CS 182 Sections 103 - 104 Eva Mok (emok@icsi.berkeley.edu) March 3, 2004 Some languages use an absolute orientation system, e.g. 1. Guugu Yimithirr (Cape York, Queensland, Australia) 2. Hai//om (Khoisan, Kalahari, Namibia, Africa) 3. Tzeltal (Mayan, Chiapas, Mexico) They describe all relations with East, South, West, North (e.g. the knife is south of the fork) Imagine using this system: The bathroom is ___________ of this classroom. I am standing ___________ of you guys. Announcements • a4 is due on Friday. • Midterm in class this Tuesday, March 7th • Be there on time! • Format: – closed books, closed notes – short answers, no blue books – up to March 2th lecture • Review Session: Monday, March 6th Time TBA (check the website), Location TBA Quick Recap • Last Week – Image Schemas and related experiments • This Week – Regier’s Model of learning spatial relation terms • Coming up – Motor control Quiz! 1. What are image schemas? What are they useful for? 2. How many senses of the English on can you think of? How would you distinguish them? 3. How is Regier’s model capable of learning without explicit negative examples? How does it learn spatial relation terms in different languages? 4. What mechanism do we use to link concepts together? How does it work? What are other models of learning? Quiz! 1. What are image schemas? What are they useful for? 2. How many senses of the English on can you think of? How would you distinguish them? 3. How is Regier’s model capable of learning without explicit negative examples? How does it learn spatial relation terms in different languages? 4. What mechanism do we use to link concepts together? How does it work? What are other models of learning? Why image schemas? • Different languages make different distinctions in their use of spatial relation terms • The weaker claim: Image schemas give us a ‘vocabulary’ to talk about the different dimensions of spatial structure that languages care about • The stronger claim: These dimensions are embodied -- our bodies constrain the way we observe and interact with the world. Therefore these schemas are universal. English AROUND ON OVER IN Bowerman & Pederson Dutch OP OM ANN BOVEN IN Bowerman & Pederson Chinese ZHOU LI SHANG Bowerman & Pederson The collection of image schemas • Trajector / Landmark (asymmetric) – The bike is near the house – ? The house is near the bike TR • Boundary / Bounded Region – a bounded region has a closed boundary LM boundary bounded region • Topological Relations – Separation, Contact, Overlap, Inclusion, Surround • Orientation – Vertical (up/down), Horizontal (left/right, front/back) – Absolute (E, S, W, N) More image schemas • Proximal / Distal – distance from center (near/far) • Part / Whole – top of the hill, cover of the magazine • Container – interior, exterior, boundary, portal • Source-Path-Goal – source, path, goal, trajector • Force-Dynamics – support, force S TR P G Quiz! 1. What are image schemas? What are they useful for? 2. How many senses of the English on can you think of? How would you distinguish them? 3. How is Regier’s model capable of learning without explicit negative examples? How does it learn spatial relation terms in different languages? 4. What mechanism do we use to link concepts together? How does it work? What are other models of learning? The English ‘on’ 1. The computer is on the desk 2. The picture is on the wall 3. The projector is on the ceiling UP TR LM DN TR/LM, verticality, contact, support TR LM TR LM TR/LM, contact, attaching force TR/LM, contact, attaching force Writing out the on construction SCHEMA TR-LM ROLES tr: trajector lm: landmark SCHEMA Contact ROLES r1: bounded-region r2: region SCHEMA Support ROLES supporter: bounded-region supportee: bounded-region The first sense of on CONSTRUCTION On FORM: Word self.f.orth “on” MEANING EVOKES TR-LM as trlm EVOKES Contact as c EVOKES Support as s trlm.tr c.r1 trlm.lm c.r2 trlm.tr s.supportee trlm.lm s.supporter Quiz! 1. What are image schemas? What are they useful for? 2. How many senses of the English on can you think of? How would you distinguish them? 3. How is Regier’s model capable of learning without explicit negative examples? How does it learn spatial relation terms in different languages? 4. What mechanism do we use to link concepts together? How does it work? What are other models of learning? Language Development in Children • 0-3 mo: prefers sounds in native language • 3-6 mo: imitation of vowel sounds only • 6-8 mo: babbling in consonant-vowel segments • 8-10 mo: word comprehension, starts to lose sensitivity to consonants outside native language • 12-13 mo: word production (naming) • 16-20 mo: word combinations, relational words (verbs, adj.) • 24-36 mo: grammaticization, inflectional morphology • 3 years – adulthood: vocab. growth, sentence-level grammar for discourse purposes Regier’s Model above below left right in out on off Learning System TR Input: above LM • Training input: configuration of TR/LM and the correct spatial relation term • Learned behavior: input TR/LM, output spatial relation Issue #1: Implicit Negatives • Children usually do not get explicit negatives • But we won’t know when to stop generalizing if we don’t have negative evidence • Yet spatial relation terms aren’t entirely mutually exclusive • The same scene can often be described with two or more spatial relation terms (e.g. above and outside) • How can we make the learning problem realistic yet learnable? Dealing with Implicit Negatives • Explicit positive for above • Implicit negatives for below, left, right, etc • in Regier: E = ½ ∑i,p (( ti,p – oi,p) * βi,p )2, where i is the node, p is the pattern, βi,p = 1 if explicit positive, βi,p < 1 if implicit negative Issue #2: Shift Invariance • Backprop cannot handle shift invariance (it cannot generalize from 0011, 0110 to 1100) • But the cup is on the table whether you see it right in the center or from the corner of your eyes (i.e. in different areas of the retina map) • What structure can we utilize to make the input shiftinvariant? Learning System dynamic relations (e.g. into) structured connectionist network (based on visual system) We’ll look at the details next lecture Quiz! 1. What are image schemas? What are they useful for? 2. How many senses of the English on can you think of? How would you distinguish them? 3. How is Regier’s model capable of learning without explicit negative examples? How does it learn spatial relation terms in different languages? 4. What mechanism do we use to link concepts together? How does it work? What are other models of learning? Models of Learning • Hebbian – coincidence • Reinforcement – delayed reward • Recruitment – one-trial • Supervised – correction (backprop) • Unsupervised – similarity Elman Nets & Jordan Nets Output 1 Output 1 Hidden Context Input α Hidden Context Input • Basically the same recurrent nets we mentioned a few weeks ago • Updating the context as we receive input • In Jordan nets we model “forgetting” as well • The recurrent connections have fixed weights • You can train these networks using good ol’ backprop Recurrent Backprop w2 a w4 b w1 c w3 unrolling 3 iterations a b c a b c a b c w1 a w2 w3 b w4 c • we’ll pretend to step through the network one iteration at a time • backprop as usual, but average equivalent weights (e.g. all 3 highlighted edges on the right are equivalent) The Idea of Recruitment Learning K Y • Suppose we want to link up concept X to concept Y • The idea is to pick the two nodes in the middle to link them up X N B F = B/N Pno link (1 F ) • Can we be sure that we can find a path to get from X to Y? BK the point is, with a fan-out of 1000, if we allow 2 intermediate layers, we can almost always find a path Remember triangle nodes? • Let’s say we are trying to remember a green circle • currently weak connections between concepts (dotted lines) has-color blue has-shape green round oval Strengthen these connections • and you end up with this picture has-color has-shape Green circle blue green round oval Demo of the Regier System on the English above above – positive examples above – negative examples above – after training above – test examples