1 00:00:00,000 --> 00:00:07,724 Let's now see whether page rank has anything to do with how we remember things 2 00:00:07,724 --> 00:00:14,606 and recall our past memories. Quite clearly, search results ordered by 3 00:00:14,606 --> 00:00:22,233 page rank have proved rather intuitive. The success of Google itself is testament 4 00:00:22,233 --> 00:00:27,318 to this fact. Might the intuitive success of page rank 5 00:00:27,318 --> 00:00:32,780 provide any insight into how we recall things from memory. 6 00:00:34,160 --> 00:00:41,408 In an interesting article in psychological science in 2007 researchers tried to 7 00:00:41,408 --> 00:00:47,736 figure this out through experiment. What they did is they asked a lot of 8 00:00:47,736 --> 00:00:54,253 people to form associations between words. So people would be asked to name the first 9 00:00:54,253 --> 00:00:59,390 word that comes to mind when hearing a word like apple, for example. 10 00:00:59,390 --> 00:01:05,523 Some people might say orange, in which case a link would get created between the 11 00:01:05,523 --> 00:01:10,813 word apple and the word orange. If more people paired these two words, 12 00:01:10,813 --> 00:01:17,877 this particular link would get stronger. So in this manner, a semantic network was 13 00:01:17,877 --> 00:01:22,425 created. Such a semantic network is very similar to 14 00:01:22,425 --> 00:01:26,528 the network of web pages linked by hyperlinks. 15 00:01:26,528 --> 00:01:34,875 That was the argument in this paper. Next people were asked to form letter word 16 00:01:34,875 --> 00:01:40,640 associations, very similar to the example we discussed a while ago. 17 00:01:41,120 --> 00:01:47,639 So what's the first word that comes to mind when your given the letter A. 18 00:01:47,639 --> 00:01:50,140 Maybe apple. The question is. 19 00:01:51,160 --> 00:01:55,898 Could the semantic network be used to predict. 20 00:01:55,898 --> 00:02:01,151 Which word comes to mind given a particular letter. 21 00:02:01,151 --> 00:02:05,580 The relationship to search is quite simple. 22 00:02:06,100 --> 00:02:11,368 In the web, when it's searching for words, the semantic met... 23 00:02:11,368 --> 00:02:17,689 Network on a searching for letters. So the question is if your algorithm 24 00:02:17,689 --> 00:02:24,626 retrieves a whole bunch of words starting with the letter a, which one would it 25 00:02:24,626 --> 00:02:29,335 order first. Does page rank in this network, instead of 26 00:02:29,335 --> 00:02:36,647 this network now, provide any clue? Turns out that the human responses in 27 00:02:36,647 --> 00:02:44,724 experiment two could be predicted very well by page rank operating on the 28 00:02:44,724 --> 00:02:50,702 semantic network. Here is an example of the results of this 29 00:02:50,702 --> 00:02:56,120 experiment. What this says is that. 30 00:02:56,120 --> 00:03:02,902 75 percent of the human responses, ended up being within the top twenty percent 31 00:03:02,902 --> 00:03:07,747 dial, of the ordering provided by the Paydrag. 32 00:03:07,747 --> 00:03:14,234 So clearly paydrag was mimicking. The human response fared well. 33 00:03:14,234 --> 00:03:21,666 The question is does this mean anything. Is human memory similar to Google's 34 00:03:21,666 --> 00:03:27,339 massive index with page reckoning on it, what do you think? 35 00:03:27,339 --> 00:03:35,921 Actually the answer is probably'no'. For example, most of us are poor at 36 00:03:35,921 --> 00:03:41,162 remembering facts. Such as when was Napoleon's defeat at 37 00:03:41,162 --> 00:03:46,043 Waterloo? Google on the other hand is excellent at 38 00:03:46,043 --> 00:03:55,251 retrieving exactly the correct fact. Secondly, we often need context to augment 39 00:03:55,251 --> 00:04:00,591 our recall. For example, we might not recognize a work 40 00:04:00,591 --> 00:04:08,059 colleague when we see them in a mall. We simply can't place them in that 41 00:04:08,059 --> 00:04:14,700 context. Further, memories are linked in time. 42 00:04:15,920 --> 00:04:22,499 For example, what one did the first thing in the morning, and thereafter, and 43 00:04:22,499 --> 00:04:29,580 thereafter, are all linked. Another example might be an incident from 44 00:04:29,580 --> 00:04:33,081 one's first day at school, college or work. 45 00:04:33,081 --> 00:04:39,416 When you put this two contexts together, the incident and the day at work or 46 00:04:39,416 --> 00:04:46,280 college get linked. Finally, memories are fuzzy. 47 00:04:46,800 --> 00:04:51,856 For example, can you recall every item in your room? 48 00:04:51,856 --> 00:04:57,309 Probably not. The picture that you form in your mind is 49 00:04:57,309 --> 00:05:03,000 very spars. On the other hand. 50 00:05:03,720 --> 00:05:10,143 Memories can be triggered by very sparse matches, such as a mere smell, or a mere 51 00:05:10,143 --> 00:05:14,640 face, can recall an entire incident or an entire picture. 52 00:05:17,680 --> 00:05:26,035 This does appear to suggest that the way memory works is quite different from how 53 00:05:26,035 --> 00:05:31,348 search works. Later in this course we'll touch upon a 54 00:05:31,348 --> 00:05:37,840 number of models which appear to be much closer to how human memory works. 55 00:05:38,140 --> 00:05:45,326 And quite different from search. And we will also see how the AI 56 00:05:45,326 --> 00:05:54,084 applications that exploit big data get better and better when one uses memory 57 00:05:54,084 --> 00:05:58,800 like mechanisms as oppose to search alone.