1 00:00:00,070 --> 00:00:02,500 The following content is provided under a Creative 2 00:00:02,500 --> 00:00:04,019 Commons license. 3 00:00:04,019 --> 00:00:06,360 Your support will help MIT OpenCourseWare 4 00:00:06,360 --> 00:00:10,730 continue to offer high quality educational resources for free. 5 00:00:10,730 --> 00:00:13,340 To make a donation or view additional materials 6 00:00:13,340 --> 00:00:17,227 from hundreds of MIT courses, visit MIT OpenCourseWare 7 00:00:17,227 --> 00:00:17,852 at ocw.mit.edu. 8 00:00:21,724 --> 00:00:25,480 PROFESSOR: Here we have a near miss learning tree. 9 00:00:25,480 --> 00:00:29,220 It's a little bit different but a little bit similar. 10 00:00:29,220 --> 00:00:30,440 We've got different types. 11 00:00:30,440 --> 00:00:33,060 We're trying to learn about different light sources. 12 00:00:33,060 --> 00:00:36,300 So we're concerned about different lamps, flashlights, 13 00:00:36,300 --> 00:00:37,620 all sorts of things. 14 00:00:37,620 --> 00:00:41,770 So these guys can have had types of support, a base or wires. 15 00:00:41,770 --> 00:00:44,400 The base can be flat, clamped, or have legs. 16 00:00:44,400 --> 00:00:46,440 The legs can be flat bottomed or wheeled. 17 00:00:46,440 --> 00:00:49,030 The light source can be incandescent, fluorescent, 18 00:00:49,030 --> 00:00:50,270 or sodium vapor. 19 00:00:50,270 --> 00:00:51,780 And the energy source that powers it 20 00:00:51,780 --> 00:00:55,410 can be electric, battery, oil, or gas. 21 00:00:55,410 --> 00:00:57,350 So there's also some other things. 22 00:00:57,350 --> 00:01:00,840 But these are the trees that we might be climbing up 23 00:01:00,840 --> 00:01:03,840 with the climb-tree heuristic. 24 00:01:03,840 --> 00:01:07,050 Our starting model is we have an incandescent bulb. 25 00:01:07,050 --> 00:01:09,870 The height is equal to 24 inches exactly. 26 00:01:09,870 --> 00:01:11,840 They have a flat base. 27 00:01:11,840 --> 00:01:14,500 There's electricity to power it. 28 00:01:14,500 --> 00:01:16,350 And it has a shade. 29 00:01:16,350 --> 00:01:23,820 So our starting model is basically like a table lamp 30 00:01:23,820 --> 00:01:25,230 that you set on the table. 31 00:01:25,230 --> 00:01:26,890 It's got the shade over the top of it. 32 00:01:26,890 --> 00:01:29,250 You know, the standard room table lamp 33 00:01:29,250 --> 00:01:30,930 is our starting model. 34 00:01:30,930 --> 00:01:34,840 So recall in arch learning, which 35 00:01:34,840 --> 00:01:40,030 is what we are using here, that there are several heuristics 36 00:01:40,030 --> 00:01:42,110 that we might want to use. 37 00:01:42,110 --> 00:01:49,930 Let's see, well, the heuristics are-- we've 38 00:01:49,930 --> 00:01:58,765 got require-link, forbid-link. 39 00:02:04,810 --> 00:02:17,600 We've got climb-tree, extend-set, close-interval. 40 00:02:25,240 --> 00:02:34,570 And we also have drop-link-- forgot about that one. 41 00:02:34,570 --> 00:02:36,130 That's sort of with the links. 42 00:02:36,130 --> 00:02:36,880 I can put it here. 43 00:02:42,802 --> 00:02:47,910 OK, so if you want to do well on these questions 44 00:02:47,910 --> 00:02:50,680 if they appear on the final, first of all, 45 00:02:50,680 --> 00:02:54,070 you better know what those six heuristics do. 46 00:02:54,070 --> 00:03:04,330 So basically require-link is a heuristic. 47 00:03:04,330 --> 00:03:09,000 It goes into your model as a requirement, 48 00:03:09,000 --> 00:03:12,450 where before your model just didn't care. 49 00:03:12,450 --> 00:03:15,240 So let's say there was something about color for your light, 50 00:03:15,240 --> 00:03:17,970 like for your lamp's base or for the lamp. 51 00:03:17,970 --> 00:03:21,850 And right, now it doesn't care about color at all. 52 00:03:21,850 --> 00:03:28,370 But let's say that you wound up having an example that was, 53 00:03:28,370 --> 00:03:32,130 for instance, blue or white or whatever these colors are. 54 00:03:32,130 --> 00:03:43,610 And it wound up getting that blue was OK for-- in its model, 55 00:03:43,610 --> 00:03:48,410 it wound up getting, OK, blue and white are OK. 56 00:03:48,410 --> 00:03:50,790 Or actually, no, an even more basic example-- 57 00:03:50,790 --> 00:03:54,590 let's say that it just kept getting blue lamps as examples, 58 00:03:54,590 --> 00:03:55,760 and all of them were blue. 59 00:03:55,760 --> 00:03:58,218 And it just didn't care about the fact that they were blue. 60 00:03:58,218 --> 00:04:00,420 But they did turn out to be blue. 61 00:04:00,420 --> 00:04:03,060 And then eventually it found a red one 62 00:04:03,060 --> 00:04:04,780 that was exactly the same as its model, 63 00:04:04,780 --> 00:04:08,320 but the red one was a near miss. 64 00:04:08,320 --> 00:04:11,840 Then it might require blue, which 65 00:04:11,840 --> 00:04:16,110 was already an acceptable, permissible thing in its model. 66 00:04:16,110 --> 00:04:19,050 But it could be required. 67 00:04:19,050 --> 00:04:21,899 Now, another thing it can do is forbid. 68 00:04:21,899 --> 00:04:23,990 Some people ask me, why would you bother saying, 69 00:04:23,990 --> 00:04:28,500 forbid red, say, if you can already just require blue? 70 00:04:28,500 --> 00:04:31,940 Together they cover the same ground. 71 00:04:31,940 --> 00:04:35,110 The answer is because this system, this arch learning 72 00:04:35,110 --> 00:04:37,870 system-- remember, it's Patrick's doctoral thesis. 73 00:04:37,870 --> 00:04:40,560 It's an old system. 74 00:04:40,560 --> 00:04:42,200 Because it was built on old systems 75 00:04:42,200 --> 00:04:45,410 and because also it's generally a good idea, 76 00:04:45,410 --> 00:04:47,902 it was very parsimonious. 77 00:04:47,902 --> 00:04:49,360 Let's say you had a set of colors-- 78 00:04:49,360 --> 00:04:53,440 red, blue, yellow, green, pink, purple, orange-- 79 00:04:53,440 --> 00:04:59,080 and you found that only orange was giving you problems, 80 00:04:59,080 --> 00:05:04,800 you could try to require that it be in this. 81 00:05:04,800 --> 00:05:06,810 Or you could even use extend-set, 82 00:05:06,810 --> 00:05:13,060 make a giant set of red, blue, yellow, green, purple, pink, 83 00:05:13,060 --> 00:05:13,800 but not orange. 84 00:05:13,800 --> 00:05:15,750 But if you just forbid orange, you've 85 00:05:15,750 --> 00:05:18,050 just saved a whole bunch of space, 86 00:05:18,050 --> 00:05:21,506 particularly if there's a large possible series of set. 87 00:05:21,506 --> 00:05:23,130 You want to have the ability to forbid. 88 00:05:23,130 --> 00:05:27,220 Because it needs to have the smallest model that it can 89 00:05:27,220 --> 00:05:31,120 that covers all of its samples. 90 00:05:31,120 --> 00:05:35,180 So we also have climb-tree. 91 00:05:35,180 --> 00:05:40,070 Climb-tree takes one of the elements in our model 92 00:05:40,070 --> 00:05:41,805 and moves up the tree one level. 93 00:05:44,770 --> 00:05:48,355 So let's say we said, only incandescent lights are good. 94 00:05:48,355 --> 00:05:50,480 But we found that fluorescent lights are also good. 95 00:05:50,480 --> 00:05:53,300 Whoop, we move up to light source, climb-tree. 96 00:05:53,300 --> 00:05:56,230 Extend-set-- we don't have any sets here. 97 00:05:56,230 --> 00:06:00,210 But let's go back to red, blue, orange, yellow. 98 00:06:00,210 --> 00:06:03,670 Extend-set is when we say, only red is good as an example. 99 00:06:03,670 --> 00:06:05,810 But we see another positive example that's yellow. 100 00:06:05,810 --> 00:06:07,790 So we say, OK, I'll extend the set. 101 00:06:07,790 --> 00:06:09,670 Red and yellow are both good. 102 00:06:12,190 --> 00:06:16,100 All right, close-interval-- close-interval 103 00:06:16,100 --> 00:06:17,990 is really obvious when you need to use it. 104 00:06:17,990 --> 00:06:22,220 Because not only is it only used for intervals like height 105 00:06:22,220 --> 00:06:25,020 equals 24 inches or some kind of numbers, 106 00:06:25,020 --> 00:06:29,420 but it is the only thing you can use for when height 107 00:06:29,420 --> 00:06:30,840 equals some kind of numbers. 108 00:06:30,840 --> 00:06:34,270 Close-interval will let you fuss around with the intervals 109 00:06:34,270 --> 00:06:38,420 and say, oh, well I guess one with a 20 inch height is OK, 110 00:06:38,420 --> 00:06:39,020 too. 111 00:06:39,020 --> 00:06:42,134 So we'll make the whole interval 20 to 24 inches good. 112 00:06:42,134 --> 00:06:44,300 Because it doesn't make any sense for it to be like, 113 00:06:44,300 --> 00:06:46,820 20 is good, 24 is good, but nothing in between is good. 114 00:06:46,820 --> 00:06:48,120 No, those are awful. 115 00:06:48,120 --> 00:06:50,930 So that close-interval covers the entire interval. 116 00:06:50,930 --> 00:06:52,800 Last but not least is drop-link. 117 00:06:55,350 --> 00:07:00,080 The thing about drop-link is that drop-link is, again, 118 00:07:00,080 --> 00:07:02,790 only really used due to the fact that the system wants 119 00:07:02,790 --> 00:07:06,560 to be as parsimonious as possible with its model. 120 00:07:06,560 --> 00:07:10,880 Drop-link is-- let's say you have a color, and you're like, 121 00:07:10,880 --> 00:07:13,034 OK, only red is acceptable. 122 00:07:13,034 --> 00:07:14,200 And then you see a blue one. 123 00:07:14,200 --> 00:07:14,700 It's OK. 124 00:07:14,700 --> 00:07:17,074 And let's say red, blue, and yellow are your only colors. 125 00:07:17,074 --> 00:07:18,690 Only red and blue are acceptable. 126 00:07:18,690 --> 00:07:19,740 Hey, wait a minute. 127 00:07:19,740 --> 00:07:22,890 No, we can just say that only yellow 128 00:07:22,890 --> 00:07:25,350 is not acceptable, or something like that, all right? 129 00:07:25,350 --> 00:07:30,090 But then you see yellow as a positive example. 130 00:07:30,090 --> 00:07:32,236 Actually, you can't even switch to the yellow being 131 00:07:32,236 --> 00:07:33,090 not acceptable, because you've only 132 00:07:33,090 --> 00:07:34,339 seen positive of red and blue. 133 00:07:34,339 --> 00:07:36,505 So you say, only red and blue are acceptable. 134 00:07:36,505 --> 00:07:37,005 Question? 135 00:07:37,005 --> 00:07:39,183 AUDIENCE: You mentioned that if we 136 00:07:39,183 --> 00:07:40,766 had seen incandescent and fluorescent, 137 00:07:40,766 --> 00:07:42,840 it's OK that we could climb the tree to light source, 138 00:07:42,840 --> 00:07:44,131 because it's more parsimonious. 139 00:07:44,131 --> 00:07:47,217 If we then see one that's sodium vapor, say a street lamp, 140 00:07:47,217 --> 00:07:49,712 and we reject it, do we have to go back down the tree, 141 00:07:49,712 --> 00:07:50,638 or do we just add a forbid-link? 142 00:07:50,638 --> 00:07:52,471 PROFESSOR: Then you would add a forbid-link. 143 00:07:52,471 --> 00:07:55,240 The question is, we climb incandescent, fluorescent, 144 00:07:55,240 --> 00:07:56,704 to light source. 145 00:07:56,704 --> 00:07:58,370 When we see sodium vapor, what do we do? 146 00:07:58,370 --> 00:08:03,240 The answer is, our system, unlike with lattice learning, 147 00:08:03,240 --> 00:08:05,680 in arch learning, our system is memoryless 148 00:08:05,680 --> 00:08:06,970 of its previous examples. 149 00:08:06,970 --> 00:08:09,400 It's incapable of going back down the tree. 150 00:08:09,400 --> 00:08:11,720 And so if you see incandescent and fluorescent, 151 00:08:11,720 --> 00:08:14,461 climb up to light source, you're only recourse, 152 00:08:14,461 --> 00:08:16,210 if you see that sodium vapor doesn't work, 153 00:08:16,210 --> 00:08:18,292 is to forbid-link sodium vapor. 154 00:08:18,292 --> 00:08:19,250 That's a good question. 155 00:08:19,250 --> 00:08:21,480 It shows good understanding of how we work. 156 00:08:21,480 --> 00:08:24,474 So drop-link, as you see, red and blue, they both are OK. 157 00:08:24,474 --> 00:08:26,890 You don't know about yellow, so you can't just switch that 158 00:08:26,890 --> 00:08:28,015 to a forbid-link to yellow. 159 00:08:28,015 --> 00:08:31,350 But then you see yellow is also OK. 160 00:08:31,350 --> 00:08:32,900 What do you do? 161 00:08:32,900 --> 00:08:34,429 You can drop the link altogether, 162 00:08:34,429 --> 00:08:36,120 save yourself space. 163 00:08:36,120 --> 00:08:39,470 If all the colors that you have in the entire world are fine, 164 00:08:39,470 --> 00:08:42,380 you can just not have part of your model be the color, 165 00:08:42,380 --> 00:08:43,809 and just say, any color is fine. 166 00:08:43,809 --> 00:08:48,500 Color must not be really part of what defines a lamp. 167 00:08:48,500 --> 00:08:50,190 And I'd say it's not. 168 00:08:50,190 --> 00:08:52,810 I mean, it's what defines a tacky lamp, I guess. 169 00:08:52,810 --> 00:08:54,890 But it's not what defines a lamp. 170 00:08:54,890 --> 00:08:58,270 So that might be why color isn't here to begin with. 171 00:08:58,270 --> 00:09:01,400 So to make yourself do better on this, 172 00:09:01,400 --> 00:09:03,320 more knowledge equals less search. 173 00:09:03,320 --> 00:09:05,080 Let's have more knowledge. 174 00:09:05,080 --> 00:09:07,980 There's only some of these are used to generalize, 175 00:09:07,980 --> 00:09:11,110 to make our model more accepting of new examples. 176 00:09:11,110 --> 00:09:12,950 And those are ones we would use after we've 177 00:09:12,950 --> 00:09:15,660 seen a positive hit, sort of generalize, 178 00:09:15,660 --> 00:09:17,380 learn more things in the model. 179 00:09:17,380 --> 00:09:20,460 Some of these are used to specialize and make 180 00:09:20,460 --> 00:09:24,850 our model refined, make our model more specific. 181 00:09:24,850 --> 00:09:27,440 You'd only use those after you saw a near miss. 182 00:09:27,440 --> 00:09:29,296 So let's actually separate those. 183 00:09:29,296 --> 00:09:30,670 That way you guys will never make 184 00:09:30,670 --> 00:09:34,630 the mistake of using one of these in the wrong situation. 185 00:09:34,630 --> 00:09:36,460 And more knowledge, less search. 186 00:09:36,460 --> 00:09:38,510 And less search means a faster quiz time. 187 00:09:38,510 --> 00:09:40,490 So require-link, what do you guys think? 188 00:09:40,490 --> 00:09:42,864 Is that a specializer or a generalizer? 189 00:09:42,864 --> 00:09:43,780 AUDIENCE: Specializer. 190 00:09:43,780 --> 00:09:45,370 PROFESSOR: Specializer, that's right. 191 00:09:45,370 --> 00:09:47,801 We'd only use it when we saw a near miss. 192 00:09:47,801 --> 00:09:49,704 Forbid-link, specializer or generalizer? 193 00:09:49,704 --> 00:09:50,620 AUDIENCE: Specializer. 194 00:09:50,620 --> 00:09:51,578 PROFESSOR: Specializer. 195 00:09:51,578 --> 00:09:54,370 We'd only forbid something if we saw a miss. 196 00:09:54,370 --> 00:09:56,176 Climb-tree, specializer or generalizer? 197 00:09:56,176 --> 00:09:57,050 AUDIENCE: Generalizer 198 00:09:57,050 --> 00:09:58,140 PROFESSOR: Generalizer, that's right. 199 00:09:58,140 --> 00:10:00,810 We'd only climb up to a more generic thing in the tree 200 00:10:00,810 --> 00:10:02,340 if we saw a positive example. 201 00:10:02,340 --> 00:10:03,364 Extend-set? 202 00:10:03,364 --> 00:10:04,280 AUDIENCE: Generalizer. 203 00:10:04,280 --> 00:10:05,238 PROFESSOR: Generalizer. 204 00:10:05,238 --> 00:10:06,920 We'd only extend the things in our set 205 00:10:06,920 --> 00:10:08,810 if we saw a positive example. 206 00:10:08,810 --> 00:10:10,124 Close-interval? 207 00:10:10,124 --> 00:10:11,040 AUDIENCE: Generalizer. 208 00:10:11,040 --> 00:10:12,373 PROFESSOR: That's a generalizer. 209 00:10:12,373 --> 00:10:15,710 We'd only make the interval bigger 210 00:10:15,710 --> 00:10:20,420 if we saw a positive example. 211 00:10:20,420 --> 00:10:26,640 Someone asks, what do you do if you have like 10 to 30, 212 00:10:26,640 --> 00:10:31,050 and then you find a negative example in 20? 213 00:10:31,050 --> 00:10:33,937 It's generally pretty annoying for this system 214 00:10:33,937 --> 00:10:35,020 to have to deal with that. 215 00:10:35,020 --> 00:10:36,300 There's a variety of things you could do 216 00:10:36,300 --> 00:10:37,510 based on your implementation. 217 00:10:37,510 --> 00:10:42,000 I've never seen us ask it in a quiz. 218 00:10:42,000 --> 00:10:46,450 But one thing you could do is forbid 20, exactly just 20, 219 00:10:46,450 --> 00:10:52,350 and just have a little hole at exactly 20. 220 00:10:52,350 --> 00:10:54,350 So then drop-link-- specializer or generalizer? 221 00:10:54,350 --> 00:10:55,310 AUDIENCE: Generalizer. 222 00:10:55,310 --> 00:10:55,830 PROFESSOR: Generalizer. 223 00:10:55,830 --> 00:10:57,310 This one's an easy one to mess up 224 00:10:57,310 --> 00:10:58,914 if you don't understand the system. 225 00:10:58,914 --> 00:11:01,330 Because dropping, you think, oh, getting rid of something. 226 00:11:01,330 --> 00:11:02,163 That's specializing. 227 00:11:02,163 --> 00:11:04,869 But actually no, you're generalizing the entire area 228 00:11:04,869 --> 00:11:06,160 saying, we can forget about it. 229 00:11:06,160 --> 00:11:08,050 Because they're all good. 230 00:11:08,050 --> 00:11:12,061 All right, so given that, we are set to do this problem. 231 00:11:12,061 --> 00:11:13,685 Hey, we have a pretty reasonable amount 232 00:11:13,685 --> 00:11:16,880 of time considering that we don't have to do our sums. 233 00:11:16,880 --> 00:11:17,670 It's great. 234 00:11:17,670 --> 00:11:21,893 So our first example is sort of one 235 00:11:21,893 --> 00:11:24,143 of those stand reading lights that you sort of have it 236 00:11:24,143 --> 00:11:27,060 on a stand, and you can adjust the little light 237 00:11:27,060 --> 00:11:28,510 bulb up or down. 238 00:11:28,510 --> 00:11:35,190 It is an incandescent bulb. 239 00:11:35,190 --> 00:11:44,440 It has a height of 11 inches. 240 00:11:44,440 --> 00:11:46,835 And it's got a flat base. 241 00:11:51,650 --> 00:11:52,290 It's electric. 242 00:11:56,490 --> 00:11:59,090 It's got a shade. 243 00:11:59,090 --> 00:12:00,820 And it's a hit. 244 00:12:00,820 --> 00:12:03,780 It's a hit, a positive example, a plus. 245 00:12:03,780 --> 00:12:05,110 It's good. 246 00:12:05,110 --> 00:12:07,740 So right away, we know that we can only use 247 00:12:07,740 --> 00:12:10,630 the S's or the G's? 248 00:12:10,630 --> 00:12:11,250 The G's. 249 00:12:11,250 --> 00:12:12,250 We can only use the G's. 250 00:12:12,250 --> 00:12:13,041 Because it's a hit. 251 00:12:13,041 --> 00:12:14,740 So we're not going to be requiring. 252 00:12:14,740 --> 00:12:16,400 We're not to be forbidding today. 253 00:12:16,400 --> 00:12:18,530 We're going to be happy and using a generalizer. 254 00:12:18,530 --> 00:12:22,230 So I'll help do the first one. 255 00:12:22,230 --> 00:12:24,060 You guys will help me do the next one. 256 00:12:24,060 --> 00:12:26,300 Or someone can say, we didn't cover this, 257 00:12:26,300 --> 00:12:30,930 and I'll do the first one, and we'll stop. 258 00:12:30,930 --> 00:12:32,480 So we've got an incandescent light. 259 00:12:32,480 --> 00:12:34,590 That's the same as our model. 260 00:12:34,590 --> 00:12:36,400 We've got a height of 11 inches. 261 00:12:36,400 --> 00:12:38,660 Oh, our model only covers 24. 262 00:12:38,660 --> 00:12:40,910 Flat base-- that's the same as our model. 263 00:12:40,910 --> 00:12:42,770 Electricity-- same as our model. 264 00:12:42,770 --> 00:12:43,860 Shade-- same as our model. 265 00:12:43,860 --> 00:12:46,860 So I say the only difference is the height. 266 00:12:46,860 --> 00:12:55,030 I say we need to close the interval, close-interval. 267 00:12:55,030 --> 00:13:02,290 And I say that our model is the same as before, 268 00:13:02,290 --> 00:13:11,965 except for that height is an element of 11 to 24. 269 00:13:16,620 --> 00:13:20,610 Simple enough-- I picked the easy one for myself. 270 00:13:20,610 --> 00:13:22,900 Next one. 271 00:13:22,900 --> 00:13:42,430 So this one is going to be-- the picture here didn't turn out. 272 00:13:42,430 --> 00:13:45,760 But it's basically a lamp that doesn't have a shade. 273 00:13:45,760 --> 00:13:51,600 It just sort of has the light shining on you. 274 00:13:51,600 --> 00:13:53,290 So it is a positive example. 275 00:13:56,040 --> 00:14:10,549 It's incandescent, height equals 11.5, flat base, 276 00:14:10,549 --> 00:14:11,215 and electricity. 277 00:14:15,350 --> 00:14:20,315 All right, so first of all, which kind 278 00:14:20,315 --> 00:14:24,302 would we use, specializing or generalizing? 279 00:14:24,302 --> 00:14:25,260 AUDIENCE: Generalizing. 280 00:14:25,260 --> 00:14:28,170 PROFESSOR: Generalizing, and particularly, 281 00:14:28,170 --> 00:14:32,020 anyone want to take a look at our model, which 282 00:14:32,020 --> 00:14:37,760 is the starting model, except for the height can go 11 to 24? 283 00:14:37,760 --> 00:14:40,626 Which heuristic do we need to use here? 284 00:14:40,626 --> 00:14:42,060 AUDIENCE: Drop-link. 285 00:14:42,060 --> 00:14:45,390 PROFESSOR: All right, people are all saying drop-link. 286 00:14:45,390 --> 00:14:47,220 Yes, that's right. 287 00:14:47,220 --> 00:14:48,230 Someone said extend-set. 288 00:14:48,230 --> 00:14:50,940 Well, I suppose we could extend the set to shade or not shade. 289 00:14:50,940 --> 00:14:52,220 But that's everything. 290 00:14:52,220 --> 00:14:53,960 So drop-link is the answer. 291 00:14:53,960 --> 00:14:54,990 We can drop the shade. 292 00:14:54,990 --> 00:15:02,190 Shade or not, this is still pretty much a lamp. 293 00:15:02,190 --> 00:15:04,035 So that's correct. 294 00:15:04,035 --> 00:15:04,640 We drop-link. 295 00:15:10,010 --> 00:15:13,331 So the model-- all right, there's 296 00:15:13,331 --> 00:15:15,955 been enough changes to it that I think I'll write it out again. 297 00:15:15,955 --> 00:15:24,070 The model is incandescent, height 298 00:15:24,070 --> 00:15:36,835 equals an element from 11 to 24, and flat base, and electricity. 299 00:15:46,620 --> 00:15:51,335 Good, so question? 300 00:15:51,335 --> 00:15:57,275 AUDIENCE: So during this all, we're 301 00:15:57,275 --> 00:16:00,740 editing our model based on the order that the model is in, 302 00:16:00,740 --> 00:16:01,706 right? 303 00:16:01,706 --> 00:16:04,205 PROFESSOR: So the question is, we're editing the model based 304 00:16:04,205 --> 00:16:05,690 on the order of the examples? 305 00:16:05,690 --> 00:16:08,165 AUDIENCE: What happens if there's 306 00:16:08,165 --> 00:16:11,135 the same thing except with no shade or something like that 307 00:16:11,135 --> 00:16:12,432 afterwards? 308 00:16:12,432 --> 00:16:13,890 PROFESSOR: So the question is, what 309 00:16:13,890 --> 00:16:19,080 happens if you have the same thing except for that there's 310 00:16:19,080 --> 00:16:23,070 one that's the same with no shades and is a near miss 311 00:16:23,070 --> 00:16:25,585 or something so that you need to-- 312 00:16:25,585 --> 00:16:28,200 or I'm sorry, you're saying, what if there's one that has 313 00:16:28,200 --> 00:16:31,500 a shade now that's a near miss? 314 00:16:31,500 --> 00:16:36,950 And so it would probably try to forbid-link shades. 315 00:16:36,950 --> 00:16:39,080 That's sort of the question? 316 00:16:39,080 --> 00:16:41,920 You're right. 317 00:16:41,920 --> 00:16:43,710 That's an inconsistency in the data. 318 00:16:43,710 --> 00:16:45,550 The system is very fragile. 319 00:16:45,550 --> 00:16:49,490 It's very fragile to ordering in particular. 320 00:16:49,490 --> 00:16:53,690 And you'd be surprised how awesome 321 00:16:53,690 --> 00:16:59,370 this does considering that it was made in, like, the '60s. 322 00:16:59,370 --> 00:17:00,930 It's pretty impressive stuff. 323 00:17:00,930 --> 00:17:02,670 But it's old. 324 00:17:02,670 --> 00:17:05,761 We're not saying that this is a way that you should do 325 00:17:05,761 --> 00:17:07,010 all of your learning nowadays. 326 00:17:07,010 --> 00:17:10,530 Because it has some serious issues. 327 00:17:10,530 --> 00:17:13,940 It's old, but it's pretty damn good for what it did. 328 00:17:13,940 --> 00:17:17,200 A newer style of this kind of learning made by me 329 00:17:17,200 --> 00:17:19,510 and another friend who's now working at Google 330 00:17:19,510 --> 00:17:23,220 is lattice learning, which has its own share of issues, 331 00:17:23,220 --> 00:17:24,869 one of which is that since it's trying 332 00:17:24,869 --> 00:17:27,060 to act like a little kid, at first it 333 00:17:27,060 --> 00:17:29,690 tries to claim that everything is OK until you show it 334 00:17:29,690 --> 00:17:31,650 some good negative examples. 335 00:17:31,650 --> 00:17:34,330 But very particularly one way gets around 336 00:17:34,330 --> 00:17:36,610 this problem is it's not memoryless. 337 00:17:36,610 --> 00:17:38,600 It in fact stores all of the examples 338 00:17:38,600 --> 00:17:40,740 it's ever seen and compares and contrasts them 339 00:17:40,740 --> 00:17:44,150 to what it sees in the new example. 340 00:17:44,150 --> 00:17:46,137 And this allows it to say something 341 00:17:46,137 --> 00:17:48,470 like-- let's say you're trying to teach it what can fly. 342 00:17:48,470 --> 00:17:51,220 In lattice learning, you say, all right, well, a blue jay 343 00:17:51,220 --> 00:17:52,810 can fly. 344 00:17:52,810 --> 00:17:54,890 And then if you ask it, can a cow fly, sure, 345 00:17:54,890 --> 00:17:56,110 a cow jumped over the moon. 346 00:17:56,110 --> 00:17:57,140 Everything can fly. 347 00:17:57,140 --> 00:17:58,590 So that's its problem. 348 00:17:58,590 --> 00:18:00,780 But if you say, actually, that was a nursery rhyme, 349 00:18:00,780 --> 00:18:04,360 a cow cannot in fact fly, it will then say, 350 00:18:04,360 --> 00:18:05,980 only birds can fly. 351 00:18:05,980 --> 00:18:08,464 Because it remembers that the blue jay can fly. 352 00:18:08,464 --> 00:18:11,005 And then you can eventually say, well, actually bats can fly, 353 00:18:11,005 --> 00:18:12,190 a fruit bat can fly. 354 00:18:12,190 --> 00:18:16,967 Well, OK, so chiropterans can fly and birds can fly. 355 00:18:16,967 --> 00:18:17,550 But that's it. 356 00:18:17,550 --> 00:18:19,280 And you can give it some other examples. 357 00:18:19,280 --> 00:18:22,070 But there are other styles of learning. 358 00:18:22,070 --> 00:18:23,670 You're absolutely right. 359 00:18:23,670 --> 00:18:25,680 In fact, if there's a multiple choice asking you 360 00:18:25,680 --> 00:18:29,760 about the strengths and weaknesses of arch learning, 361 00:18:29,760 --> 00:18:33,000 one of the weaknesses-- it's very vulnerable to ordering 362 00:18:33,000 --> 00:18:35,530 of what you teach it. 363 00:18:35,530 --> 00:18:36,490 But think about it. 364 00:18:36,490 --> 00:18:38,720 If our first example was one that's 365 00:18:38,720 --> 00:18:41,000 exactly the same as this but with a shade, 366 00:18:41,000 --> 00:18:44,400 and it was a miss, the system would just yell at you. 367 00:18:44,400 --> 00:18:47,070 Like, you're kind of being an asshole to it to give it two 368 00:18:47,070 --> 00:18:48,530 things that are inconsistent. 369 00:18:48,530 --> 00:18:51,340 You might say in the real world that happens. 370 00:18:51,340 --> 00:18:56,770 Well, arch learning, not great with real world messy data. 371 00:18:56,770 --> 00:18:58,980 It just goes ballistic. 372 00:18:58,980 --> 00:18:59,750 It's very OCD. 373 00:18:59,750 --> 00:19:02,107 It wants everything to match up. 374 00:19:02,107 --> 00:19:03,940 If it gets two things that are inconsistent, 375 00:19:03,940 --> 00:19:05,402 it'll just yell at you. 376 00:19:05,402 --> 00:19:06,490 It's like, you're wrong. 377 00:19:06,490 --> 00:19:08,020 This can't be true. 378 00:19:08,020 --> 00:19:10,390 Because you said it's OK with the shade. 379 00:19:10,390 --> 00:19:11,970 So that's a very good question. 380 00:19:11,970 --> 00:19:13,083 Another question? 381 00:19:13,083 --> 00:19:15,055 AUDIENCE: So if you have a sample 382 00:19:15,055 --> 00:19:18,506 that had two different things for the model, 383 00:19:18,506 --> 00:19:19,985 and it's a near miss-- 384 00:19:19,985 --> 00:19:21,464 PROFESSOR: It's not a near miss. 385 00:19:21,464 --> 00:19:23,247 The question is, if you have something 386 00:19:23,247 --> 00:19:25,830 with two different things in the model, and it's a near miss-- 387 00:19:25,830 --> 00:19:28,350 and I'm sorry to cut you off. 388 00:19:28,350 --> 00:19:30,510 Based on timing, I'm just going to say, 389 00:19:30,510 --> 00:19:32,030 whatever you said afterwards doesn't 390 00:19:32,030 --> 00:19:34,590 apply due to the fact at that point it's not a near miss. 391 00:19:34,590 --> 00:19:35,820 It's very important. 392 00:19:35,820 --> 00:19:38,780 It's only a near miss if there's only one change. 393 00:19:38,780 --> 00:19:40,300 Otherwise, it's just a miss. 394 00:19:40,300 --> 00:19:42,670 And if it's a miss with more than one change, 395 00:19:42,670 --> 00:19:44,750 you were probably going to say, how do you know what to change? 396 00:19:44,750 --> 00:19:46,500 The answer is, you can't change anything. 397 00:19:46,500 --> 00:19:48,700 Because any of them could have been what was wrong. 398 00:19:48,700 --> 00:19:51,380 That's why for a hit, oh, it can have as many differences 399 00:19:51,380 --> 00:19:51,970 as you want. 400 00:19:51,970 --> 00:19:53,420 You'll just generalize them all. 401 00:19:53,420 --> 00:19:55,847 For a miss, it can only have one thing different. 402 00:19:55,847 --> 00:19:58,180 In fact, there might be one here that's not a near miss, 403 00:19:58,180 --> 00:19:59,870 and you just say, oh, we don't do anything. 404 00:19:59,870 --> 00:20:00,578 We'll keep going. 405 00:20:00,578 --> 00:20:02,070 We might see it if we have time. 406 00:20:02,070 --> 00:20:03,650 So question? 407 00:20:03,650 --> 00:20:05,725 AUDIENCE: Because ordering is an important factor 408 00:20:05,725 --> 00:20:10,343 for this bigger thing, and you encounter an item 409 00:20:10,343 --> 00:20:16,752 which has two discrepancies in values, 410 00:20:16,752 --> 00:20:20,203 but if the ordering were different such 411 00:20:20,203 --> 00:20:21,939 that that would be accounted for, 412 00:20:21,939 --> 00:20:24,147 do you remember if you turn back to that [INAUDIBLE]? 413 00:20:24,147 --> 00:20:25,872 PROFESSOR: So the question is, let's say 414 00:20:25,872 --> 00:20:29,100 you have a non-near miss. 415 00:20:29,100 --> 00:20:31,580 Can you hold onto it and use it later 416 00:20:31,580 --> 00:20:33,480 when it would be a near miss? 417 00:20:33,480 --> 00:20:35,800 The answer is, nope. 418 00:20:35,800 --> 00:20:38,030 If we're going here, for a system that 419 00:20:38,030 --> 00:20:41,580 doesn't remember things, you can put in some kludges. 420 00:20:41,580 --> 00:20:43,000 That definitely makes it smarter. 421 00:20:43,000 --> 00:20:44,220 It uses all its data. 422 00:20:44,220 --> 00:20:48,520 But Patrick in the '60s making this up strove for elegance. 423 00:20:48,520 --> 00:20:51,800 And the elegant solution is, let's just be memoryless 424 00:20:51,800 --> 00:20:55,050 completely, the idea being, well, little babies 425 00:20:55,050 --> 00:20:57,650 can't tell you every experience they had with playing 426 00:20:57,650 --> 00:20:59,140 with blocks with an arch. 427 00:20:59,140 --> 00:21:01,170 So let's remember nothing. 428 00:21:01,170 --> 00:21:04,230 With lattice learning, our idea is 429 00:21:04,230 --> 00:21:07,280 that people sort of somewhere in the subconscious 430 00:21:07,280 --> 00:21:10,342 maybe do store all the examples, or at least a lot 431 00:21:10,342 --> 00:21:11,800 more than you give them credit for. 432 00:21:11,800 --> 00:21:13,500 So why not store them? 433 00:21:13,500 --> 00:21:16,560 But with arch learning, it's, little babies can't tell you, 434 00:21:16,560 --> 00:21:20,540 oh yeah, there's that one time I played with a lamp, 435 00:21:20,540 --> 00:21:23,302 and it didn't have a shade, and so 436 00:21:23,302 --> 00:21:25,010 it wasn't a lamp, or something like that. 437 00:21:25,010 --> 00:21:26,718 They're not going to be able to store it. 438 00:21:26,718 --> 00:21:29,892 They're not going to be able to save it for later. 439 00:21:29,892 --> 00:21:31,350 Sort of in the style of Turing, who 440 00:21:31,350 --> 00:21:33,970 believed a human would be a teacher to a computer, 441 00:21:33,970 --> 00:21:36,780 arch learning really focuses on the fact 442 00:21:36,780 --> 00:21:39,320 that the human is a kind and good teacher who 443 00:21:39,320 --> 00:21:42,740 offers examples that are exactly appropriate at the time. 444 00:21:42,740 --> 00:21:45,500 And that puts a lot of pressure on you 445 00:21:45,500 --> 00:21:48,330 as the trainer of an arch learning system. 446 00:21:48,330 --> 00:21:49,230 Good questions, all. 447 00:21:49,230 --> 00:21:51,720 Let's continue working this guy out. 448 00:21:51,720 --> 00:21:59,910 So the next example is basically another one 449 00:21:59,910 --> 00:22:01,420 of these stand lamps. 450 00:22:01,420 --> 00:22:04,340 It's got a light shining down, which is fluorescent. 451 00:22:04,340 --> 00:22:05,165 It looks like this. 452 00:22:08,690 --> 00:22:10,760 It's fluorescent. 453 00:22:10,760 --> 00:22:22,250 Height equals 13 inches, flat base, electric, shade. 454 00:22:27,670 --> 00:22:30,970 And yeah, as I put here, it's a hit. 455 00:22:30,970 --> 00:22:34,250 So as you guys have told me countless times, 456 00:22:34,250 --> 00:22:35,820 so I believe you that you remember, 457 00:22:35,820 --> 00:22:39,840 we're going to generalize for the hit. 458 00:22:39,840 --> 00:22:44,630 So what should we do here? 459 00:22:49,660 --> 00:22:52,620 So some people are saying extend-set. 460 00:22:52,620 --> 00:22:53,560 AUDIENCE: Climb-tree. 461 00:22:53,560 --> 00:22:56,770 PROFESSOR: The people who are saying climb-tree are correct. 462 00:22:56,770 --> 00:22:59,010 Why do we climb-tree instead of extend-set? 463 00:22:59,010 --> 00:23:04,620 Because of the fact that it's a tree, not a set. 464 00:23:04,620 --> 00:23:08,060 And it is more parsimonious to climb 465 00:23:08,060 --> 00:23:11,750 the tree than it is to treat it as a set and extend the set. 466 00:23:11,750 --> 00:23:13,210 Sets don't have a hierarchy. 467 00:23:13,210 --> 00:23:16,560 It's just like red comma blue comma yellow, or shade comma 468 00:23:16,560 --> 00:23:17,950 not shade. 469 00:23:17,950 --> 00:23:19,830 Whenever you have a tree, it's a fact 470 00:23:19,830 --> 00:23:21,670 that it's more parsimonious to just climb it 471 00:23:21,670 --> 00:23:24,170 than it is to treat it as a set and extend it, 472 00:23:24,170 --> 00:23:26,290 even though you could do that, I suppose. 473 00:23:26,290 --> 00:23:29,320 And arch learning, it's parsimonious. 474 00:23:29,320 --> 00:23:30,050 It's elegant. 475 00:23:30,050 --> 00:23:31,000 It's simple. 476 00:23:31,000 --> 00:23:33,710 And the elegant thing to do-- just climb up. 477 00:23:33,710 --> 00:23:36,260 So yes, we're going to climb up the tree. 478 00:23:41,540 --> 00:23:51,810 So our model is the same as last time except that this time it 479 00:23:51,810 --> 00:24:02,200 has a light source instead of saying, incandescent. 480 00:24:02,200 --> 00:24:04,720 It just climbed up to light source, sure. 481 00:24:04,720 --> 00:24:07,880 All right, so the next one is pretty cool. 482 00:24:07,880 --> 00:24:11,130 It's another one of those fluorescent shine down lamps. 483 00:24:11,130 --> 00:24:14,060 But in space, it has a tripod with three legs. 484 00:24:14,060 --> 00:24:16,890 And those legs all have wheels. 485 00:24:16,890 --> 00:24:21,050 So it is also a hit. 486 00:24:21,050 --> 00:24:39,000 So it is fluorescent, height equals 14 inches, wheeled legs, 487 00:24:39,000 --> 00:24:41,950 and electric. 488 00:24:41,950 --> 00:24:45,900 So obviously we're generalizing. 489 00:24:45,900 --> 00:24:47,867 What heuristic will we use this time? 490 00:24:47,867 --> 00:24:49,120 AUDIENCE: Climb-tree. 491 00:24:49,120 --> 00:24:51,370 PROFESSOR: Climb-tree again, that's right. 492 00:24:51,370 --> 00:24:58,640 We're going to climb from flat base to base support itself. 493 00:24:58,640 --> 00:25:02,266 Because wheeled legs and flat base, 494 00:25:02,266 --> 00:25:04,200 base support is an ancestor of both. 495 00:25:07,450 --> 00:25:11,677 So our model is the same as before. 496 00:25:11,677 --> 00:25:13,260 But at this point, it's changed enough 497 00:25:13,260 --> 00:25:15,290 that I might as well write it out. 498 00:25:15,290 --> 00:25:30,510 So we'll say that it is light source, height between 11 499 00:25:30,510 --> 00:25:38,644 and 24 inches, base support, some sort of support, 500 00:25:38,644 --> 00:25:39,310 and electricity. 501 00:25:44,320 --> 00:25:47,360 Awesome, now, come on parity. 502 00:25:52,950 --> 00:25:56,890 Ah, no, back up, you. 503 00:25:56,890 --> 00:25:58,000 There we go. 504 00:25:58,000 --> 00:26:01,120 So that down there, don't get confused. 505 00:26:01,120 --> 00:26:04,000 That down there is our most recent model. 506 00:26:04,000 --> 00:26:16,520 We now have-- oh, oops, hehe. 507 00:26:16,520 --> 00:26:21,200 OK, I actually missed one of the examples. 508 00:26:21,200 --> 00:26:23,250 However, we were actually completely correct, 509 00:26:23,250 --> 00:26:26,850 which gives away quite a lot. 510 00:26:26,850 --> 00:26:36,070 We had a miss, incandescent bulb-- back in the old days 511 00:26:36,070 --> 00:26:38,740 when incandescent was the only kind of light we had. 512 00:26:41,330 --> 00:26:51,250 Height equals 60 inches, flat, back in the old days 513 00:26:51,250 --> 00:26:56,890 when flat was the only one we had, electric, 514 00:26:56,890 --> 00:26:59,470 back in the old days when we had electric. 515 00:26:59,470 --> 00:27:02,470 And we also have shade after we've 516 00:27:02,470 --> 00:27:03,660 dropped the link for shade. 517 00:27:03,660 --> 00:27:07,160 So we know that shade or not shade are both OK. 518 00:27:07,160 --> 00:27:17,190 So the question is, what do we do? 519 00:27:17,190 --> 00:27:20,335 First of all, do we specialize or generalize with this miss? 520 00:27:20,335 --> 00:27:21,210 AUDIENCE: Specialize. 521 00:27:21,210 --> 00:27:23,230 PROFESSOR: So obviously we're only to be able to require 522 00:27:23,230 --> 00:27:23,900 or forbid. 523 00:27:29,000 --> 00:27:34,890 So standardly, traditionally, the answer would be nothing. 524 00:27:34,890 --> 00:27:38,405 However, we accept it as sane the possibility 525 00:27:38,405 --> 00:27:41,780 that you might put a require onto the height. 526 00:27:41,780 --> 00:27:44,730 Some students did that, like say, oh 527 00:27:44,730 --> 00:27:47,810 gosh, that's a hard set, 11 to 24. 528 00:27:47,810 --> 00:27:50,180 But that's not what you would normally do. 529 00:27:50,180 --> 00:27:53,300 So just to show that we tried to understand that there might 530 00:27:53,300 --> 00:27:56,640 be multiple ideas in a weird situation, 531 00:27:56,640 --> 00:27:58,750 we did accept that that one time. 532 00:27:58,750 --> 00:28:01,080 But generally, there's not anything you have to do. 533 00:28:01,080 --> 00:28:06,690 This system at the time, and still now, said 11 to 24. 534 00:28:06,690 --> 00:28:09,750 This is spookily exactly the correct place 535 00:28:09,750 --> 00:28:11,700 where this example would go. 536 00:28:11,700 --> 00:28:14,190 But you don't have to do anything to the model. 537 00:28:14,190 --> 00:28:16,220 You don't have to use any heuristics. 538 00:28:16,220 --> 00:28:23,301 OK, so last step, -- the question is-- oh, question. 539 00:28:23,301 --> 00:28:28,768 AUDIENCE: So here, if we did it in this particular order, 540 00:28:28,768 --> 00:28:31,253 having it be the next [INAUDIBLE] 541 00:28:31,253 --> 00:28:34,235 in our example list, then it only 542 00:28:34,235 --> 00:28:37,714 differs by one from our existing model, right? 543 00:28:37,714 --> 00:28:39,482 PROFESSOR: No, it differs by a lot. 544 00:28:39,482 --> 00:28:41,690 It actually wouldn't be a near miss if we did it out. 545 00:28:41,690 --> 00:28:43,815 AUDIENCE: It would be a near miss if we did it out. 546 00:28:43,815 --> 00:28:47,720 PROFESSOR: No, incandescent is different than light source. 547 00:28:47,720 --> 00:28:49,510 Height is different than height. 548 00:28:49,510 --> 00:28:53,500 Flat is different than base support. 549 00:28:53,500 --> 00:28:58,422 So it would differ by a huge number, 550 00:28:58,422 --> 00:28:59,844 and it would not be a near miss. 551 00:28:59,844 --> 00:29:01,896 AUDIENCE: OK, then here's a question. 552 00:29:01,896 --> 00:29:06,693 Even though on the tree incandescent 553 00:29:06,693 --> 00:29:14,045 is a child branching off of light source, 554 00:29:14,045 --> 00:29:18,130 if we instead encountered-- because our current model says 555 00:29:18,130 --> 00:29:21,272 light source comma whatnot, if instead we 556 00:29:21,272 --> 00:29:26,180 got a negative thing that said, fluorescent, again, 557 00:29:26,180 --> 00:29:28,720 would that be considered not a near miss, 558 00:29:28,720 --> 00:29:31,116 because fluorescent differs from light source? 559 00:29:33,630 --> 00:29:35,770 PROFESSOR: So the question is, would it 560 00:29:35,770 --> 00:29:40,505 not be considered a near miss if you had, say, fluorescent 561 00:29:40,505 --> 00:29:42,380 rather than light source, because fluorescent 562 00:29:42,380 --> 00:29:44,460 differs from light source? 563 00:29:44,460 --> 00:29:47,798 So the answer there is-- 564 00:29:47,798 --> 00:29:49,401 AUDIENCE: Since it's memoryless, it 565 00:29:49,401 --> 00:29:52,380 doesn't know that fluorescent-- 566 00:29:52,380 --> 00:29:54,180 PROFESSOR: Well sure, it is memoryless. 567 00:29:54,180 --> 00:29:55,971 It doesn't know that fluorescent used to be 568 00:29:55,971 --> 00:29:57,300 a positive example before. 569 00:29:57,300 --> 00:29:59,620 Unfortunately, let's say we did have an example that 570 00:29:59,620 --> 00:30:03,600 was fluorescent, height is an element of 11 to 24, 571 00:30:03,600 --> 00:30:05,220 base support, electric. 572 00:30:05,220 --> 00:30:07,760 Actually, let's say we have one that was sodium vapor, 573 00:30:07,760 --> 00:30:11,050 height is 13. 574 00:30:11,050 --> 00:30:14,820 Sodium vapor, height is 13, base support-- 575 00:30:14,820 --> 00:30:18,032 for some reason it was called base support and electric. 576 00:30:18,032 --> 00:30:18,990 That's possible, right? 577 00:30:18,990 --> 00:30:20,930 We haven't seen a positive of sodium vapor. 578 00:30:20,930 --> 00:30:23,570 Let's say that sodium vapor was not allowed. 579 00:30:23,570 --> 00:30:26,570 If we don't say that that's a near miss and then forbid-link 580 00:30:26,570 --> 00:30:28,130 sodium vapor, how are we ever going 581 00:30:28,130 --> 00:30:30,090 to get rid of sodium vapor? 582 00:30:30,090 --> 00:30:33,000 Unfortunately, because of that, we 583 00:30:33,000 --> 00:30:35,770 lose some amount of expressiveness. 584 00:30:35,770 --> 00:30:41,572 Actually, what you're saying is very logical and makes sense, 585 00:30:41,572 --> 00:30:43,530 that you would want to have the ability to say, 586 00:30:43,530 --> 00:30:44,560 OK, this is a subset. 587 00:30:44,560 --> 00:30:45,567 This should be OK. 588 00:30:45,567 --> 00:30:47,150 But unfortunately, because of the fact 589 00:30:47,150 --> 00:30:48,566 that we climbed the tree even when 590 00:30:48,566 --> 00:30:50,990 we haven't seen a positive example for all things 591 00:30:50,990 --> 00:30:54,670 in the bottom on that tree, we actually lose that ability. 592 00:30:54,670 --> 00:30:56,180 It's a trade-off. 593 00:30:56,180 --> 00:30:59,870 It is most certainly a weakness of arch learning 594 00:30:59,870 --> 00:31:01,070 that you point out here. 595 00:31:01,070 --> 00:31:05,080 Lattice learning fixes that, but it has its own problems. 596 00:31:05,080 --> 00:31:08,190 So you're right, that you'd want to be able to do that. 597 00:31:08,190 --> 00:31:08,920 But you can't. 598 00:31:08,920 --> 00:31:10,000 Because it's memoryless. 599 00:31:10,000 --> 00:31:12,960 So last but not least, which model 600 00:31:12,960 --> 00:31:16,850 would we present in order to teach the system, right 601 00:31:16,850 --> 00:31:22,280 now from its final system, that a lamp requires a base support? 602 00:31:22,280 --> 00:31:25,590 So it would need to be require-link for the base 603 00:31:25,590 --> 00:31:26,310 support. 604 00:31:26,310 --> 00:31:28,710 So really fast-- incandescent bulb 605 00:31:28,710 --> 00:31:33,340 with a height of 6, a flat base, and electricity. 606 00:31:33,340 --> 00:31:34,320 Would that do it? 607 00:31:34,320 --> 00:31:35,820 Incandescent bulb with a height of 8 608 00:31:35,820 --> 00:31:39,792 and a battery, incandescent bulb with a height of 12, wire 609 00:31:39,792 --> 00:31:41,750 support, and electricity-- oh yeah, by the way, 610 00:31:41,750 --> 00:31:44,700 this will have to be a miss to teach a require. 611 00:31:44,700 --> 00:31:48,510 Fluorescent bulb, height of 21, clamp base, and electricity, 612 00:31:48,510 --> 00:31:50,170 or incandescent bulb, height of 12, 613 00:31:50,170 --> 00:31:52,620 flat bottomed legs, and electricity-- 614 00:31:52,620 --> 00:31:55,300 so did anyone pick that out over the fastness of me saying it? 615 00:31:55,300 --> 00:31:57,300 Because [INAUDIBLE] is going to want to come in. 616 00:31:57,300 --> 00:31:58,279 Question or answer? 617 00:31:58,279 --> 00:31:59,195 AUDIENCE: [INAUDIBLE]. 618 00:32:02,038 --> 00:32:03,787 PROFESSOR: It's going to have a miss. 619 00:32:03,787 --> 00:32:05,370 AUDIENCE: If it's on the wire support? 620 00:32:05,370 --> 00:32:06,330 PROFESSOR: Yes. 621 00:32:06,330 --> 00:32:08,480 So I know it was fast. 622 00:32:08,480 --> 00:32:10,224 You can look through quiz 2007 final. 623 00:32:10,224 --> 00:32:12,390 Because [INAUDIBLE] is going to come in in a minute. 624 00:32:12,390 --> 00:32:13,973 But it would be a miss on the one that 625 00:32:13,973 --> 00:32:17,275 was correct on everything except for it was a wire support. 626 00:32:17,275 --> 00:32:19,150 Then the question you have is, wait a minute, 627 00:32:19,150 --> 00:32:20,440 it says incandescent bulb. 628 00:32:20,440 --> 00:32:22,360 That is now a subset of light source. 629 00:32:22,360 --> 00:32:23,930 You have his question again. 630 00:32:23,930 --> 00:32:29,760 The answer to that is they must have decided here 631 00:32:29,760 --> 00:32:31,172 that that was OK. 632 00:32:31,172 --> 00:32:32,630 But on the other hand, how would it 633 00:32:32,630 --> 00:32:35,970 learn sodium vapor if that wasn't decided to be OK? 634 00:32:35,970 --> 00:32:38,570 That's going to have to be an implementation detail. 635 00:32:38,570 --> 00:32:42,590 So one thing you can do is try to treat subsets 636 00:32:42,590 --> 00:32:44,330 as OK in the logical way. 637 00:32:44,330 --> 00:32:46,200 Because questions may do that. 638 00:32:46,200 --> 00:32:47,960 And then I guess you've lost the ability 639 00:32:47,960 --> 00:32:50,730 to forbid, say, sodium vapor. 640 00:32:50,730 --> 00:32:53,110 So you have to lose one or the other. 641 00:32:53,110 --> 00:32:56,320 I think you're definitely OK to ask TAs or someone who's 642 00:32:56,320 --> 00:32:57,960 proctoring if a situation like that 643 00:32:57,960 --> 00:33:00,570 comes up, which implementation detail is chosen. 644 00:33:00,570 --> 00:33:02,820 Because you could choose either when you're running 645 00:33:02,820 --> 00:33:04,530 arch learning.