451 1043 using a gradient over the closest known sensory data points,
rlm@451 1044 averages can be misleading. It is certainly possible to create an
rlm@451 1045 impossible sensory state by averaging two possible sensory states.
rlm@451 1046 Therefore, I simply replicate the most recent sensory experience to
rlm@451 1047 fill in the gaps.
rlm@449 1048
rlm@449 1049 #+caption: Fill in blanks in sensory experience by replicating the most
rlm@449 1050 #+caption: recent experience.
rlm@449 1051 #+name: infer-nils
rlm@452 1052 #+attr_latex: [htpb]
rlm@452 1053 #+begin_listing clojure
rlm@449 1054 #+begin_src clojure
rlm@449 1055 (defn infer-nils
rlm@449 1056 "Replace nils with the next available non-nil element in the
rlm@449 1057 sequence, or barring that, 0."
rlm@449 1058 [s]
rlm@449 1059 (loop [i (dec (count s))
rlm@449 1060 v (transient s)]
rlm@449 1061 (if (zero? i) (persistent! v)
rlm@449 1062 (if-let [cur (v i)]
rlm@449 1063 (if (get v (dec i) 0)
rlm@449 1064 (recur (dec i) v)
rlm@449 1065 (recur (dec i) (assoc! v (dec i) cur)))
rlm@449 1066 (recur i (assoc! v i 0))))))
rlm@449 1067 #+end_src
rlm@449 1068 #+end_listing
rlm@435 1069
rlm@441 1070 ** Efficient action recognition with =EMPATH=
rlm@451 1071
rlm@451 1072 To use =EMPATH= with the worm, I first need to gather a set of
rlm@451 1073 experiences from the worm that includes the actions I want to
rlm@452 1074 recognize. The =generate-phi-space= program (listing
rlm@451 1075 \ref{generate-phi-space} runs the worm through a series of
rlm@451 1076 exercices and gatheres those experiences into a vector. The
rlm@451 1077 =do-all-the-things= program is a routine expressed in a simple
rlm@452 1078 muscle contraction script language for automated worm control. It
rlm@452 1079 causes the worm to rest, curl, and wiggle over about 700 frames
rlm@452 1080 (approx. 11 seconds).
rlm@425 1081
rlm@451 1082 #+caption: Program to gather the worm's experiences into a vector for
rlm@451 1083 #+caption: further processing. The =motor-control-program= line uses
rlm@451 1084 #+caption: a motor control script that causes the worm to execute a series
rlm@451 1085 #+caption: of ``exercices'' that include all the action predicates.
rlm@451 1086 #+name: generate-phi-space
rlm@452 1087 #+attr_latex: [htpb]
rlm@452 1088 #+begin_listing clojure
rlm@451 1089 #+begin_src clojure
rlm@451 1090 (def do-all-the-things
rlm@451 1091 (concat
rlm@451 1092 curl-script
rlm@451 1093 [[300 :d-ex 40]
rlm@451 1094 [320 :d-ex 0]]
rlm@451 1095 (shift-script 280 (take 16 wiggle-script))))
rlm@451 1096
rlm@451 1097 (defn generate-phi-space []
rlm@451 1098 (let [experiences (atom [])]
rlm@451 1099 (run-world
rlm@451 1100 (apply-map
rlm@451 1101 worm-world
rlm@451 1102 (merge
rlm@451 1103 (worm-world-defaults)
rlm@451 1104 {:end-frame 700
rlm@451 1105 :motor-control
rlm@451 1106 (motor-control-program worm-muscle-labels do-all-the-things)
rlm@451 1107 :experiences experiences})))
rlm@451 1108 @experiences))
rlm@451 1109 #+end_src
rlm@451 1110 #+end_listing
rlm@451 1111
rlm@451 1112 #+caption: Use longest thread and a phi-space generated from a short
rlm@451 1113 #+caption: exercise routine to interpret actions during free play.
rlm@451 1114 #+name: empathy-debug
rlm@452 1115 #+attr_latex: [htpb]
rlm@452 1116 #+begin_listing clojure
rlm@451 1117 #+begin_src clojure
rlm@451 1118 (defn init []
rlm@451 1119 (def phi-space (generate-phi-space))
rlm@451 1120 (def phi-scan (gen-phi-scan phi-space)))
rlm@451 1121
rlm@451 1122 (defn empathy-demonstration []
rlm@451 1123 (let [proprio (atom ())]
rlm@451 1124 (fn
rlm@451 1125 [experiences text]
rlm@451 1126 (let [phi-indices (phi-scan (:proprioception (peek experiences)))]
rlm@451 1127 (swap! proprio (partial cons phi-indices))
rlm@451 1128 (let [exp-thread (longest-thread (take 300 @proprio))
rlm@451 1129 empathy (mapv phi-space (infer-nils exp-thread))]
rlm@451 1130 (println-repl (vector:last-n exp-thread 22))
rlm@451 1131 (cond
rlm@451 1132 (grand-circle? empathy) (.setText text "Grand Circle")
rlm@451 1133 (curled? empathy) (.setText text "Curled")
rlm@451 1134 (wiggling? empathy) (.setText text "Wiggling")
rlm@451 1135 (resting? empathy) (.setText text "Resting")
rlm@451 1136 :else (.setText text "Unknown")))))))
rlm@451 1137
rlm@451 1138 (defn empathy-experiment [record]
rlm@451 1139 (.start (worm-world :experience-watch (debug-experience-phi)
rlm@451 1140 :record record :worm worm*)))
rlm@451 1141 #+end_src
rlm@451 1142 #+end_listing
rlm@451 1143
rlm@451 1144 The result of running =empathy-experiment= is that the system is
rlm@451 1145 generally able to interpret worm actions using the action-predicates
rlm@451 1146 on simulated sensory data just as well as with actual data. Figure
rlm@451 1147 \ref{empathy-debug-image} was generated using =empathy-experiment=:
rlm@451 1148
rlm@451 1149 #+caption: From only proprioceptive data, =EMPATH= was able to infer
rlm@451 1150 #+caption: the complete sensory experience and classify four poses
rlm@451 1151 #+caption: (The last panel shows a composite image of \emph{wriggling},
rlm@451 1152 #+caption: a dynamic pose.)
rlm@451 1153 #+name: empathy-debug-image
rlm@451 1154 #+ATTR_LaTeX: :width 10cm :placement [H]
rlm@451 1155 [[./images/empathy-1.png]]
rlm@451 1156
rlm@451 1157 One way to measure the performance of =EMPATH= is to compare the
rlm@451 1158 sutiability of the imagined sense experience to trigger the same
rlm@451 1159 action predicates as the real sensory experience.
rlm@451 1160
rlm@451 1161 #+caption: Determine how closely empathy approximates actual
rlm@451 1162 #+caption: sensory data.
rlm@451 1163 #+name: test-empathy-accuracy
rlm@452 1164 #+attr_latex: [htpb]
rlm@452 1165 #+begin_listing clojure
rlm@451 1166 #+begin_src clojure
rlm@451 1167 (def worm-action-label
rlm@451 1168 (juxt grand-circle? curled? wiggling?))
rlm@451 1169
rlm@451 1170 (defn compare-empathy-with-baseline [matches]
rlm@451 1171 (let [proprio (atom ())]
rlm@451 1172 (fn
rlm@451 1173 [experiences text]
rlm@451 1174 (let [phi-indices (phi-scan (:proprioception (peek experiences)))]
rlm@451 1175 (swap! proprio (partial cons phi-indices))
rlm@451 1176 (let [exp-thread (longest-thread (take 300 @proprio))
rlm@451 1177 empathy (mapv phi-space (infer-nils exp-thread))
rlm@451 1178 experience-matches-empathy
rlm@451 1179 (= (worm-action-label experiences)
rlm@451 1180 (worm-action-label empathy))]
rlm@451 1181 (println-repl experience-matches-empathy)
rlm@451 1182 (swap! matches #(conj % experience-matches-empathy)))))))
rlm@451 1183
rlm@451 1184 (defn accuracy [v]
rlm@451 1185 (float (/ (count (filter true? v)) (count v))))
rlm@451 1186
rlm@451 1187 (defn test-empathy-accuracy []
rlm@451 1188 (let [res (atom [])]
rlm@451 1189 (run-world
rlm@451 1190 (worm-world :experience-watch
rlm@451 1191 (compare-empathy-with-baseline res)
rlm@451 1192 :worm worm*))
rlm@451 1193 (accuracy @res)))
rlm@451 1194 #+end_src
rlm@451 1195 #+end_listing
rlm@451 1196
rlm@451 1197 Running =test-empathy-accuracy= using the very short exercise
rlm@451 1198 program defined in listing \ref{generate-phi-space}, and then doing
rlm@451 1199 a similar pattern of activity manually yeilds an accuracy of around
rlm@451 1200 73%. This is based on very limited worm experience. By training the
rlm@451 1201 worm for longer, the accuracy dramatically improves.
rlm@451 1202
rlm@451 1203 #+caption: Program to generate \Phi-space using manual training.
rlm@451 1204 #+name: manual-phi-space
rlm@452 1205 #+attr_latex: [htpb]
rlm@451 1206 #+begin_listing clojure
rlm@451 1207 #+begin_src clojure
rlm@451 1208 (defn init-interactive []
rlm@451 1209 (def phi-space
rlm@451 1210 (let [experiences (atom [])]
rlm@451 1211 (run-world
rlm@451 1212 (apply-map
rlm@451 1213 worm-world
rlm@451 1214 (merge
rlm@451 1215 (worm-world-defaults)
rlm@451 1216 {:experiences experiences})))
rlm@451 1217 @experiences))
rlm@451 1218 (def phi-scan (gen-phi-scan phi-space)))
rlm@451 1219 #+end_src
rlm@451 1220 #+end_listing
rlm@451 1221
rlm@451 1222 After about 1 minute of manual training, I was able to achieve 95%
rlm@451 1223 accuracy on manual testing of the worm using =init-interactive= and
rlm@452 1224 =test-empathy-accuracy=. The majority of errors are near the
rlm@452 1225 boundaries of transitioning from one type of action to another.
rlm@452 1226 During these transitions the exact label for the action is more open
rlm@452 1227 to interpretation, and dissaggrement between empathy and experience
rlm@452 1228 is more excusable.
rlm@450 1229
rlm@449 1230 ** Digression: bootstrapping touch using free exploration
rlm@449 1231
rlm@452 1232 In the previous section I showed how to compute actions in terms of
rlm@452 1233 body-centered predicates which relied averate touch activation of
rlm@452 1234 pre-defined regions of the worm's skin. What if, instead of recieving
rlm@452 1235 touch pre-grouped into the six faces of each worm segment, the true
rlm@452 1236 topology of the worm's skin was unknown? This is more similiar to how
rlm@452 1237 a nerve fiber bundle might be arranged. While two fibers that are
rlm@452 1238 close in a nerve bundle /might/ correspond to two touch sensors that
rlm@452 1239 are close together on the skin, the process of taking a complicated
rlm@452 1240 surface and forcing it into essentially a circle requires some cuts
rlm@452 1241 and rerragenments.
rlm@452 1242
rlm@452 1243 In this section I show how to automatically learn the skin-topology of
rlm@452 1244 a worm segment by free exploration. As the worm rolls around on the
rlm@452 1245 floor, large sections of its surface get activated. If the worm has
rlm@452 1246 stopped moving, then whatever region of skin that is touching the
rlm@452 1247 floor is probably an important region, and should be recorded.
rlm@452 1248
rlm@452 1249 #+caption: Program to detect whether the worm is in a resting state
rlm@452 1250 #+caption: with one face touching the floor.
rlm@452 1251 #+name: pure-touch
rlm@452 1252 #+begin_listing clojure
rlm@452 1253 #+begin_src clojure
rlm@452 1254 (def full-contact [(float 0.0) (float 0.1)])
rlm@452 1255
rlm@452 1256 (defn pure-touch?
rlm@452 1257 "This is worm specific code to determine if a large region of touch
rlm@452 1258 sensors is either all on or all off."
rlm@452 1259 [[coords touch :as touch-data]]
rlm@452 1260 (= (set (map first touch)) (set full-contact)))
rlm@452 1261 #+end_src
rlm@452 1262 #+end_listing
rlm@452 1263
rlm@452 1264 After collecting these important regions, there will many nearly
rlm@452 1265 similiar touch regions. While for some purposes the subtle
rlm@452 1266 differences between these regions will be important, for my
rlm@452 1267 purposes I colapse them into mostly non-overlapping sets using
rlm@452 1268 =remove-similiar= in listing \ref{remove-similiar}
rlm@452 1269
rlm@452 1270 #+caption: Program to take a lits of set of points and ``collapse them''
rlm@452 1271 #+caption: so that the remaining sets in the list are siginificantly
rlm@452 1272 #+caption: different from each other. Prefer smaller sets to larger ones.
rlm@452 1273 #+name: remove-similiar
rlm@452 1274 #+begin_listing clojure
rlm@452 1275 #+begin_src clojure
rlm@452 1276 (defn remove-similar
rlm@452 1277 [coll]
rlm@452 1278 (loop [result () coll (sort-by (comp - count) coll)]
rlm@452 1279 (if (empty? coll) result
rlm@452 1280 (let [[x & xs] coll
rlm@452 1281 c (count x)]
rlm@452 1282 (if (some
rlm@452 1283 (fn [other-set]
rlm@452 1284 (let [oc (count other-set)]
rlm@452 1285 (< (- (count (union other-set x)) c) (* oc 0.1))))
rlm@452 1286 xs)
rlm@452 1287 (recur result xs)
rlm@452 1288 (recur (cons x result) xs))))))
rlm@452 1289 #+end_src
rlm@452 1290 #+end_listing
rlm@452 1291
rlm@452 1292 Actually running this simulation is easy given =CORTEX='s facilities.
rlm@452 1293
rlm@452 1294 #+caption: Collect experiences while the worm moves around. Filter the touch
rlm@452 1295 #+caption: sensations by stable ones, collapse similiar ones together,
rlm@452 1296 #+caption: and report the regions learned.
rlm@452 1297 #+name: learn-touch
rlm@452 1298 #+begin_listing clojure
rlm@452 1299 #+begin_src clojure
rlm@452 1300 (defn learn-touch-regions []
rlm@452 1301 (let [experiences (atom [])
rlm@452 1302 world (apply-map
rlm@452 1303 worm-world
rlm@452 1304 (assoc (worm-segment-defaults)
rlm@452 1305 :experiences experiences))]
rlm@452 1306 (run-world world)
rlm@452 1307 (->>
rlm@452 1308 @experiences
rlm@452 1309 (drop 175)
rlm@452 1310 ;; access the single segment's touch data
rlm@452 1311 (map (comp first :touch))
rlm@452 1312 ;; only deal with "pure" touch data to determine surfaces
rlm@452 1313 (filter pure-touch?)
rlm@452 1314 ;; associate coordinates with touch values
rlm@452 1315 (map (partial apply zipmap))
rlm@452 1316 ;; select those regions where contact is being made
rlm@452 1317 (map (partial group-by second))
rlm@452 1318 (map #(get % full-contact))
rlm@452 1319 (map (partial map first))
rlm@452 1320 ;; remove redundant/subset regions
rlm@452 1321 (map set)
rlm@452 1322 remove-similar)))
rlm@452 1323
rlm@452 1324 (defn learn-and-view-touch-regions []
rlm@452 1325 (map view-touch-region
rlm@452 1326 (learn-touch-regions)))
rlm@452 1327 #+end_src
rlm@452 1328 #+end_listing
rlm@452 1329
rlm@452 1330 The only thing remining to define is the particular motion the worm
rlm@452 1331 must take. I accomplish this with a simple motor control program.
rlm@452 1332
rlm@452 1333 #+caption: Motor control program for making the worm roll on the ground.
rlm@452 1334 #+caption: This could also be replaced with random motion.
rlm@452 1335 #+name: worm-roll
rlm@452 1336 #+begin_listing clojure
rlm@452 1337 #+begin_src clojure
rlm@452 1338 (defn touch-kinesthetics []
rlm@452 1339 [[170 :lift-1 40]
rlm@452 1340 [190 :lift-1 19]
rlm@452 1341 [206 :lift-1 0]
rlm@452 1342
rlm@452 1343 [400 :lift-2 40]
rlm@452 1344 [410 :lift-2 0]
rlm@452 1345
rlm@452 1346 [570 :lift-2 40]
rlm@452 1347 [590 :lift-2 21]
rlm@452 1348 [606 :lift-2 0]
rlm@452 1349
rlm@452 1350 [800 :lift-1 30]
rlm@452 1351 [809 :lift-1 0]
rlm@452 1352
rlm@452 1353 [900 :roll-2 40]
rlm@452 1354 [905 :roll-2 20]
rlm@452 1355 [910 :roll-2 0]
rlm@452 1356
rlm@452 1357 [1000 :roll-2 40]
rlm@452 1358 [1005 :roll-2 20]
rlm@452 1359 [1010 :roll-2 0]
rlm@452 1360
rlm@452 1361 [1100 :roll-2 40]
rlm@452 1362 [1105 :roll-2 20]
rlm@452 1363 [1110 :roll-2 0]
rlm@452 1364 ])
rlm@452 1365 #+end_src
rlm@452 1366 #+end_listing
rlm@452 1367
rlm@452 1368
rlm@452 1369 #+caption: The small worm rolls around on the floor, driven
rlm@452 1370 #+caption: by the motor control program in listing \ref{worm-roll}.
rlm@452 1371 #+name: worm-roll
rlm@452 1372 #+ATTR_LaTeX: :width 12cm
rlm@452 1373 [[./images/worm-roll.png]]
rlm@452 1374
rlm@452 1375
rlm@452 1376 #+caption: After completing its adventures, the worm now knows
rlm@452 1377 #+caption: how its touch sensors are arranged along its skin. These
rlm@452 1378 #+caption: are the regions that were deemed important by
rlm@452 1379 #+caption: =learn-touch-regions=. Note that the worm has discovered
rlm@452 1380 #+caption: that it has six sides.
rlm@452 1381 #+name: worm-touch-map
rlm@452 1382 #+ATTR_LaTeX: :width 12cm
rlm@452 1383 [[./images/touch-learn.png]]
rlm@452 1384
rlm@452 1385 While simple, =learn-touch-regions= exploits regularities in both
rlm@452 1386 the worm's physiology and the worm's environment to correctly
rlm@452 1387 deduce that the worm has six sides. Note that =learn-touch-regions=
rlm@452 1388 would work just as well even if the worm's touch sense data were
rlm@452 1389 completely scrambled. The cross shape is just for convienence. This
rlm@452 1390 example justifies the use of pre-defined touch regions in =EMPATH=.
rlm@452 1391
rlm@465 1392 * COMMENT Contributions
rlm@454 1393
rlm@461 1394 In this thesis you have seen the =CORTEX= system, a complete
rlm@461 1395 environment for creating simulated creatures. You have seen how to
rlm@461 1396 implement five senses including touch, proprioception, hearing,
rlm@461 1397 vision, and muscle tension. You have seen how to create new creatues
rlm@461 1398 using blender, a 3D modeling tool. I hope that =CORTEX= will be
rlm@461 1399 useful in further research projects. To this end I have included the
rlm@461 1400 full source to =CORTEX= along with a large suite of tests and
rlm@461 1401 examples. I have also created a user guide for =CORTEX= which is
rlm@461 1402 inculded in an appendix to this thesis.
rlm@447 1403
rlm@461 1404 You have also seen how I used =CORTEX= as a platform to attach the
rlm@461 1405 /action recognition/ problem, which is the problem of recognizing
rlm@461 1406 actions in video. You saw a simple system called =EMPATH= which
rlm@461 1407 ientifies actions by first describing actions in a body-centerd,
rlm@461 1408 rich sense language, then infering a full range of sensory
rlm@461 1409 experience from limited data using previous experience gained from
rlm@461 1410 free play.
rlm@447 1411
rlm@461 1412 As a minor digression, you also saw how I used =CORTEX= to enable a
rlm@461 1413 tiny worm to discover the topology of its skin simply by rolling on
rlm@461 1414 the ground.
rlm@461 1415
rlm@461 1416 In conclusion, the main contributions of this thesis are:
rlm@461 1417
rlm@461 1418 - =CORTEX=, a system for creating simulated creatures with rich
rlm@461 1419 senses.
rlm@461 1420 - =EMPATH=, a program for recognizing actions by imagining sensory
rlm@461 1421 experience.
rlm@447 1422
rlm@447 1423 # An anatomical joke:
rlm@447 1424 # - Training
rlm@447 1425 # - Skeletal imitation
rlm@447 1426 # - Sensory fleshing-out
rlm@447 1427 # - Classification