Mercurial > cortex

annotate thesis/cortex.org @ 462:bb81cef09ad7

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
stuff about simulation vs reality.
author Robert McIntyre <rlm@mit.edu>
date Thu, 27 Mar 2014 20:18:51 -0400
parents b345650a0baa
children 6d55ac73bc6f
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@453 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@448 44 How could you ``see'' the chair in figure \ref{hidden-chair}?
rlm@441 45
rlm@441 46 #+caption: The chair in this image is quite obvious to humans, but I
rlm@448 47 #+caption: doubt that any modern computer vision program can find it.
rlm@441 48 #+name: hidden-chair
rlm@441 49 #+ATTR_LaTeX: :width 10cm
rlm@441 50 [[./images/fat-person-sitting-at-desk.jpg]]
rlm@441 51
rlm@441 52 Finally, how is it that you can easily tell the difference between
rlm@441 53 how the girls /muscles/ are working in figure \ref{girl}?
rlm@441 54
rlm@441 55 #+caption: The mysterious ``common sense'' appears here as you are able
rlm@441 56 #+caption: to discern the difference in how the girl's arm muscles
rlm@441 57 #+caption: are activated between the two images.
rlm@441 58 #+name: girl
rlm@448 59 #+ATTR_LaTeX: :width 7cm
rlm@441 60 [[./images/wall-push.png]]
rlm@437 61
rlm@441 62 Each of these examples tells us something about what might be going
rlm@441 63 on in our minds as we easily solve these recognition problems.
rlm@441 64
rlm@441 65 The hidden chairs show us that we are strongly triggered by cues
rlm@447 66 relating to the position of human bodies, and that we can determine
rlm@447 67 the overall physical configuration of a human body even if much of
rlm@447 68 that body is occluded.
rlm@437 69
rlm@441 70 The picture of the girl pushing against the wall tells us that we
rlm@441 71 have common sense knowledge about the kinetics of our own bodies.
rlm@441 72 We know well how our muscles would have to work to maintain us in
rlm@441 73 most positions, and we can easily project this self-knowledge to
rlm@441 74 imagined positions triggered by images of the human body.
rlm@441 75
rlm@441 76 ** =EMPATH= neatly solves recognition problems
rlm@441 77
rlm@441 78 I propose a system that can express the types of recognition
rlm@441 79 problems above in a form amenable to computation. It is split into
rlm@441 80 four parts:
rlm@441 81
rlm@448 82 - Free/Guided Play :: The creature moves around and experiences the
rlm@448 83 world through its unique perspective. Many otherwise
rlm@448 84 complicated actions are easily described in the language of a
rlm@448 85 full suite of body-centered, rich senses. For example,
rlm@448 86 drinking is the feeling of water sliding down your throat, and
rlm@448 87 cooling your insides. It's often accompanied by bringing your
rlm@448 88 hand close to your face, or bringing your face close to water.
rlm@448 89 Sitting down is the feeling of bending your knees, activating
rlm@448 90 your quadriceps, then feeling a surface with your bottom and
rlm@448 91 relaxing your legs. These body-centered action descriptions
rlm@448 92 can be either learned or hard coded.
rlm@448 93 - Posture Imitation :: When trying to interpret a video or image,
rlm@448 94 the creature takes a model of itself and aligns it with
rlm@448 95 whatever it sees. This alignment can even cross species, as
rlm@448 96 when humans try to align themselves with things like ponies,
rlm@448 97 dogs, or other humans with a different body type.
rlm@448 98 - Empathy :: The alignment triggers associations with
rlm@448 99 sensory data from prior experiences. For example, the
rlm@448 100 alignment itself easily maps to proprioceptive data. Any
rlm@448 101 sounds or obvious skin contact in the video can to a lesser
rlm@448 102 extent trigger previous experience. Segments of previous
rlm@448 103 experiences are stitched together to form a coherent and
rlm@448 104 complete sensory portrait of the scene.
rlm@448 105 - Recognition :: With the scene described in terms of first
rlm@448 106 person sensory events, the creature can now run its
rlm@447 107 action-identification programs on this synthesized sensory
rlm@447 108 data, just as it would if it were actually experiencing the
rlm@447 109 scene first-hand. If previous experience has been accurately
rlm@447 110 retrieved, and if it is analogous enough to the scene, then
rlm@447 111 the creature will correctly identify the action in the scene.
rlm@447 112
rlm@441 113 For example, I think humans are able to label the cat video as
rlm@447 114 ``drinking'' because they imagine /themselves/ as the cat, and
rlm@441 115 imagine putting their face up against a stream of water and
rlm@441 116 sticking out their tongue. In that imagined world, they can feel
rlm@441 117 the cool water hitting their tongue, and feel the water entering
rlm@447 118 their body, and are able to recognize that /feeling/ as drinking.
rlm@447 119 So, the label of the action is not really in the pixels of the
rlm@447 120 image, but is found clearly in a simulation inspired by those
rlm@447 121 pixels. An imaginative system, having been trained on drinking and
rlm@447 122 non-drinking examples and learning that the most important
rlm@447 123 component of drinking is the feeling of water sliding down one's
rlm@447 124 throat, would analyze a video of a cat drinking in the following
rlm@447 125 manner:
rlm@441 126
rlm@447 127 1. Create a physical model of the video by putting a ``fuzzy''
rlm@447 128 model of its own body in place of the cat. Possibly also create
rlm@447 129 a simulation of the stream of water.
rlm@441 130
rlm@441 131 2. Play out this simulated scene and generate imagined sensory
rlm@441 132 experience. This will include relevant muscle contractions, a
rlm@441 133 close up view of the stream from the cat's perspective, and most
rlm@441 134 importantly, the imagined feeling of water entering the
rlm@443 135 mouth. The imagined sensory experience can come from a
rlm@441 136 simulation of the event, but can also be pattern-matched from
rlm@441 137 previous, similar embodied experience.
rlm@441 138
rlm@441 139 3. The action is now easily identified as drinking by the sense of
rlm@441 140 taste alone. The other senses (such as the tongue moving in and
rlm@441 141 out) help to give plausibility to the simulated action. Note that
rlm@441 142 the sense of vision, while critical in creating the simulation,
rlm@441 143 is not critical for identifying the action from the simulation.
rlm@441 144
rlm@441 145 For the chair examples, the process is even easier:
rlm@441 146
rlm@441 147 1. Align a model of your body to the person in the image.
rlm@441 148
rlm@441 149 2. Generate proprioceptive sensory data from this alignment.
rlm@437 150
rlm@441 151 3. Use the imagined proprioceptive data as a key to lookup related
rlm@441 152 sensory experience associated with that particular proproceptive
rlm@441 153 feeling.
rlm@437 154
rlm@443 155 4. Retrieve the feeling of your bottom resting on a surface, your
rlm@443 156 knees bent, and your leg muscles relaxed.
rlm@437 157
rlm@441 158 5. This sensory information is consistent with the =sitting?=
rlm@441 159 sensory predicate, so you (and the entity in the image) must be
rlm@441 160 sitting.
rlm@440 161
rlm@441 162 6. There must be a chair-like object since you are sitting.
rlm@440 163
rlm@441 164 Empathy offers yet another alternative to the age-old AI
rlm@441 165 representation question: ``What is a chair?'' --- A chair is the
rlm@441 166 feeling of sitting.
rlm@441 167
rlm@441 168 My program, =EMPATH= uses this empathic problem solving technique
rlm@441 169 to interpret the actions of a simple, worm-like creature.
rlm@437 170
rlm@441 171 #+caption: The worm performs many actions during free play such as
rlm@441 172 #+caption: curling, wiggling, and resting.
rlm@441 173 #+name: worm-intro
rlm@446 174 #+ATTR_LaTeX: :width 15cm
rlm@445 175 [[./images/worm-intro-white.png]]
rlm@437 176
rlm@462 177 #+caption: =EMPATH= recognized and classified each of these
rlm@462 178 #+caption: poses by inferring the complete sensory experience
rlm@462 179 #+caption: from proprioceptive data.
rlm@441 180 #+name: worm-recognition-intro
rlm@446 181 #+ATTR_LaTeX: :width 15cm
rlm@445 182 [[./images/worm-poses.png]]
rlm@441 183
rlm@441 184 One powerful advantage of empathic problem solving is that it
rlm@441 185 factors the action recognition problem into two easier problems. To
rlm@441 186 use empathy, you need an /aligner/, which takes the video and a
rlm@441 187 model of your body, and aligns the model with the video. Then, you
rlm@441 188 need a /recognizer/, which uses the aligned model to interpret the
rlm@441 189 action. The power in this method lies in the fact that you describe
rlm@448 190 all actions form a body-centered viewpoint. You are less tied to
rlm@447 191 the particulars of any visual representation of the actions. If you
rlm@441 192 teach the system what ``running'' is, and you have a good enough
rlm@441 193 aligner, the system will from then on be able to recognize running
rlm@441 194 from any point of view, even strange points of view like above or
rlm@441 195 underneath the runner. This is in contrast to action recognition
rlm@448 196 schemes that try to identify actions using a non-embodied approach.
rlm@448 197 If these systems learn about running as viewed from the side, they
rlm@448 198 will not automatically be able to recognize running from any other
rlm@448 199 viewpoint.
rlm@441 200
rlm@441 201 Another powerful advantage is that using the language of multiple
rlm@441 202 body-centered rich senses to describe body-centerd actions offers a
rlm@441 203 massive boost in descriptive capability. Consider how difficult it
rlm@441 204 would be to compose a set of HOG filters to describe the action of
rlm@447 205 a simple worm-creature ``curling'' so that its head touches its
rlm@447 206 tail, and then behold the simplicity of describing thus action in a
rlm@441 207 language designed for the task (listing \ref{grand-circle-intro}):
rlm@441 208
rlm@446 209 #+caption: Body-centerd actions are best expressed in a body-centered
rlm@446 210 #+caption: language. This code detects when the worm has curled into a
rlm@446 211 #+caption: full circle. Imagine how you would replicate this functionality
rlm@446 212 #+caption: using low-level pixel features such as HOG filters!
rlm@446 213 #+name: grand-circle-intro
rlm@452 214 #+attr_latex: [htpb]
rlm@452 215 #+begin_listing clojure
rlm@446 216 #+begin_src clojure
rlm@446 217 (defn grand-circle?
rlm@446 218 "Does the worm form a majestic circle (one end touching the other)?"
rlm@446 219 [experiences]
rlm@446 220 (and (curled? experiences)
rlm@446 221 (let [worm-touch (:touch (peek experiences))
rlm@446 222 tail-touch (worm-touch 0)
rlm@446 223 head-touch (worm-touch 4)]
rlm@462 224 (and (< 0.2 (contact worm-segment-bottom-tip tail-touch))
rlm@462 225 (< 0.2 (contact worm-segment-top-tip head-touch))))))
rlm@446 226 #+end_src
rlm@446 227 #+end_listing
rlm@446 228
rlm@435 229
rlm@449 230 ** =CORTEX= is a toolkit for building sensate creatures
rlm@435 231
rlm@448 232 I built =CORTEX= to be a general AI research platform for doing
rlm@448 233 experiments involving multiple rich senses and a wide variety and
rlm@448 234 number of creatures. I intend it to be useful as a library for many
rlm@462 235 more projects than just this thesis. =CORTEX= was necessary to meet
rlm@462 236 a need among AI researchers at CSAIL and beyond, which is that
rlm@462 237 people often will invent neat ideas that are best expressed in the
rlm@448 238 language of creatures and senses, but in order to explore those
rlm@448 239 ideas they must first build a platform in which they can create
rlm@448 240 simulated creatures with rich senses! There are many ideas that
rlm@448 241 would be simple to execute (such as =EMPATH=), but attached to them
rlm@448 242 is the multi-month effort to make a good creature simulator. Often,
rlm@448 243 that initial investment of time proves to be too much, and the
rlm@448 244 project must make do with a lesser environment.
rlm@435 245
rlm@448 246 =CORTEX= is well suited as an environment for embodied AI research
rlm@448 247 for three reasons:
rlm@448 248
rlm@448 249 - You can create new creatures using Blender, a popular 3D modeling
rlm@448 250 program. Each sense can be specified using special blender nodes
rlm@448 251 with biologically inspired paramaters. You need not write any
rlm@448 252 code to create a creature, and can use a wide library of
rlm@448 253 pre-existing blender models as a base for your own creatures.
rlm@448 254
rlm@448 255 - =CORTEX= implements a wide variety of senses, including touch,
rlm@448 256 proprioception, vision, hearing, and muscle tension. Complicated
rlm@448 257 senses like touch, and vision involve multiple sensory elements
rlm@448 258 embedded in a 2D surface. You have complete control over the
rlm@448 259 distribution of these sensor elements through the use of simple
rlm@448 260 png image files. In particular, =CORTEX= implements more
rlm@448 261 comprehensive hearing than any other creature simulation system
rlm@448 262 available.
rlm@448 263
rlm@448 264 - =CORTEX= supports any number of creatures and any number of
rlm@448 265 senses. Time in =CORTEX= dialates so that the simulated creatures
rlm@448 266 always precieve a perfectly smooth flow of time, regardless of
rlm@448 267 the actual computational load.
rlm@448 268
rlm@448 269 =CORTEX= is built on top of =jMonkeyEngine3=, which is a video game
rlm@448 270 engine designed to create cross-platform 3D desktop games. =CORTEX=
rlm@448 271 is mainly written in clojure, a dialect of =LISP= that runs on the
rlm@448 272 java virtual machine (JVM). The API for creating and simulating
rlm@449 273 creatures and senses is entirely expressed in clojure, though many
rlm@449 274 senses are implemented at the layer of jMonkeyEngine or below. For
rlm@449 275 example, for the sense of hearing I use a layer of clojure code on
rlm@449 276 top of a layer of java JNI bindings that drive a layer of =C++=
rlm@449 277 code which implements a modified version of =OpenAL= to support
rlm@449 278 multiple listeners. =CORTEX= is the only simulation environment
rlm@449 279 that I know of that can support multiple entities that can each
rlm@449 280 hear the world from their own perspective. Other senses also
rlm@449 281 require a small layer of Java code. =CORTEX= also uses =bullet=, a
rlm@449 282 physics simulator written in =C=.
rlm@448 283
rlm@448 284 #+caption: Here is the worm from above modeled in Blender, a free
rlm@448 285 #+caption: 3D-modeling program. Senses and joints are described
rlm@448 286 #+caption: using special nodes in Blender.
rlm@448 287 #+name: worm-recognition-intro
rlm@448 288 #+ATTR_LaTeX: :width 12cm
rlm@448 289 [[./images/blender-worm.png]]
rlm@448 290
rlm@449 291 Here are some thing I anticipate that =CORTEX= might be used for:
rlm@449 292
rlm@449 293 - exploring new ideas about sensory integration
rlm@449 294 - distributed communication among swarm creatures
rlm@449 295 - self-learning using free exploration,
rlm@449 296 - evolutionary algorithms involving creature construction
rlm@449 297 - exploration of exoitic senses and effectors that are not possible
rlm@449 298 in the real world (such as telekenisis or a semantic sense)
rlm@449 299 - imagination using subworlds
rlm@449 300
rlm@451 301 During one test with =CORTEX=, I created 3,000 creatures each with
rlm@448 302 their own independent senses and ran them all at only 1/80 real
rlm@448 303 time. In another test, I created a detailed model of my own hand,
rlm@448 304 equipped with a realistic distribution of touch (more sensitive at
rlm@448 305 the fingertips), as well as eyes and ears, and it ran at around 1/4
rlm@451 306 real time.
rlm@448 307
rlm@451 308 #+BEGIN_LaTeX
rlm@449 309 \begin{sidewaysfigure}
rlm@449 310 \includegraphics[width=9.5in]{images/full-hand.png}
rlm@451 311 \caption{
rlm@451 312 I modeled my own right hand in Blender and rigged it with all the
rlm@451 313 senses that {\tt CORTEX} supports. My simulated hand has a
rlm@451 314 biologically inspired distribution of touch sensors. The senses are
rlm@451 315 displayed on the right, and the simulation is displayed on the
rlm@451 316 left. Notice that my hand is curling its fingers, that it can see
rlm@451 317 its own finger from the eye in its palm, and that it can feel its
rlm@451 318 own thumb touching its palm.}
rlm@449 319 \end{sidewaysfigure}
rlm@451 320 #+END_LaTeX
rlm@448 321
rlm@437 322 ** Contributions
rlm@435 323
rlm@451 324 - I built =CORTEX=, a comprehensive platform for embodied AI
rlm@451 325 experiments. =CORTEX= supports many features lacking in other
rlm@451 326 systems, such proper simulation of hearing. It is easy to create
rlm@451 327 new =CORTEX= creatures using Blender, a free 3D modeling program.
rlm@449 328
rlm@451 329 - I built =EMPATH=, which uses =CORTEX= to identify the actions of
rlm@451 330 a worm-like creature using a computational model of empathy.
rlm@449 331
rlm@436 332 * Building =CORTEX=
rlm@435 333
rlm@462 334 I intend for =CORTEX= to be used as a general purpose library for
rlm@462 335 building creatures and outfitting them with senses, so that it will
rlm@462 336 be useful for other researchers who want to test out ideas of their
rlm@462 337 own. To this end, wherver I have had to make archetictural choices
rlm@462 338 about =CORTEX=, I have chosen to give as much freedom to the user as
rlm@462 339 possible, so that =CORTEX= may be used for things I have not
rlm@462 340 forseen.
rlm@462 341
rlm@462 342 ** Simulation or Reality?
rlm@462 343
rlm@462 344 The most important archetictural decision of all is the choice to
rlm@462 345 use a computer-simulated environemnt in the first place! The world
rlm@462 346 is a vast and rich place, and for now simulations are a very poor
rlm@462 347 reflection of its complexity. It may be that there is a significant
rlm@462 348 qualatative difference between dealing with senses in the real
rlm@462 349 world and dealing with pale facilimilies of them in a
rlm@462 350 simulation. What are the advantages and disadvantages of a
rlm@462 351 simulation vs. reality?
rlm@462 352
rlm@462 353 *** Simulation
rlm@462 354
rlm@462 355 The advantages of virtual reality are that when everything is a
rlm@462 356 simulation, experiments in that simulation are absolutely
rlm@462 357 reproducible. It's also easier to change the character and world
rlm@462 358 to explore new situations and different sensory combinations.
rlm@462 359
rlm@462 360 If the world is to be simulated on a computer, then not only do
rlm@462 361 you have to worry about whether the character's senses are rich
rlm@462 362 enough to learn from the world, but whether the world itself is
rlm@462 363 rendered with enough detail and realism to give enough working
rlm@462 364 material to the character's senses. To name just a few
rlm@462 365 difficulties facing modern physics simulators: destructibility of
rlm@462 366 the environment, simulation of water/other fluids, large areas,
rlm@462 367 nonrigid bodies, lots of objects, smoke. I don't know of any
rlm@462 368 computer simulation that would allow a character to take a rock
rlm@462 369 and grind it into fine dust, then use that dust to make a clay
rlm@462 370 sculpture, at least not without spending years calculating the
rlm@462 371 interactions of every single small grain of dust. Maybe a
rlm@462 372 simulated world with today's limitations doesn't provide enough
rlm@462 373 richness for real intelligence to evolve.
rlm@462 374
rlm@462 375 *** Reality
rlm@462 376
rlm@462 377 The other approach for playing with senses is to hook your
rlm@462 378 software up to real cameras, microphones, robots, etc., and let it
rlm@462 379 loose in the real world. This has the advantage of eliminating
rlm@462 380 concerns about simulating the world at the expense of increasing
rlm@462 381 the complexity of implementing the senses. Instead of just
rlm@462 382 grabbing the current rendered frame for processing, you have to
rlm@462 383 use an actual camera with real lenses and interact with photons to
rlm@462 384 get an image. It is much harder to change the character, which is
rlm@462 385 now partly a physical robot of some sort, since doing so involves
rlm@462 386 changing things around in the real world instead of modifying
rlm@462 387 lines of code. While the real world is very rich and definitely
rlm@462 388 provides enough stimulation for intelligence to develop as
rlm@462 389 evidenced by our own existence, it is also uncontrollable in the
rlm@462 390 sense that a particular situation cannot be recreated perfectly or
rlm@462 391 saved for later use. It is harder to conduct science because it is
rlm@462 392 harder to repeat an experiment. The worst thing about using the
rlm@462 393 real world instead of a simulation is the matter of time. Instead
rlm@462 394 of simulated time you get the constant and unstoppable flow of
rlm@462 395 real time. This severely limits the sorts of software you can use
rlm@462 396 to program the AI because all sense inputs must be handled in real
rlm@462 397 time. Complicated ideas may have to be implemented in hardware or
rlm@462 398 may simply be impossible given the current speed of our
rlm@462 399 processors. Contrast this with a simulation, in which the flow of
rlm@462 400 time in the simulated world can be slowed down to accommodate the
rlm@462 401 limitations of the character's programming. In terms of cost,
rlm@462 402 doing everything in software is far cheaper than building custom
rlm@462 403 real-time hardware. All you need is a laptop and some patience.
rlm@435 404
rlm@436 405 ** Because of Time, simulation is perferable to reality
rlm@435 406
rlm@462 407 I envision =CORTEX= being used to support rapid prototyping and
rlm@462 408 iteration of ideas. Even if I could put together a well constructed
rlm@462 409 kit for creating robots, it would still not be enough because of
rlm@462 410 the scourge of real-time processing. Anyone who wants to test their
rlm@462 411 ideas in the real world must always worry about getting their
rlm@462 412 algorithms to run fast enough to process information in real
rlm@462 413 time. The need for real time processing only increases if multiple
rlm@462 414 senses are involved. In the extreme case, even simple algorithms
rlm@462 415 will have to be accelerated by ASIC chips or FPGAs, turning what
rlm@462 416 would otherwise be a few lines of code and a 10x speed penality
rlm@462 417 into a multi-month ordeal. For this reason, =CORTEX= supports
rlm@462 418 /time-dialiation/, which scales back the framerate of the
rlm@462 419 simulation in proportion to the amount of processing each
rlm@462 420 frame. From the perspective of the creatures inside the simulation,
rlm@462 421 time always appears to flow at a constant rate, regardless of how
rlm@462 422 complicated the envorimnent becomes or how many creatures are in
rlm@462 423 the simulation. The cost is that =CORTEX= can sometimes run slower
rlm@462 424 than real time. This can also be an advantage, however ---
rlm@462 425 simulations of very simple creatures in =CORTEX= generally run at
rlm@462 426 40x on my machine!
rlm@462 427
rlm@436 428 ** Video game engines are a great starting point
rlm@462 429
rlm@462 430 I did not need to write my own physics simulation code or shader to
rlm@462 431 build =CORTEX=. Doing so would lead to a system that is impossible
rlm@462 432 for anyone but myself to use anyway. Instead, I use a video game
rlm@462 433 engine as a base and modify it to accomodate the additional needs
rlm@462 434 of =CORTEX=. Video game engines are an ideal starting point to
rlm@462 435 build =CORTEX=, because they are not far from being creature
rlm@462 436 building systems themselves.
rlm@462 437
rlm@462 438 First off, general purpose video game engines come with a physics
rlm@462 439 engine and lighting / sound system. The physics system provides
rlm@462 440 tools that can be co-opted to serve as touch, proprioception, and
rlm@462 441 muscles. Since some games support split screen views, a good video
rlm@462 442 game engine will allow you to efficiently create multiple cameras
rlm@462 443 in the simulated world that can be used as eyes.
rlm@435 444
rlm@436 445 ** Bodies are composed of segments connected by joints
rlm@435 446
rlm@436 447 ** Eyes reuse standard video game components
rlm@436 448
rlm@436 449 ** Hearing is hard; =CORTEX= does it right
rlm@436 450
rlm@436 451 ** Touch uses hundreds of hair-like elements
rlm@436 452
rlm@440 453 ** Proprioception is the sense that makes everything ``real''
rlm@436 454
rlm@436 455 ** Muscles are both effectors and sensors
rlm@436 456
rlm@436 457 ** =CORTEX= brings complex creatures to life!
rlm@436 458
rlm@436 459 ** =CORTEX= enables many possiblities for further research
rlm@435 460
rlm@435 461 * Empathy in a simulated worm
rlm@435 462
rlm@449 463 Here I develop a computational model of empathy, using =CORTEX= as a
rlm@449 464 base. Empathy in this context is the ability to observe another
rlm@449 465 creature and infer what sorts of sensations that creature is
rlm@449 466 feeling. My empathy algorithm involves multiple phases. First is
rlm@449 467 free-play, where the creature moves around and gains sensory
rlm@449 468 experience. From this experience I construct a representation of the
rlm@449 469 creature's sensory state space, which I call \Phi-space. Using
rlm@449 470 \Phi-space, I construct an efficient function which takes the
rlm@449 471 limited data that comes from observing another creature and enriches
rlm@449 472 it full compliment of imagined sensory data. I can then use the
rlm@449 473 imagined sensory data to recognize what the observed creature is
rlm@449 474 doing and feeling, using straightforward embodied action predicates.
rlm@449 475 This is all demonstrated with using a simple worm-like creature, and
rlm@449 476 recognizing worm-actions based on limited data.
rlm@449 477
rlm@449 478 #+caption: Here is the worm with which we will be working.
rlm@449 479 #+caption: It is composed of 5 segments. Each segment has a
rlm@449 480 #+caption: pair of extensor and flexor muscles. Each of the
rlm@449 481 #+caption: worm's four joints is a hinge joint which allows
rlm@451 482 #+caption: about 30 degrees of rotation to either side. Each segment
rlm@449 483 #+caption: of the worm is touch-capable and has a uniform
rlm@449 484 #+caption: distribution of touch sensors on each of its faces.
rlm@449 485 #+caption: Each joint has a proprioceptive sense to detect
rlm@449 486 #+caption: relative positions. The worm segments are all the
rlm@449 487 #+caption: same except for the first one, which has a much
rlm@449 488 #+caption: higher weight than the others to allow for easy
rlm@449 489 #+caption: manual motor control.
rlm@449 490 #+name: basic-worm-view
rlm@449 491 #+ATTR_LaTeX: :width 10cm
rlm@449 492 [[./images/basic-worm-view.png]]
rlm@449 493
rlm@449 494 #+caption: Program for reading a worm from a blender file and
rlm@449 495 #+caption: outfitting it with the senses of proprioception,
rlm@449 496 #+caption: touch, and the ability to move, as specified in the
rlm@449 497 #+caption: blender file.
rlm@449 498 #+name: get-worm
rlm@449 499 #+begin_listing clojure
rlm@449 500 #+begin_src clojure
rlm@449 501 (defn worm []
rlm@449 502 (let [model (load-blender-model "Models/worm/worm.blend")]
rlm@449 503 {:body (doto model (body!))
rlm@449 504 :touch (touch! model)
rlm@449 505 :proprioception (proprioception! model)
rlm@449 506 :muscles (movement! model)}))
rlm@449 507 #+end_src
rlm@449 508 #+end_listing
rlm@452 509
rlm@436 510 ** Embodiment factors action recognition into managable parts
rlm@435 511
rlm@449 512 Using empathy, I divide the problem of action recognition into a
rlm@449 513 recognition process expressed in the language of a full compliment
rlm@449 514 of senses, and an imaganitive process that generates full sensory
rlm@449 515 data from partial sensory data. Splitting the action recognition
rlm@449 516 problem in this manner greatly reduces the total amount of work to
rlm@449 517 recognize actions: The imaganitive process is mostly just matching
rlm@449 518 previous experience, and the recognition process gets to use all
rlm@449 519 the senses to directly describe any action.
rlm@449 520
rlm@436 521 ** Action recognition is easy with a full gamut of senses
rlm@435 522
rlm@449 523 Embodied representations using multiple senses such as touch,
rlm@449 524 proprioception, and muscle tension turns out be be exceedingly
rlm@449 525 efficient at describing body-centered actions. It is the ``right
rlm@449 526 language for the job''. For example, it takes only around 5 lines
rlm@449 527 of LISP code to describe the action of ``curling'' using embodied
rlm@451 528 primitives. It takes about 10 lines to describe the seemingly
rlm@449 529 complicated action of wiggling.
rlm@449 530
rlm@449 531 The following action predicates each take a stream of sensory
rlm@449 532 experience, observe however much of it they desire, and decide
rlm@449 533 whether the worm is doing the action they describe. =curled?=
rlm@449 534 relies on proprioception, =resting?= relies on touch, =wiggling?=
rlm@449 535 relies on a fourier analysis of muscle contraction, and
rlm@449 536 =grand-circle?= relies on touch and reuses =curled?= as a gaurd.
rlm@449 537
rlm@449 538 #+caption: Program for detecting whether the worm is curled. This is the
rlm@449 539 #+caption: simplest action predicate, because it only uses the last frame
rlm@449 540 #+caption: of sensory experience, and only uses proprioceptive data. Even
rlm@449 541 #+caption: this simple predicate, however, is automatically frame
rlm@449 542 #+caption: independent and ignores vermopomorphic differences such as
rlm@449 543 #+caption: worm textures and colors.
rlm@449 544 #+name: curled
rlm@452 545 #+attr_latex: [htpb]
rlm@452 546 #+begin_listing clojure
rlm@449 547 #+begin_src clojure
rlm@449 548 (defn curled?
rlm@449 549 "Is the worm curled up?"
rlm@449 550 [experiences]
rlm@449 551 (every?
rlm@449 552 (fn [[_ _ bend]]
rlm@449 553 (> (Math/sin bend) 0.64))
rlm@449 554 (:proprioception (peek experiences))))
rlm@449 555 #+end_src
rlm@449 556 #+end_listing
rlm@449 557
rlm@449 558 #+caption: Program for summarizing the touch information in a patch
rlm@449 559 #+caption: of skin.
rlm@449 560 #+name: touch-summary
rlm@452 561 #+attr_latex: [htpb]
rlm@452 562
rlm@452 563 #+begin_listing clojure
rlm@449 564 #+begin_src clojure
rlm@449 565 (defn contact
rlm@449 566 "Determine how much contact a particular worm segment has with
rlm@449 567 other objects. Returns a value between 0 and 1, where 1 is full
rlm@449 568 contact and 0 is no contact."
rlm@449 569 [touch-region [coords contact :as touch]]
rlm@449 570 (-> (zipmap coords contact)
rlm@449 571 (select-keys touch-region)
rlm@449 572 (vals)
rlm@449 573 (#(map first %))
rlm@449 574 (average)
rlm@449 575 (* 10)
rlm@449 576 (- 1)
rlm@449 577 (Math/abs)))
rlm@449 578 #+end_src
rlm@449 579 #+end_listing
rlm@449 580
rlm@449 581
rlm@449 582 #+caption: Program for detecting whether the worm is at rest. This program
rlm@449 583 #+caption: uses a summary of the tactile information from the underbelly
rlm@449 584 #+caption: of the worm, and is only true if every segment is touching the
rlm@449 585 #+caption: floor. Note that this function contains no references to
rlm@449 586 #+caption: proprioction at all.
rlm@449 587 #+name: resting
rlm@452 588 #+attr_latex: [htpb]
rlm@452 589 #+begin_listing clojure
rlm@449 590 #+begin_src clojure
rlm@449 591 (def worm-segment-bottom (rect-region [8 15] [14 22]))
rlm@449 592
rlm@449 593 (defn resting?
rlm@449 594 "Is the worm resting on the ground?"
rlm@449 595 [experiences]
rlm@449 596 (every?
rlm@449 597 (fn [touch-data]
rlm@449 598 (< 0.9 (contact worm-segment-bottom touch-data)))
rlm@449 599 (:touch (peek experiences))))
rlm@449 600 #+end_src
rlm@449 601 #+end_listing
rlm@449 602
rlm@449 603 #+caption: Program for detecting whether the worm is curled up into a
rlm@449 604 #+caption: full circle. Here the embodied approach begins to shine, as
rlm@449 605 #+caption: I am able to both use a previous action predicate (=curled?=)
rlm@449 606 #+caption: as well as the direct tactile experience of the head and tail.
rlm@449 607 #+name: grand-circle
rlm@452 608 #+attr_latex: [htpb]
rlm@452 609 #+begin_listing clojure
rlm@449 610 #+begin_src clojure
rlm@449 611 (def worm-segment-bottom-tip (rect-region [15 15] [22 22]))
rlm@449 612
rlm@449 613 (def worm-segment-top-tip (rect-region [0 15] [7 22]))
rlm@449 614
rlm@449 615 (defn grand-circle?
rlm@449 616 "Does the worm form a majestic circle (one end touching the other)?"
rlm@449 617 [experiences]
rlm@449 618 (and (curled? experiences)
rlm@449 619 (let [worm-touch (:touch (peek experiences))
rlm@449 620 tail-touch (worm-touch 0)
rlm@449 621 head-touch (worm-touch 4)]
rlm@449 622 (and (< 0.55 (contact worm-segment-bottom-tip tail-touch))
rlm@449 623 (< 0.55 (contact worm-segment-top-tip head-touch))))))
rlm@449 624 #+end_src
rlm@449 625 #+end_listing
rlm@449 626
rlm@449 627
rlm@449 628 #+caption: Program for detecting whether the worm has been wiggling for
rlm@449 629 #+caption: the last few frames. It uses a fourier analysis of the muscle
rlm@449 630 #+caption: contractions of the worm's tail to determine wiggling. This is
rlm@449 631 #+caption: signigicant because there is no particular frame that clearly
rlm@449 632 #+caption: indicates that the worm is wiggling --- only when multiple frames
rlm@449 633 #+caption: are analyzed together is the wiggling revealed. Defining
rlm@449 634 #+caption: wiggling this way also gives the worm an opportunity to learn
rlm@449 635 #+caption: and recognize ``frustrated wiggling'', where the worm tries to
rlm@449 636 #+caption: wiggle but can't. Frustrated wiggling is very visually different
rlm@449 637 #+caption: from actual wiggling, but this definition gives it to us for free.
rlm@449 638 #+name: wiggling
rlm@452 639 #+attr_latex: [htpb]
rlm@452 640 #+begin_listing clojure
rlm@449 641 #+begin_src clojure
rlm@449 642 (defn fft [nums]
rlm@449 643 (map
rlm@449 644 #(.getReal %)
rlm@449 645 (.transform
rlm@449 646 (FastFourierTransformer. DftNormalization/STANDARD)
rlm@449 647 (double-array nums) TransformType/FORWARD)))
rlm@449 648
rlm@449 649 (def indexed (partial map-indexed vector))
rlm@449 650
rlm@449 651 (defn max-indexed [s]
rlm@449 652 (first (sort-by (comp - second) (indexed s))))
rlm@449 653
rlm@449 654 (defn wiggling?
rlm@449 655 "Is the worm wiggling?"
rlm@449 656 [experiences]
rlm@449 657 (let [analysis-interval 0x40]
rlm@449 658 (when (> (count experiences) analysis-interval)
rlm@449 659 (let [a-flex 3
rlm@449 660 a-ex 2
rlm@449 661 muscle-activity
rlm@449 662 (map :muscle (vector:last-n experiences analysis-interval))
rlm@449 663 base-activity
rlm@449 664 (map #(- (% a-flex) (% a-ex)) muscle-activity)]
rlm@449 665 (= 2
rlm@449 666 (first
rlm@449 667 (max-indexed
rlm@449 668 (map #(Math/abs %)
rlm@449 669 (take 20 (fft base-activity))))))))))
rlm@449 670 #+end_src
rlm@449 671 #+end_listing
rlm@449 672
rlm@449 673 With these action predicates, I can now recognize the actions of
rlm@449 674 the worm while it is moving under my control and I have access to
rlm@449 675 all the worm's senses.
rlm@449 676
rlm@449 677 #+caption: Use the action predicates defined earlier to report on
rlm@449 678 #+caption: what the worm is doing while in simulation.
rlm@449 679 #+name: report-worm-activity
rlm@452 680 #+attr_latex: [htpb]
rlm@452 681 #+begin_listing clojure
rlm@449 682 #+begin_src clojure
rlm@449 683 (defn debug-experience
rlm@449 684 [experiences text]
rlm@449 685 (cond
rlm@449 686 (grand-circle? experiences) (.setText text "Grand Circle")
rlm@449 687 (curled? experiences) (.setText text "Curled")
rlm@449 688 (wiggling? experiences) (.setText text "Wiggling")
rlm@449 689 (resting? experiences) (.setText text "Resting")))
rlm@449 690 #+end_src
rlm@449 691 #+end_listing
rlm@449 692
rlm@449 693 #+caption: Using =debug-experience=, the body-centered predicates
rlm@449 694 #+caption: work together to classify the behaviour of the worm.
rlm@451 695 #+caption: the predicates are operating with access to the worm's
rlm@451 696 #+caption: full sensory data.
rlm@449 697 #+name: basic-worm-view
rlm@449 698 #+ATTR_LaTeX: :width 10cm
rlm@449 699 [[./images/worm-identify-init.png]]
rlm@449 700
rlm@449 701 These action predicates satisfy the recognition requirement of an
rlm@451 702 empathic recognition system. There is power in the simplicity of
rlm@451 703 the action predicates. They describe their actions without getting
rlm@451 704 confused in visual details of the worm. Each one is frame
rlm@451 705 independent, but more than that, they are each indepent of
rlm@449 706 irrelevant visual details of the worm and the environment. They
rlm@449 707 will work regardless of whether the worm is a different color or
rlm@451 708 hevaily textured, or if the environment has strange lighting.
rlm@449 709
rlm@449 710 The trick now is to make the action predicates work even when the
rlm@449 711 sensory data on which they depend is absent. If I can do that, then
rlm@449 712 I will have gained much,
rlm@435 713
rlm@436 714 ** \Phi-space describes the worm's experiences
rlm@449 715
rlm@449 716 As a first step towards building empathy, I need to gather all of
rlm@449 717 the worm's experiences during free play. I use a simple vector to
rlm@449 718 store all the experiences.
rlm@449 719
rlm@449 720 Each element of the experience vector exists in the vast space of
rlm@449 721 all possible worm-experiences. Most of this vast space is actually
rlm@449 722 unreachable due to physical constraints of the worm's body. For
rlm@449 723 example, the worm's segments are connected by hinge joints that put
rlm@451 724 a practical limit on the worm's range of motions without limiting
rlm@451 725 its degrees of freedom. Some groupings of senses are impossible;
rlm@451 726 the worm can not be bent into a circle so that its ends are
rlm@451 727 touching and at the same time not also experience the sensation of
rlm@451 728 touching itself.
rlm@449 729
rlm@451 730 As the worm moves around during free play and its experience vector
rlm@451 731 grows larger, the vector begins to define a subspace which is all
rlm@451 732 the sensations the worm can practicaly experience during normal
rlm@451 733 operation. I call this subspace \Phi-space, short for
rlm@451 734 physical-space. The experience vector defines a path through
rlm@451 735 \Phi-space. This path has interesting properties that all derive
rlm@451 736 from physical embodiment. The proprioceptive components are
rlm@451 737 completely smooth, because in order for the worm to move from one
rlm@451 738 position to another, it must pass through the intermediate
rlm@451 739 positions. The path invariably forms loops as actions are repeated.
rlm@451 740 Finally and most importantly, proprioception actually gives very
rlm@451 741 strong inference about the other senses. For example, when the worm
rlm@451 742 is flat, you can infer that it is touching the ground and that its
rlm@451 743 muscles are not active, because if the muscles were active, the
rlm@451 744 worm would be moving and would not be perfectly flat. In order to
rlm@451 745 stay flat, the worm has to be touching the ground, or it would
rlm@451 746 again be moving out of the flat position due to gravity. If the
rlm@451 747 worm is positioned in such a way that it interacts with itself,
rlm@451 748 then it is very likely to be feeling the same tactile feelings as
rlm@451 749 the last time it was in that position, because it has the same body
rlm@451 750 as then. If you observe multiple frames of proprioceptive data,
rlm@451 751 then you can become increasingly confident about the exact
rlm@451 752 activations of the worm's muscles, because it generally takes a
rlm@451 753 unique combination of muscle contractions to transform the worm's
rlm@451 754 body along a specific path through \Phi-space.
rlm@449 755
rlm@449 756 There is a simple way of taking \Phi-space and the total ordering
rlm@449 757 provided by an experience vector and reliably infering the rest of
rlm@449 758 the senses.
rlm@435 759
rlm@436 760 ** Empathy is the process of tracing though \Phi-space
rlm@449 761
rlm@450 762 Here is the core of a basic empathy algorithm, starting with an
rlm@451 763 experience vector:
rlm@451 764
rlm@451 765 First, group the experiences into tiered proprioceptive bins. I use
rlm@451 766 powers of 10 and 3 bins, and the smallest bin has an approximate
rlm@451 767 size of 0.001 radians in all proprioceptive dimensions.
rlm@450 768
rlm@450 769 Then, given a sequence of proprioceptive input, generate a set of
rlm@451 770 matching experience records for each input, using the tiered
rlm@451 771 proprioceptive bins.
rlm@449 772
rlm@450 773 Finally, to infer sensory data, select the longest consective chain
rlm@451 774 of experiences. Conecutive experience means that the experiences
rlm@451 775 appear next to each other in the experience vector.
rlm@449 776
rlm@450 777 This algorithm has three advantages:
rlm@450 778
rlm@450 779 1. It's simple
rlm@450 780
rlm@451 781 3. It's very fast -- retrieving possible interpretations takes
rlm@451 782 constant time. Tracing through chains of interpretations takes
rlm@451 783 time proportional to the average number of experiences in a
rlm@451 784 proprioceptive bin. Redundant experiences in \Phi-space can be
rlm@451 785 merged to save computation.
rlm@450 786
rlm@450 787 2. It protects from wrong interpretations of transient ambiguous
rlm@451 788 proprioceptive data. For example, if the worm is flat for just
rlm@450 789 an instant, this flattness will not be interpreted as implying
rlm@450 790 that the worm has its muscles relaxed, since the flattness is
rlm@450 791 part of a longer chain which includes a distinct pattern of
rlm@451 792 muscle activation. Markov chains or other memoryless statistical
rlm@451 793 models that operate on individual frames may very well make this
rlm@451 794 mistake.
rlm@450 795
rlm@450 796 #+caption: Program to convert an experience vector into a
rlm@450 797 #+caption: proprioceptively binned lookup function.
rlm@450 798 #+name: bin
rlm@452 799 #+attr_latex: [htpb]
rlm@452 800 #+begin_listing clojure
rlm@450 801 #+begin_src clojure
rlm@449 802 (defn bin [digits]
rlm@449 803 (fn [angles]
rlm@449 804 (->> angles
rlm@449 805 (flatten)
rlm@449 806 (map (juxt #(Math/sin %) #(Math/cos %)))
rlm@449 807 (flatten)
rlm@449 808 (mapv #(Math/round (* % (Math/pow 10 (dec digits))))))))
rlm@449 809
rlm@449 810 (defn gen-phi-scan
rlm@450 811 "Nearest-neighbors with binning. Only returns a result if
rlm@450 812 the propriceptive data is within 10% of a previously recorded
rlm@450 813 result in all dimensions."
rlm@450 814 [phi-space]
rlm@449 815 (let [bin-keys (map bin [3 2 1])
rlm@449 816 bin-maps
rlm@449 817 (map (fn [bin-key]
rlm@449 818 (group-by
rlm@449 819 (comp bin-key :proprioception phi-space)
rlm@449 820 (range (count phi-space)))) bin-keys)
rlm@449 821 lookups (map (fn [bin-key bin-map]
rlm@450 822 (fn [proprio] (bin-map (bin-key proprio))))
rlm@450 823 bin-keys bin-maps)]
rlm@449 824 (fn lookup [proprio-data]
rlm@449 825 (set (some #(% proprio-data) lookups)))))
rlm@450 826 #+end_src
rlm@450 827 #+end_listing
rlm@449 828
rlm@451 829 #+caption: =longest-thread= finds the longest path of consecutive
rlm@451 830 #+caption: experiences to explain proprioceptive worm data.
rlm@451 831 #+name: phi-space-history-scan
rlm@451 832 #+ATTR_LaTeX: :width 10cm
rlm@451 833 [[./images/aurellem-gray.png]]
rlm@451 834
rlm@451 835 =longest-thread= infers sensory data by stitching together pieces
rlm@451 836 from previous experience. It prefers longer chains of previous
rlm@451 837 experience to shorter ones. For example, during training the worm
rlm@451 838 might rest on the ground for one second before it performs its
rlm@451 839 excercises. If during recognition the worm rests on the ground for
rlm@451 840 five seconds, =longest-thread= will accomodate this five second
rlm@451 841 rest period by looping the one second rest chain five times.
rlm@451 842
rlm@451 843 =longest-thread= takes time proportinal to the average number of
rlm@451 844 entries in a proprioceptive bin, because for each element in the
rlm@451 845 starting bin it performes a series of set lookups in the preceeding
rlm@451 846 bins. If the total history is limited, then this is only a constant
rlm@451 847 multiple times the number of entries in the starting bin. This
rlm@451 848 analysis also applies even if the action requires multiple longest
rlm@451 849 chains -- it's still the average number of entries in a
rlm@451 850 proprioceptive bin times the desired chain length. Because
rlm@451 851 =longest-thread= is so efficient and simple, I can interpret
rlm@451 852 worm-actions in real time.
rlm@449 853
rlm@450 854 #+caption: Program to calculate empathy by tracing though \Phi-space
rlm@450 855 #+caption: and finding the longest (ie. most coherent) interpretation
rlm@450 856 #+caption: of the data.
rlm@450 857 #+name: longest-thread
rlm@452 858 #+attr_latex: [htpb]
rlm@452 859 #+begin_listing clojure
rlm@450 860 #+begin_src clojure
rlm@449 861 (defn longest-thread
rlm@449 862 "Find the longest thread from phi-index-sets. The index sets should
rlm@449 863 be ordered from most recent to least recent."
rlm@449 864 [phi-index-sets]
rlm@449 865 (loop [result '()
rlm@449 866 [thread-bases & remaining :as phi-index-sets] phi-index-sets]
rlm@449 867 (if (empty? phi-index-sets)
rlm@449 868 (vec result)
rlm@449 869 (let [threads
rlm@449 870 (for [thread-base thread-bases]
rlm@449 871 (loop [thread (list thread-base)
rlm@449 872 remaining remaining]
rlm@449 873 (let [next-index (dec (first thread))]
rlm@449 874 (cond (empty? remaining) thread
rlm@449 875 (contains? (first remaining) next-index)
rlm@449 876 (recur
rlm@449 877 (cons next-index thread) (rest remaining))
rlm@449 878 :else thread))))
rlm@449 879 longest-thread
rlm@449 880 (reduce (fn [thread-a thread-b]
rlm@449 881 (if (> (count thread-a) (count thread-b))
rlm@449 882 thread-a thread-b))
rlm@449 883 '(nil)
rlm@449 884 threads)]
rlm@449 885 (recur (concat longest-thread result)
rlm@449 886 (drop (count longest-thread) phi-index-sets))))))
rlm@450 887 #+end_src
rlm@450 888 #+end_listing
rlm@450 889
rlm@451 890 There is one final piece, which is to replace missing sensory data
rlm@451 891 with a best-guess estimate. While I could fill in missing data by
rlm@451 892 using a gradient over the closest known sensory data points,
rlm@451 893 averages can be misleading. It is certainly possible to create an
rlm@451 894 impossible sensory state by averaging two possible sensory states.
rlm@451 895 Therefore, I simply replicate the most recent sensory experience to
rlm@451 896 fill in the gaps.
rlm@449 897
rlm@449 898 #+caption: Fill in blanks in sensory experience by replicating the most
rlm@449 899 #+caption: recent experience.
rlm@449 900 #+name: infer-nils
rlm@452 901 #+attr_latex: [htpb]
rlm@452 902 #+begin_listing clojure
rlm@449 903 #+begin_src clojure
rlm@449 904 (defn infer-nils
rlm@449 905 "Replace nils with the next available non-nil element in the
rlm@449 906 sequence, or barring that, 0."
rlm@449 907 [s]
rlm@449 908 (loop [i (dec (count s))
rlm@449 909 v (transient s)]
rlm@449 910 (if (zero? i) (persistent! v)
rlm@449 911 (if-let [cur (v i)]
rlm@449 912 (if (get v (dec i) 0)
rlm@449 913 (recur (dec i) v)
rlm@449 914 (recur (dec i) (assoc! v (dec i) cur)))
rlm@449 915 (recur i (assoc! v i 0))))))
rlm@449 916 #+end_src
rlm@449 917 #+end_listing
rlm@435 918
rlm@441 919 ** Efficient action recognition with =EMPATH=
rlm@451 920
rlm@451 921 To use =EMPATH= with the worm, I first need to gather a set of
rlm@451 922 experiences from the worm that includes the actions I want to
rlm@452 923 recognize. The =generate-phi-space= program (listing
rlm@451 924 \ref{generate-phi-space} runs the worm through a series of
rlm@451 925 exercices and gatheres those experiences into a vector. The
rlm@451 926 =do-all-the-things= program is a routine expressed in a simple
rlm@452 927 muscle contraction script language for automated worm control. It
rlm@452 928 causes the worm to rest, curl, and wiggle over about 700 frames
rlm@452 929 (approx. 11 seconds).
rlm@425 930
rlm@451 931 #+caption: Program to gather the worm's experiences into a vector for
rlm@451 932 #+caption: further processing. The =motor-control-program= line uses
rlm@451 933 #+caption: a motor control script that causes the worm to execute a series
rlm@451 934 #+caption: of ``exercices'' that include all the action predicates.
rlm@451 935 #+name: generate-phi-space
rlm@452 936 #+attr_latex: [htpb]
rlm@452 937 #+begin_listing clojure
rlm@451 938 #+begin_src clojure
rlm@451 939 (def do-all-the-things
rlm@451 940 (concat
rlm@451 941 curl-script
rlm@451 942 [[300 :d-ex 40]
rlm@451 943 [320 :d-ex 0]]
rlm@451 944 (shift-script 280 (take 16 wiggle-script))))
rlm@451 945
rlm@451 946 (defn generate-phi-space []
rlm@451 947 (let [experiences (atom [])]
rlm@451 948 (run-world
rlm@451 949 (apply-map
rlm@451 950 worm-world
rlm@451 951 (merge
rlm@451 952 (worm-world-defaults)
rlm@451 953 {:end-frame 700
rlm@451 954 :motor-control
rlm@451 955 (motor-control-program worm-muscle-labels do-all-the-things)
rlm@451 956 :experiences experiences})))
rlm@451 957 @experiences))
rlm@451 958 #+end_src
rlm@451 959 #+end_listing
rlm@451 960
rlm@451 961 #+caption: Use longest thread and a phi-space generated from a short
rlm@451 962 #+caption: exercise routine to interpret actions during free play.
rlm@451 963 #+name: empathy-debug
rlm@452 964 #+attr_latex: [htpb]
rlm@452 965 #+begin_listing clojure
rlm@451 966 #+begin_src clojure
rlm@451 967 (defn init []
rlm@451 968 (def phi-space (generate-phi-space))
rlm@451 969 (def phi-scan (gen-phi-scan phi-space)))
rlm@451 970
rlm@451 971 (defn empathy-demonstration []
rlm@451 972 (let [proprio (atom ())]
rlm@451 973 (fn
rlm@451 974 [experiences text]
rlm@451 975 (let [phi-indices (phi-scan (:proprioception (peek experiences)))]
rlm@451 976 (swap! proprio (partial cons phi-indices))
rlm@451 977 (let [exp-thread (longest-thread (take 300 @proprio))
rlm@451 978 empathy (mapv phi-space (infer-nils exp-thread))]
rlm@451 979 (println-repl (vector:last-n exp-thread 22))
rlm@451 980 (cond
rlm@451 981 (grand-circle? empathy) (.setText text "Grand Circle")
rlm@451 982 (curled? empathy) (.setText text "Curled")
rlm@451 983 (wiggling? empathy) (.setText text "Wiggling")
rlm@451 984 (resting? empathy) (.setText text "Resting")
rlm@451 985 :else (.setText text "Unknown")))))))
rlm@451 986
rlm@451 987 (defn empathy-experiment [record]
rlm@451 988 (.start (worm-world :experience-watch (debug-experience-phi)
rlm@451 989 :record record :worm worm*)))
rlm@451 990 #+end_src
rlm@451 991 #+end_listing
rlm@451 992
rlm@451 993 The result of running =empathy-experiment= is that the system is
rlm@451 994 generally able to interpret worm actions using the action-predicates
rlm@451 995 on simulated sensory data just as well as with actual data. Figure
rlm@451 996 \ref{empathy-debug-image} was generated using =empathy-experiment=:
rlm@451 997
rlm@451 998 #+caption: From only proprioceptive data, =EMPATH= was able to infer
rlm@451 999 #+caption: the complete sensory experience and classify four poses
rlm@451 1000 #+caption: (The last panel shows a composite image of \emph{wriggling},
rlm@451 1001 #+caption: a dynamic pose.)
rlm@451 1002 #+name: empathy-debug-image
rlm@451 1003 #+ATTR_LaTeX: :width 10cm :placement [H]
rlm@451 1004 [[./images/empathy-1.png]]
rlm@451 1005
rlm@451 1006 One way to measure the performance of =EMPATH= is to compare the
rlm@451 1007 sutiability of the imagined sense experience to trigger the same
rlm@451 1008 action predicates as the real sensory experience.
rlm@451 1009
rlm@451 1010 #+caption: Determine how closely empathy approximates actual
rlm@451 1011 #+caption: sensory data.
rlm@451 1012 #+name: test-empathy-accuracy
rlm@452 1013 #+attr_latex: [htpb]
rlm@452 1014 #+begin_listing clojure
rlm@451 1015 #+begin_src clojure
rlm@451 1016 (def worm-action-label
rlm@451 1017 (juxt grand-circle? curled? wiggling?))
rlm@451 1018
rlm@451 1019 (defn compare-empathy-with-baseline [matches]
rlm@451 1020 (let [proprio (atom ())]
rlm@451 1021 (fn
rlm@451 1022 [experiences text]
rlm@451 1023 (let [phi-indices (phi-scan (:proprioception (peek experiences)))]
rlm@451 1024 (swap! proprio (partial cons phi-indices))
rlm@451 1025 (let [exp-thread (longest-thread (take 300 @proprio))
rlm@451 1026 empathy (mapv phi-space (infer-nils exp-thread))
rlm@451 1027 experience-matches-empathy
rlm@451 1028 (= (worm-action-label experiences)
rlm@451 1029 (worm-action-label empathy))]
rlm@451 1030 (println-repl experience-matches-empathy)
rlm@451 1031 (swap! matches #(conj % experience-matches-empathy)))))))
rlm@451 1032
rlm@451 1033 (defn accuracy [v]
rlm@451 1034 (float (/ (count (filter true? v)) (count v))))
rlm@451 1035
rlm@451 1036 (defn test-empathy-accuracy []
rlm@451 1037 (let [res (atom [])]
rlm@451 1038 (run-world
rlm@451 1039 (worm-world :experience-watch
rlm@451 1040 (compare-empathy-with-baseline res)
rlm@451 1041 :worm worm*))
rlm@451 1042 (accuracy @res)))
rlm@451 1043 #+end_src
rlm@451 1044 #+end_listing
rlm@451 1045
rlm@451 1046 Running =test-empathy-accuracy= using the very short exercise
rlm@451 1047 program defined in listing \ref{generate-phi-space}, and then doing
rlm@451 1048 a similar pattern of activity manually yeilds an accuracy of around
rlm@451 1049 73%. This is based on very limited worm experience. By training the
rlm@451 1050 worm for longer, the accuracy dramatically improves.
rlm@451 1051
rlm@451 1052 #+caption: Program to generate \Phi-space using manual training.
rlm@451 1053 #+name: manual-phi-space
rlm@452 1054 #+attr_latex: [htpb]
rlm@451 1055 #+begin_listing clojure
rlm@451 1056 #+begin_src clojure
rlm@451 1057 (defn init-interactive []
rlm@451 1058 (def phi-space
rlm@451 1059 (let [experiences (atom [])]
rlm@451 1060 (run-world
rlm@451 1061 (apply-map
rlm@451 1062 worm-world
rlm@451 1063 (merge
rlm@451 1064 (worm-world-defaults)
rlm@451 1065 {:experiences experiences})))
rlm@451 1066 @experiences))
rlm@451 1067 (def phi-scan (gen-phi-scan phi-space)))
rlm@451 1068 #+end_src
rlm@451 1069 #+end_listing
rlm@451 1070
rlm@451 1071 After about 1 minute of manual training, I was able to achieve 95%
rlm@451 1072 accuracy on manual testing of the worm using =init-interactive= and
rlm@452 1073 =test-empathy-accuracy=. The majority of errors are near the
rlm@452 1074 boundaries of transitioning from one type of action to another.
rlm@452 1075 During these transitions the exact label for the action is more open
rlm@452 1076 to interpretation, and dissaggrement between empathy and experience
rlm@452 1077 is more excusable.
rlm@450 1078
rlm@449 1079 ** Digression: bootstrapping touch using free exploration
rlm@449 1080
rlm@452 1081 In the previous section I showed how to compute actions in terms of
rlm@452 1082 body-centered predicates which relied averate touch activation of
rlm@452 1083 pre-defined regions of the worm's skin. What if, instead of recieving
rlm@452 1084 touch pre-grouped into the six faces of each worm segment, the true
rlm@452 1085 topology of the worm's skin was unknown? This is more similiar to how
rlm@452 1086 a nerve fiber bundle might be arranged. While two fibers that are
rlm@452 1087 close in a nerve bundle /might/ correspond to two touch sensors that
rlm@452 1088 are close together on the skin, the process of taking a complicated
rlm@452 1089 surface and forcing it into essentially a circle requires some cuts
rlm@452 1090 and rerragenments.
rlm@452 1091
rlm@452 1092 In this section I show how to automatically learn the skin-topology of
rlm@452 1093 a worm segment by free exploration. As the worm rolls around on the
rlm@452 1094 floor, large sections of its surface get activated. If the worm has
rlm@452 1095 stopped moving, then whatever region of skin that is touching the
rlm@452 1096 floor is probably an important region, and should be recorded.
rlm@452 1097
rlm@452 1098 #+caption: Program to detect whether the worm is in a resting state
rlm@452 1099 #+caption: with one face touching the floor.
rlm@452 1100 #+name: pure-touch
rlm@452 1101 #+begin_listing clojure
rlm@452 1102 #+begin_src clojure
rlm@452 1103 (def full-contact [(float 0.0) (float 0.1)])
rlm@452 1104
rlm@452 1105 (defn pure-touch?
rlm@452 1106 "This is worm specific code to determine if a large region of touch
rlm@452 1107 sensors is either all on or all off."
rlm@452 1108 [[coords touch :as touch-data]]
rlm@452 1109 (= (set (map first touch)) (set full-contact)))
rlm@452 1110 #+end_src
rlm@452 1111 #+end_listing
rlm@452 1112
rlm@452 1113 After collecting these important regions, there will many nearly
rlm@452 1114 similiar touch regions. While for some purposes the subtle
rlm@452 1115 differences between these regions will be important, for my
rlm@452 1116 purposes I colapse them into mostly non-overlapping sets using
rlm@452 1117 =remove-similiar= in listing \ref{remove-similiar}
rlm@452 1118
rlm@452 1119 #+caption: Program to take a lits of set of points and ``collapse them''
rlm@452 1120 #+caption: so that the remaining sets in the list are siginificantly
rlm@452 1121 #+caption: different from each other. Prefer smaller sets to larger ones.
rlm@452 1122 #+name: remove-similiar
rlm@452 1123 #+begin_listing clojure
rlm@452 1124 #+begin_src clojure
rlm@452 1125 (defn remove-similar
rlm@452 1126 [coll]
rlm@452 1127 (loop [result () coll (sort-by (comp - count) coll)]
rlm@452 1128 (if (empty? coll) result
rlm@452 1129 (let [[x & xs] coll
rlm@452 1130 c (count x)]
rlm@452 1131 (if (some
rlm@452 1132 (fn [other-set]
rlm@452 1133 (let [oc (count other-set)]
rlm@452 1134 (< (- (count (union other-set x)) c) (* oc 0.1))))
rlm@452 1135 xs)
rlm@452 1136 (recur result xs)
rlm@452 1137 (recur (cons x result) xs))))))
rlm@452 1138 #+end_src
rlm@452 1139 #+end_listing
rlm@452 1140
rlm@452 1141 Actually running this simulation is easy given =CORTEX='s facilities.
rlm@452 1142
rlm@452 1143 #+caption: Collect experiences while the worm moves around. Filter the touch
rlm@452 1144 #+caption: sensations by stable ones, collapse similiar ones together,
rlm@452 1145 #+caption: and report the regions learned.
rlm@452 1146 #+name: learn-touch
rlm@452 1147 #+begin_listing clojure
rlm@452 1148 #+begin_src clojure
rlm@452 1149 (defn learn-touch-regions []
rlm@452 1150 (let [experiences (atom [])
rlm@452 1151 world (apply-map
rlm@452 1152 worm-world
rlm@452 1153 (assoc (worm-segment-defaults)
rlm@452 1154 :experiences experiences))]
rlm@452 1155 (run-world world)
rlm@452 1156 (->>
rlm@452 1157 @experiences
rlm@452 1158 (drop 175)
rlm@452 1159 ;; access the single segment's touch data
rlm@452 1160 (map (comp first :touch))
rlm@452 1161 ;; only deal with "pure" touch data to determine surfaces
rlm@452 1162 (filter pure-touch?)
rlm@452 1163 ;; associate coordinates with touch values
rlm@452 1164 (map (partial apply zipmap))
rlm@452 1165 ;; select those regions where contact is being made
rlm@452 1166 (map (partial group-by second))
rlm@452 1167 (map #(get % full-contact))
rlm@452 1168 (map (partial map first))
rlm@452 1169 ;; remove redundant/subset regions
rlm@452 1170 (map set)
rlm@452 1171 remove-similar)))
rlm@452 1172
rlm@452 1173 (defn learn-and-view-touch-regions []
rlm@452 1174 (map view-touch-region
rlm@452 1175 (learn-touch-regions)))
rlm@452 1176 #+end_src
rlm@452 1177 #+end_listing
rlm@452 1178
rlm@452 1179 The only thing remining to define is the particular motion the worm
rlm@452 1180 must take. I accomplish this with a simple motor control program.
rlm@452 1181
rlm@452 1182 #+caption: Motor control program for making the worm roll on the ground.
rlm@452 1183 #+caption: This could also be replaced with random motion.
rlm@452 1184 #+name: worm-roll
rlm@452 1185 #+begin_listing clojure
rlm@452 1186 #+begin_src clojure
rlm@452 1187 (defn touch-kinesthetics []
rlm@452 1188 [[170 :lift-1 40]
rlm@452 1189 [190 :lift-1 19]
rlm@452 1190 [206 :lift-1 0]
rlm@452 1191
rlm@452 1192 [400 :lift-2 40]
rlm@452 1193 [410 :lift-2 0]
rlm@452 1194
rlm@452 1195 [570 :lift-2 40]
rlm@452 1196 [590 :lift-2 21]
rlm@452 1197 [606 :lift-2 0]
rlm@452 1198
rlm@452 1199 [800 :lift-1 30]
rlm@452 1200 [809 :lift-1 0]
rlm@452 1201
rlm@452 1202 [900 :roll-2 40]
rlm@452 1203 [905 :roll-2 20]
rlm@452 1204 [910 :roll-2 0]
rlm@452 1205
rlm@452 1206 [1000 :roll-2 40]
rlm@452 1207 [1005 :roll-2 20]
rlm@452 1208 [1010 :roll-2 0]
rlm@452 1209
rlm@452 1210 [1100 :roll-2 40]
rlm@452 1211 [1105 :roll-2 20]
rlm@452 1212 [1110 :roll-2 0]
rlm@452 1213 ])
rlm@452 1214 #+end_src
rlm@452 1215 #+end_listing
rlm@452 1216
rlm@452 1217
rlm@452 1218 #+caption: The small worm rolls around on the floor, driven
rlm@452 1219 #+caption: by the motor control program in listing \ref{worm-roll}.
rlm@452 1220 #+name: worm-roll
rlm@452 1221 #+ATTR_LaTeX: :width 12cm
rlm@452 1222 [[./images/worm-roll.png]]
rlm@452 1223
rlm@452 1224
rlm@452 1225 #+caption: After completing its adventures, the worm now knows
rlm@452 1226 #+caption: how its touch sensors are arranged along its skin. These
rlm@452 1227 #+caption: are the regions that were deemed important by
rlm@452 1228 #+caption: =learn-touch-regions=. Note that the worm has discovered
rlm@452 1229 #+caption: that it has six sides.
rlm@452 1230 #+name: worm-touch-map
rlm@452 1231 #+ATTR_LaTeX: :width 12cm
rlm@452 1232 [[./images/touch-learn.png]]
rlm@452 1233
rlm@452 1234 While simple, =learn-touch-regions= exploits regularities in both
rlm@452 1235 the worm's physiology and the worm's environment to correctly
rlm@452 1236 deduce that the worm has six sides. Note that =learn-touch-regions=
rlm@452 1237 would work just as well even if the worm's touch sense data were
rlm@452 1238 completely scrambled. The cross shape is just for convienence. This
rlm@452 1239 example justifies the use of pre-defined touch regions in =EMPATH=.
rlm@452 1240
rlm@432 1241 * Contributions
rlm@454 1242
rlm@461 1243 In this thesis you have seen the =CORTEX= system, a complete
rlm@461 1244 environment for creating simulated creatures. You have seen how to
rlm@461 1245 implement five senses including touch, proprioception, hearing,
rlm@461 1246 vision, and muscle tension. You have seen how to create new creatues
rlm@461 1247 using blender, a 3D modeling tool. I hope that =CORTEX= will be
rlm@461 1248 useful in further research projects. To this end I have included the
rlm@461 1249 full source to =CORTEX= along with a large suite of tests and
rlm@461 1250 examples. I have also created a user guide for =CORTEX= which is
rlm@461 1251 inculded in an appendix to this thesis.
rlm@447 1252
rlm@461 1253 You have also seen how I used =CORTEX= as a platform to attach the
rlm@461 1254 /action recognition/ problem, which is the problem of recognizing
rlm@461 1255 actions in video. You saw a simple system called =EMPATH= which
rlm@461 1256 ientifies actions by first describing actions in a body-centerd,
rlm@461 1257 rich sense language, then infering a full range of sensory
rlm@461 1258 experience from limited data using previous experience gained from
rlm@461 1259 free play.
rlm@447 1260
rlm@461 1261 As a minor digression, you also saw how I used =CORTEX= to enable a
rlm@461 1262 tiny worm to discover the topology of its skin simply by rolling on
rlm@461 1263 the ground.
rlm@461 1264
rlm@461 1265 In conclusion, the main contributions of this thesis are:
rlm@461 1266
rlm@461 1267 - =CORTEX=, a system for creating simulated creatures with rich
rlm@461 1268 senses.
rlm@461 1269 - =EMPATH=, a program for recognizing actions by imagining sensory
rlm@461 1270 experience.
rlm@447 1271
rlm@447 1272 # An anatomical joke:
rlm@447 1273 # - Training
rlm@447 1274 # - Skeletal imitation
rlm@447 1275 # - Sensory fleshing-out
rlm@447 1276 # - Classification