1 00:00:04,400 --> 00:00:07,800 Speaker 1: Welcome to Tech Stuff, a production from I Heart Radio. 2 00:00:11,840 --> 00:00:14,040 Speaker 1: Hey there, and welcome to tech Stuff. I'm your host, 3 00:00:14,120 --> 00:00:16,759 Speaker 1: Jonathan Strickland. I'm an executive producer with iHeart Radio. And 4 00:00:16,800 --> 00:00:20,479 Speaker 1: how the tech are you? Alright? Well, I'm still on vacation. 5 00:00:20,600 --> 00:00:24,040 Speaker 1: I'll be coming back soon, so tomorrow you should expect 6 00:00:24,040 --> 00:00:27,920 Speaker 1: a brand new episode unless something goes wrong while I'm 7 00:00:27,920 --> 00:00:31,000 Speaker 1: trying to get back. Hopefully nothing like that happens, And 8 00:00:31,440 --> 00:00:33,920 Speaker 1: so we thought we'd have a little rerun. This episode 9 00:00:33,960 --> 00:00:38,080 Speaker 1: originally published in April one, so just last year. It 10 00:00:38,200 --> 00:00:42,400 Speaker 1: is titled machine Learning one oh one. And I wanted 11 00:00:42,440 --> 00:00:45,360 Speaker 1: to do this one because, as always, we hear a 12 00:00:45,440 --> 00:00:48,919 Speaker 1: lot about artificial intelligence and machine learning in the news 13 00:00:48,960 --> 00:00:54,040 Speaker 1: and in media, and often those topics get a little confusing. 14 00:00:54,120 --> 00:00:59,520 Speaker 1: They can come across more broad than some people intend, 15 00:00:59,840 --> 00:01:04,640 Speaker 1: or or they can be somewhat misguided in their interpretations. 16 00:01:04,680 --> 00:01:06,480 Speaker 1: So I thought it would be useful to have a 17 00:01:06,480 --> 00:01:10,399 Speaker 1: little refresher course on machine learning and artificial intelligence to 18 00:01:10,400 --> 00:01:13,920 Speaker 1: hope you enjoy, uh and I will be back at 19 00:01:13,920 --> 00:01:20,200 Speaker 1: the end. Back in nineteen eighties, six comedy science fiction 20 00:01:20,280 --> 00:01:24,200 Speaker 1: film that I saw in the theater about a robot, 21 00:01:24,600 --> 00:01:28,560 Speaker 1: the game sentience and becomes a total goofball what it will. 22 00:01:28,600 --> 00:01:31,039 Speaker 1: It hit theaters in eighties six and it was called 23 00:01:31,640 --> 00:01:36,080 Speaker 1: Short Circuit. The movie starred Steve Gutenberg, Ali Sheety, and 24 00:01:36,360 --> 00:01:40,440 Speaker 1: lamentably a white actor named Fisher Stevens playing a non 25 00:01:40,480 --> 00:01:44,520 Speaker 1: white character, someone who is Indian. I should add that's 26 00:01:44,520 --> 00:01:48,240 Speaker 1: not Steven's fault. I mean, he auditioned to be in 27 00:01:48,240 --> 00:01:50,640 Speaker 1: a movie and he got a gig. He didn't cast 28 00:01:50,720 --> 00:01:53,080 Speaker 1: himself in the film, and he has since talked about 29 00:01:53,120 --> 00:01:56,720 Speaker 1: his experiences, realizing the problems with a white man playing 30 00:01:56,720 --> 00:01:59,760 Speaker 1: a non white character, but setting aside all the problematic 31 00:01:59,760 --> 00:02:04,080 Speaker 1: white washing, the movie showed this robot, who in the 32 00:02:04,080 --> 00:02:08,440 Speaker 1: course of the film names itself Johnny five learning. It 33 00:02:08,560 --> 00:02:11,560 Speaker 1: learns about the world around it, it learns about people, 34 00:02:12,080 --> 00:02:16,960 Speaker 1: It learns about human concepts like humor and emotion, and 35 00:02:17,000 --> 00:02:20,919 Speaker 1: the general idea was pretty cute. Now, the nifty thing 36 00:02:21,040 --> 00:02:25,680 Speaker 1: is machines actually can learn. In fact, machine learning is 37 00:02:25,720 --> 00:02:29,320 Speaker 1: a really important field of study these days, complete with 38 00:02:29,360 --> 00:02:32,959 Speaker 1: its own challenges and risks. I've talked about machine learning 39 00:02:33,240 --> 00:02:35,040 Speaker 1: a few times in the past, but I figured we 40 00:02:35,040 --> 00:02:38,400 Speaker 1: could do a deeper dive to understand what machine learning 41 00:02:38,560 --> 00:02:42,160 Speaker 1: is what it isn't how people are leveraging machine learning 42 00:02:42,240 --> 00:02:45,919 Speaker 1: and why I said that it does come with risks, 43 00:02:45,919 --> 00:02:53,280 Speaker 1: So let's learn about machines learning. It will be impossible 44 00:02:53,360 --> 00:02:56,800 Speaker 1: to talk about machine learning without also talking about artificial 45 00:02:56,840 --> 00:03:01,840 Speaker 1: intelligence or AI. And this term artificial intelligence is a 46 00:03:02,000 --> 00:03:06,520 Speaker 1: real doozy. It trips people up, even people who have 47 00:03:06,680 --> 00:03:11,560 Speaker 1: dedicated their lives to researching and developing artificial intelligence. You 48 00:03:11,600 --> 00:03:16,200 Speaker 1: can get two experts in AI talking about AI and 49 00:03:16,240 --> 00:03:19,000 Speaker 1: find out that because they have slightly different takes on 50 00:03:19,160 --> 00:03:24,680 Speaker 1: what AI is, there are some communication issues. It's not 51 00:03:24,760 --> 00:03:27,480 Speaker 1: as simple as red versus blue would have you think 52 00:03:28,080 --> 00:03:33,680 Speaker 1: what does the A stand for? So when you really 53 00:03:34,120 --> 00:03:36,440 Speaker 1: boil it down, it comes out as as no big 54 00:03:36,480 --> 00:03:39,480 Speaker 1: surprise that there's a lot of ambiguity here. After all, 55 00:03:39,840 --> 00:03:44,880 Speaker 1: how would you define intelligence just intelligence, not artificial intelligence, 56 00:03:45,240 --> 00:03:49,880 Speaker 1: just intelligence? Well? Would it be the ability to learn, 57 00:03:50,240 --> 00:03:54,480 Speaker 1: that is, to acquire skills and knowledge? Or is it 58 00:03:54,560 --> 00:03:57,920 Speaker 1: the application of learning? Is it problems solving? Is it 59 00:03:58,400 --> 00:04:01,680 Speaker 1: being able to think ahead and make plans in order 60 00:04:01,720 --> 00:04:05,960 Speaker 1: to achieve a specific goal? Is it the ability to 61 00:04:06,240 --> 00:04:09,800 Speaker 1: examine a problem and deconstructed in order to figure out 62 00:04:09,840 --> 00:04:12,840 Speaker 1: the best solution. A more specific version of problem solving. 63 00:04:13,480 --> 00:04:18,800 Speaker 1: Is it the ability to recognize, understand, and navigate emotional scenarios? Now, 64 00:04:18,920 --> 00:04:24,200 Speaker 1: arguably it's all of these things and more. We all 65 00:04:24,240 --> 00:04:28,640 Speaker 1: have kind of an intuitive grasp on what intelligence is, 66 00:04:29,560 --> 00:04:34,240 Speaker 1: but defining it in a simple way tends to feel 67 00:04:34,240 --> 00:04:37,680 Speaker 1: reductive and it leaves out a lot of important details. 68 00:04:37,720 --> 00:04:43,440 Speaker 1: So if defining just general intelligence is hard, it stands 69 00:04:43,440 --> 00:04:46,720 Speaker 1: for a reason that defining artificial intelligence is also a 70 00:04:46,760 --> 00:04:50,600 Speaker 1: tough job. Heck, even coming up with a number of 71 00:04:50,640 --> 00:04:54,680 Speaker 1: different types of a I is tricky. And if you 72 00:04:54,720 --> 00:04:59,159 Speaker 1: don't believe me, just google the phrase different types of 73 00:04:59,279 --> 00:05:03,400 Speaker 1: artificial intelligence. Never mind, you don't. You don't really actually 74 00:05:03,440 --> 00:05:06,119 Speaker 1: have to do that. I already did it, though, Feel 75 00:05:06,160 --> 00:05:08,640 Speaker 1: free to do it yourself and check my work if 76 00:05:08,680 --> 00:05:13,360 Speaker 1: you like. When I googled that phrase different types of AI, 77 00:05:13,520 --> 00:05:16,400 Speaker 1: some of the top results included a blog post on 78 00:05:16,600 --> 00:05:21,480 Speaker 1: BMC software titled four types of artificial Intelligence. But then 79 00:05:21,520 --> 00:05:24,279 Speaker 1: there was also an article on code bots that was 80 00:05:24,320 --> 00:05:27,680 Speaker 1: titled what are the three types of AI? And then 81 00:05:27,720 --> 00:05:31,440 Speaker 1: there was an article from Forbes titled seven types of 82 00:05:31,520 --> 00:05:35,600 Speaker 1: artificial intelligence. See, we can't even agree on how many 83 00:05:35,720 --> 00:05:39,200 Speaker 1: versions of a EI there are because defining a I 84 00:05:40,080 --> 00:05:44,040 Speaker 1: is really hard. It largely depends upon how you view 85 00:05:44,200 --> 00:05:46,720 Speaker 1: AI and then how you break it down into different 86 00:05:46,760 --> 00:05:51,599 Speaker 1: realms of intelligence. Now we could go super high level, 87 00:05:51,920 --> 00:05:55,159 Speaker 1: because a classic way to look at AI is strong 88 00:05:55,760 --> 00:06:02,240 Speaker 1: versus weak Artificial intelligence stro on AI UH sometimes called 89 00:06:02,440 --> 00:06:08,760 Speaker 1: artificial general intelligence, would be a machine that processes information 90 00:06:09,040 --> 00:06:13,400 Speaker 1: and at least appears to have some form of consciousness 91 00:06:13,480 --> 00:06:17,440 Speaker 1: and self awareness and the ability to both have experiences 92 00:06:17,480 --> 00:06:21,359 Speaker 1: and to be aware that it is having experiences. It 93 00:06:21,440 --> 00:06:25,599 Speaker 1: might even feel emotion, though maybe not emotions that we 94 00:06:25,680 --> 00:06:29,480 Speaker 1: could easily identify or sympathize with. So this would be 95 00:06:30,080 --> 00:06:33,840 Speaker 1: the kind of machine that would think in a way 96 00:06:34,000 --> 00:06:36,840 Speaker 1: similar to humans. It would be able to sense its 97 00:06:36,920 --> 00:06:40,640 Speaker 1: environment and not just react, but really process what is 98 00:06:40,680 --> 00:06:43,839 Speaker 1: going on and build and understanding. It's the type of 99 00:06:43,880 --> 00:06:46,880 Speaker 1: AI that we see a lot in science fiction. A's 100 00:06:46,920 --> 00:06:50,000 Speaker 1: the type of AI of Johnny five from Short Circuit 101 00:06:50,480 --> 00:06:53,719 Speaker 1: or how from two thousand one, or the droids in 102 00:06:53,800 --> 00:06:57,880 Speaker 1: Star Wars. It's also a type of artificial intelligence that 103 00:06:57,960 --> 00:07:01,480 Speaker 1: we have yet to actually achieve in the real world. 104 00:07:02,000 --> 00:07:06,520 Speaker 1: So then what is week AI. Well, you could say 105 00:07:06,520 --> 00:07:10,120 Speaker 1: it's everything else, or you could say it's the building 106 00:07:10,160 --> 00:07:16,080 Speaker 1: blocks that maybe collectively will lead to strong AI week. 107 00:07:16,240 --> 00:07:21,160 Speaker 1: AI involves processes that allow machines to complete tasks, So, 108 00:07:21,240 --> 00:07:25,640 Speaker 1: for example, image recognition software could fall into this category. 109 00:07:25,960 --> 00:07:29,640 Speaker 1: Once upon a time, in order to search photos effectively, 110 00:07:30,160 --> 00:07:34,680 Speaker 1: you needed to actually add meta data like tags to 111 00:07:34,880 --> 00:07:40,040 Speaker 1: those photos. So, for example, I might tag pictures of 112 00:07:40,080 --> 00:07:44,080 Speaker 1: my dog with the meta tag dog, and then if 113 00:07:44,080 --> 00:07:46,920 Speaker 1: I wanted to see photos of my pooch, then I 114 00:07:46,920 --> 00:07:49,920 Speaker 1: would pull up my photo app and search the term dog, 115 00:07:50,440 --> 00:07:52,920 Speaker 1: and all the photos that I had tagged with the 116 00:07:52,960 --> 00:07:55,320 Speaker 1: word dog would show up. But if I had failed 117 00:07:55,480 --> 00:07:59,520 Speaker 1: to tag some pictures of my dog, those pictures wouldn't 118 00:07:59,560 --> 00:08:02,200 Speaker 1: pop up in search because the computer program wasn't actually 119 00:08:02,280 --> 00:08:05,200 Speaker 1: looking for dogs in my photos. It was just looking 120 00:08:05,200 --> 00:08:08,720 Speaker 1: for photos that had that particular meta tag attached to it. 121 00:08:09,480 --> 00:08:12,320 Speaker 1: But now we've reached a point where at least some 122 00:08:12,400 --> 00:08:16,720 Speaker 1: photo apps are using image recognition to analyze photos, and 123 00:08:16,760 --> 00:08:20,120 Speaker 1: these will return results that the algorithm has identified as 124 00:08:20,160 --> 00:08:23,560 Speaker 1: having a reasonable chance of meeting your search query. So 125 00:08:23,840 --> 00:08:26,280 Speaker 1: if I used an app like that and I put 126 00:08:26,320 --> 00:08:29,480 Speaker 1: in dog as my search term, it could pull up 127 00:08:29,480 --> 00:08:32,640 Speaker 1: photos that had no meta tags attached to them at all. 128 00:08:33,120 --> 00:08:36,520 Speaker 1: Because the search is relying on image recognition. Now, this 129 00:08:36,640 --> 00:08:40,680 Speaker 1: also means that if the image recognition algorithm isn't very good, 130 00:08:40,720 --> 00:08:42,960 Speaker 1: I could get some images that don't have a dog 131 00:08:43,000 --> 00:08:46,480 Speaker 1: in them at all, or it might miss other images 132 00:08:46,520 --> 00:08:48,960 Speaker 1: that have my dog in them. But my point is 133 00:08:49,000 --> 00:08:52,080 Speaker 1: that the ability to identify whether or not a dog 134 00:08:52,160 --> 00:08:56,000 Speaker 1: is in a particular photo represents a kind of weak 135 00:08:56,160 --> 00:09:01,560 Speaker 1: artificial intelligence. You wouldn't say that the photo search tool 136 00:09:01,720 --> 00:09:05,560 Speaker 1: possesses humanlike intelligence, because really it only does one thing. 137 00:09:06,120 --> 00:09:10,200 Speaker 1: It's analyzing photos and looks for matches to specific search queries, 138 00:09:10,559 --> 00:09:14,360 Speaker 1: but it can't do anything outside of that use case. However, 139 00:09:14,400 --> 00:09:17,080 Speaker 1: that's just one little example. There are all sorts of 140 00:09:17,080 --> 00:09:23,120 Speaker 1: other ones, like voice recognition, environmental sensing, course plotting, that 141 00:09:23,200 --> 00:09:25,760 Speaker 1: kind of thing, and in some circles, as we get 142 00:09:25,800 --> 00:09:30,320 Speaker 1: better at making machines and systems that can do these things, 143 00:09:31,120 --> 00:09:34,120 Speaker 1: those elements seem to kind of drift away from the 144 00:09:34,200 --> 00:09:38,960 Speaker 1: ongoing conversation about artificial intelligence. A guy named Larry Tessler, 145 00:09:39,160 --> 00:09:41,320 Speaker 1: who was a computer scientist who worked at lots of 146 00:09:41,320 --> 00:09:46,320 Speaker 1: really important places like Xerox, Park and Amazon and Apple, 147 00:09:46,840 --> 00:09:52,200 Speaker 1: he once observed, quote, intelligence is whatever machines haven't done yet. 148 00:09:52,559 --> 00:09:55,920 Speaker 1: End quote. So his point was that the reason that 149 00:09:56,000 --> 00:09:58,560 Speaker 1: AI is really hard to talk about is that the 150 00:09:58,600 --> 00:10:04,160 Speaker 1: goal post for why actually is artificial intelligence is constantly moving. 151 00:10:06,000 --> 00:10:08,560 Speaker 1: Now this pretty much mirrors how we think about things 152 00:10:08,600 --> 00:10:13,439 Speaker 1: like consciousness. Lots of people study consciousness, and the general 153 00:10:13,480 --> 00:10:16,040 Speaker 1: sense I get is that it's a lot easier for 154 00:10:16,080 --> 00:10:20,160 Speaker 1: people to talk about what isn't consciousness rather than what 155 00:10:20,520 --> 00:10:25,080 Speaker 1: consciousness actually is. And it seems like artificial intelligence is 156 00:10:25,120 --> 00:10:28,640 Speaker 1: in a similar place, which really isn't that big of 157 00:10:28,640 --> 00:10:33,640 Speaker 1: a surprise as we closely associate intelligence with consciousness. Now 158 00:10:33,679 --> 00:10:36,959 Speaker 1: this leads us to why there are so many different 159 00:10:37,040 --> 00:10:41,000 Speaker 1: takes on how many types of AI there are. It 160 00:10:41,000 --> 00:10:45,400 Speaker 1: all depends on how you classify different disciplines in artificial intelligence, 161 00:10:45,720 --> 00:10:48,920 Speaker 1: and over time, a lot of disciplines that were previously 162 00:10:49,080 --> 00:10:53,480 Speaker 1: distinct from AI have sort of converged into becoming part 163 00:10:53,600 --> 00:10:56,840 Speaker 1: of the AI discussion. Machine learning, as it turns out, 164 00:10:57,360 --> 00:11:00,880 Speaker 1: was part of the AI discussion, branch off from it, 165 00:11:01,120 --> 00:11:05,480 Speaker 1: and then rejoined the AI discussion years later. So I 166 00:11:05,520 --> 00:11:08,000 Speaker 1: am not going to go down all the different approaches 167 00:11:08,040 --> 00:11:10,640 Speaker 1: to classification because I don't know that they would be 168 00:11:10,760 --> 00:11:13,840 Speaker 1: that valuable to us. They would really just illustrate that 169 00:11:13,880 --> 00:11:16,280 Speaker 1: there are a lot of different ways to look at 170 00:11:16,320 --> 00:11:21,560 Speaker 1: the subject. So if you ever find yourself in a 171 00:11:21,600 --> 00:11:25,760 Speaker 1: conversation about AI, it might be a good idea to 172 00:11:25,800 --> 00:11:29,400 Speaker 1: set a few ground rules as to what everyone means 173 00:11:29,840 --> 00:11:33,320 Speaker 1: when they use the term artificial intelligence. That can help 174 00:11:33,559 --> 00:11:38,360 Speaker 1: with expectations and understanding. Or you could just run for 175 00:11:38,400 --> 00:11:41,560 Speaker 1: the nearest exit, which is what people tend to do 176 00:11:41,640 --> 00:11:48,120 Speaker 1: whenever I start talking about it anyway. What about machine learning, Well, 177 00:11:48,200 --> 00:11:51,240 Speaker 1: from one perspective, you could say machine learning is a 178 00:11:51,360 --> 00:11:55,520 Speaker 1: sub discipline of artificial intelligence, although like I said, it 179 00:11:55,600 --> 00:11:59,679 Speaker 1: hasn't always been viewed as such. I think most people 180 00:11:59,760 --> 00:12:02,880 Speaker 1: would say that the ability to learn that is to 181 00:12:03,200 --> 00:12:07,520 Speaker 1: take information and experience and then have some form of 182 00:12:07,640 --> 00:12:11,120 Speaker 1: understanding of those things so that you can apply that 183 00:12:11,200 --> 00:12:15,200 Speaker 1: to future tasks, potentially getting better over time. I would 184 00:12:15,240 --> 00:12:18,880 Speaker 1: say most people would call that part of intelligence. But 185 00:12:19,480 --> 00:12:21,400 Speaker 1: you could also be a bit more wishy washy and 186 00:12:21,440 --> 00:12:25,000 Speaker 1: say it's related to, you know, artificial intelligence, as opposed 187 00:12:25,040 --> 00:12:28,080 Speaker 1: to being part of AI, since the definition of AI 188 00:12:28,240 --> 00:12:33,320 Speaker 1: is let's say, fluid. Either way of classifying machine learning works. 189 00:12:33,360 --> 00:12:37,960 Speaker 1: As far as I'm concerned, machine learning boils down to 190 00:12:38,000 --> 00:12:41,520 Speaker 1: the idea of creating a system that can learn as 191 00:12:41,559 --> 00:12:45,360 Speaker 1: it performs a task. It can learn what works and 192 00:12:45,520 --> 00:12:49,280 Speaker 1: more importantly, what does not work. You may have heard 193 00:12:49,360 --> 00:12:51,920 Speaker 1: that we learn a lot more from our mistakes than 194 00:12:51,960 --> 00:12:56,320 Speaker 1: we do from our successes, which there's pretty much true 195 00:12:56,360 --> 00:13:00,480 Speaker 1: in my experience. When something goes wrong, it's usually, but 196 00:13:00,800 --> 00:13:05,640 Speaker 1: not always, possible to trace the event or events that 197 00:13:05,800 --> 00:13:09,920 Speaker 1: led to the failure. You can identify decisions that we're 198 00:13:09,960 --> 00:13:13,400 Speaker 1: probably the wrong ones or that led to a bad outcome, 199 00:13:14,120 --> 00:13:17,640 Speaker 1: But if you have a success, it's hard to figure 200 00:13:17,679 --> 00:13:22,600 Speaker 1: out which decisions were key to that successful outcome. Did 201 00:13:22,640 --> 00:13:25,199 Speaker 1: your decision at step two set you on the right path, 202 00:13:25,600 --> 00:13:28,720 Speaker 1: or was your choice at step three so good that 203 00:13:28,800 --> 00:13:31,840 Speaker 1: it helped correct a mistake that you made it step two. 204 00:13:32,360 --> 00:13:35,319 Speaker 1: But a good approach to machine learning involves a system 205 00:13:35,480 --> 00:13:38,560 Speaker 1: that can adjust things on its own to reduce mistakes 206 00:13:38,960 --> 00:13:41,839 Speaker 1: and increase the success rate. And another way of putting 207 00:13:41,880 --> 00:13:44,959 Speaker 1: it is that instead of programming a system to arrive 208 00:13:45,000 --> 00:13:48,920 Speaker 1: at a specific outcome, you are training the system to 209 00:13:49,080 --> 00:13:52,480 Speaker 1: learn how to do it by itself. And that sounds 210 00:13:52,480 --> 00:13:55,240 Speaker 1: a bit magical when you put it that way, doesn't it? 211 00:13:55,800 --> 00:13:59,040 Speaker 1: It sounds like someone just took a computer and showed 212 00:13:59,040 --> 00:14:01,840 Speaker 1: it pictures of cat and then expected the computer to 213 00:14:01,880 --> 00:14:05,200 Speaker 1: know what a cat was. And this actually does mirror 214 00:14:05,360 --> 00:14:09,000 Speaker 1: an actual project that really did do that, But I'm 215 00:14:09,080 --> 00:14:13,320 Speaker 1: leaving out some big important information in the middle. Now, 216 00:14:13,840 --> 00:14:17,679 Speaker 1: one big step is that computers and machines can't just 217 00:14:17,800 --> 00:14:20,880 Speaker 1: magically learn by default. People first had to come up 218 00:14:20,920 --> 00:14:24,240 Speaker 1: with a methodology that allows machines to go through the 219 00:14:24,280 --> 00:14:27,960 Speaker 1: process of completing a task, then making adjustments to the 220 00:14:28,080 --> 00:14:32,920 Speaker 1: process of doing that task, which would then improve future results. 221 00:14:33,440 --> 00:14:36,960 Speaker 1: We have to lay the groundwork in architecture and theory 222 00:14:37,160 --> 00:14:41,160 Speaker 1: and algorithms. We have to build the logical pathways that 223 00:14:41,200 --> 00:14:44,760 Speaker 1: computers can follow in order for them to learn. A 224 00:14:44,800 --> 00:14:49,680 Speaker 1: lot of machine learning revolves around patterns and pattern recognition. 225 00:14:50,080 --> 00:14:52,400 Speaker 1: So what do I mean by patterns? Well, I mean 226 00:14:52,560 --> 00:14:58,680 Speaker 1: some form of regularity and predictability. Machine learning models analyze 227 00:14:58,720 --> 00:15:03,040 Speaker 1: patterns and attempt to draw conclusions based on those patterns. 228 00:15:03,760 --> 00:15:07,120 Speaker 1: This in itself is tricky stuff. So why is that? Well, 229 00:15:07,160 --> 00:15:11,720 Speaker 1: it's because sometimes we might think there's a pattern when 230 00:15:11,720 --> 00:15:17,040 Speaker 1: in reality there is not. We humans are pretty good 231 00:15:17,320 --> 00:15:22,160 Speaker 1: at recognizing patterns, which makes sense. It's a survival mechanism. 232 00:15:22,200 --> 00:15:25,280 Speaker 1: If you were to look at tall grass and you 233 00:15:25,480 --> 00:15:28,800 Speaker 1: see patterns that suggest the presence of a predator like 234 00:15:29,000 --> 00:15:33,200 Speaker 1: a tiger, well you would know that danger is nearby, 235 00:15:33,240 --> 00:15:36,120 Speaker 1: and you would have the opportunity to do something about 236 00:15:36,160 --> 00:15:40,200 Speaker 1: that to help your chances of survival. If, however, you 237 00:15:40,320 --> 00:15:44,400 Speaker 1: remained blissfully unaware of the danger, you'd be far more 238 00:15:44,480 --> 00:15:48,240 Speaker 1: likely to fall prey to that hungry tiger. So recognizing 239 00:15:48,320 --> 00:15:51,280 Speaker 1: patterns is one of the abilities that gave humans a 240 00:15:51,360 --> 00:15:55,080 Speaker 1: chance to live another day, and, from an evolutionary standpoint, 241 00:15:55,120 --> 00:16:00,240 Speaker 1: a chance to make more humans. But sometimes we wins 242 00:16:00,280 --> 00:16:05,360 Speaker 1: will perceive a pattern where none actually exists. A simple 243 00:16:05,360 --> 00:16:08,760 Speaker 1: example of this is the fun exercise of laying on 244 00:16:08,800 --> 00:16:13,000 Speaker 1: your back outside, looking up at the clouds and saying, 245 00:16:13,040 --> 00:16:16,600 Speaker 1: what does that cloud remind you? Of? The shapes of clouds, 246 00:16:16,680 --> 00:16:21,120 Speaker 1: which have no significance and are the product of environmental factors, 247 00:16:21,560 --> 00:16:25,040 Speaker 1: can seem to suggest patterns to us. We might see 248 00:16:25,040 --> 00:16:28,840 Speaker 1: a dog, or a car or a face, but we 249 00:16:28,920 --> 00:16:32,880 Speaker 1: know that what we're really seeing with just the appearance 250 00:16:33,000 --> 00:16:35,400 Speaker 1: of a pattern, it's it's not evidence of a pattern 251 00:16:35,480 --> 00:16:40,000 Speaker 1: actually being there. It's noise, not signal. But it could 252 00:16:40,040 --> 00:16:44,200 Speaker 1: be misinterpreted as signal. Well, it turns out that in 253 00:16:44,280 --> 00:16:47,440 Speaker 1: machine learning applications this is also an issue. I'll talk 254 00:16:47,480 --> 00:16:50,520 Speaker 1: about it more towards the end of this episode. Computers 255 00:16:50,560 --> 00:16:55,400 Speaker 1: can sometimes misinterpret data and determine something represents a pattern 256 00:16:55,480 --> 00:16:58,760 Speaker 1: when it really doesn't. When that happens, a system relying 257 00:16:58,760 --> 00:17:02,760 Speaker 1: on machine learning can whose false positives, and the consequences 258 00:17:02,800 --> 00:17:06,159 Speaker 1: can sometimes be funny, like hey, this image recognition software 259 00:17:06,200 --> 00:17:09,119 Speaker 1: thinks this coffee mug is actually a kidney cat. Or 260 00:17:09,160 --> 00:17:12,640 Speaker 1: they can be really serious and potentially harmful. Hey, this 261 00:17:12,800 --> 00:17:17,120 Speaker 1: facial recognition software has misidentified a person, marking them as, say, 262 00:17:17,200 --> 00:17:20,240 Speaker 1: a person of interest in a criminal case. And it's 263 00:17:20,240 --> 00:17:23,280 Speaker 1: all because this facial recognition software isn't very good at 264 00:17:23,320 --> 00:17:29,040 Speaker 1: differentiating people of color. That's a real problem that really happens. Now, 265 00:17:29,040 --> 00:17:31,800 Speaker 1: when we come back, I'll give a little overview of 266 00:17:31,880 --> 00:17:35,080 Speaker 1: the evolution of machine learning. But before we do that, 267 00:17:35,720 --> 00:17:46,560 Speaker 1: let's take a quick break to talk about the history 268 00:17:46,760 --> 00:17:50,080 Speaker 1: of machine learning. We first have to look back much 269 00:17:50,560 --> 00:17:54,080 Speaker 1: much earlier, long before the era of computers, and talk 270 00:17:54,160 --> 00:17:58,480 Speaker 1: about how thinkers like Thomas Bayes thought about the act 271 00:17:58,720 --> 00:18:03,400 Speaker 1: of problem solving. Bays was born way back in two, 272 00:18:03,440 --> 00:18:06,320 Speaker 1: so quite a bit before we were thinking about machine learning, 273 00:18:06,720 --> 00:18:11,400 Speaker 1: but he was interested in problem solving for problems involving probabilities, 274 00:18:11,840 --> 00:18:16,480 Speaker 1: and specifically the relationship between different probabilities. I think it's 275 00:18:16,520 --> 00:18:19,440 Speaker 1: easier to talk about if I give you an example. 276 00:18:20,040 --> 00:18:22,520 Speaker 1: So let's make a silly one, all right, So let's 277 00:18:22,560 --> 00:18:27,200 Speaker 1: say we got ourselves a plucky podcaster. Hey there, everybody, 278 00:18:27,440 --> 00:18:31,960 Speaker 1: It's Jonathan Strickland, and it's Tuesday as I record this, 279 00:18:32,160 --> 00:18:35,040 Speaker 1: And because of who I am, you know who this 280 00:18:35,119 --> 00:18:39,800 Speaker 1: podcaster is. And because it's Tuesday, there is a chance 281 00:18:39,960 --> 00:18:42,840 Speaker 1: I am wearing a they might be Giants T shirt. 282 00:18:43,320 --> 00:18:48,080 Speaker 1: And we also know that if this podcaster is wearing 283 00:18:48,280 --> 00:18:51,800 Speaker 1: a they might be Giants T shirt on a Tuesday, 284 00:18:52,000 --> 00:18:55,639 Speaker 1: there's a sixty chance that I'm going to end up 285 00:18:55,640 --> 00:18:59,720 Speaker 1: wearing pajamas on Wednesday. But we also know that if 286 00:18:59,760 --> 00:19:04,280 Speaker 1: I did not where they might be Giant's shirt on Tuesday, 287 00:19:04,480 --> 00:19:08,359 Speaker 1: and remember there's a six chance I didn't, then we 288 00:19:08,440 --> 00:19:10,879 Speaker 1: know there's an eighty percent chance I'm going to be 289 00:19:10,920 --> 00:19:15,359 Speaker 1: wearing pajamas on Wednesday. Will Bays worked out a way 290 00:19:15,440 --> 00:19:20,240 Speaker 1: that described the sort of probability relationship between different discrete 291 00:19:20,320 --> 00:19:24,320 Speaker 1: events and using his reasoning, you can work forward or 292 00:19:24,440 --> 00:19:29,000 Speaker 1: backward based on probabilities. Theys would describe wearing a they 293 00:19:29,080 --> 00:19:32,240 Speaker 1: Might be Giant shirt on Tuesday as one event and 294 00:19:32,280 --> 00:19:36,360 Speaker 1: wearing pajamas on Wednesday as a separate event, and then 295 00:19:36,400 --> 00:19:39,399 Speaker 1: describe the two not only determining how likely it is 296 00:19:39,440 --> 00:19:43,760 Speaker 1: I'll wear pajamas on Wednesday, but if we start with 297 00:19:43,880 --> 00:19:46,439 Speaker 1: the later event, in other words, that we start with 298 00:19:46,480 --> 00:19:50,199 Speaker 1: the fact that it's Wednesday and I'm wearing pajamas, we 299 00:19:50,240 --> 00:19:55,360 Speaker 1: could work out how likely it was that yesterday, on Tuesday, 300 00:19:55,440 --> 00:19:58,719 Speaker 1: I was wearing they Might be Giants shirt. That was 301 00:19:58,800 --> 00:20:01,240 Speaker 1: his his contribution, that you can work this in either 302 00:20:01,359 --> 00:20:04,919 Speaker 1: direction if you know these different variables. Now, Bay has 303 00:20:05,000 --> 00:20:08,480 Speaker 1: never published his thoughts, but rather send an essay explaining 304 00:20:08,520 --> 00:20:11,280 Speaker 1: it to a friend of his, who then made sure 305 00:20:11,359 --> 00:20:13,879 Speaker 1: that the work was published. After Bays had passed away, 306 00:20:14,160 --> 00:20:18,280 Speaker 1: and a few decades later, Pierre Simon Laplace would take 307 00:20:18,359 --> 00:20:20,800 Speaker 1: this work that Bays had done and flesh it out 308 00:20:20,840 --> 00:20:25,520 Speaker 1: into an actual formal theorem. It's an important example of 309 00:20:25,600 --> 00:20:30,080 Speaker 1: conditional probability, and a lot of what machine learning is 310 00:20:30,880 --> 00:20:36,000 Speaker 1: really boiled down to is dealing with different probabilities, not certainties, which, 311 00:20:36,040 --> 00:20:37,399 Speaker 1: when you get down to it, is what most of 312 00:20:37,440 --> 00:20:39,360 Speaker 1: us are doing most of the time. Right. We make 313 00:20:39,400 --> 00:20:44,720 Speaker 1: decisions based on at least perceived probabilities. Sometimes these decisions 314 00:20:44,800 --> 00:20:48,200 Speaker 1: might feel like they're a coin flip situation, that any 315 00:20:48,320 --> 00:20:51,639 Speaker 1: choice is equally likely to precipitate a good outcome or 316 00:20:51,680 --> 00:20:54,640 Speaker 1: a bad outcome. Other Times we might make a choice 317 00:20:54,680 --> 00:20:58,240 Speaker 1: because we feel the probabilities are stacked favorably one way 318 00:20:58,320 --> 00:21:02,080 Speaker 1: over another. Sometimes we will make a choice to back 319 00:21:02,240 --> 00:21:07,720 Speaker 1: the least probable outcome, because well, humans are not always superrational. 320 00:21:07,760 --> 00:21:10,960 Speaker 1: In hex sometimes the long shot does pay off, so 321 00:21:11,920 --> 00:21:16,120 Speaker 1: that keeps Vegas in business. Bayes' theorem is just one 322 00:21:16,160 --> 00:21:19,639 Speaker 1: example of ways that mathematicians and philosophers figured out ways 323 00:21:19,680 --> 00:21:24,639 Speaker 1: to mathematically express problem solving and decision making, And a 324 00:21:24,680 --> 00:21:26,879 Speaker 1: lot of this was figuring out if there were a 325 00:21:26,920 --> 00:21:29,880 Speaker 1: way to boil down things that most of us approached 326 00:21:29,960 --> 00:21:34,359 Speaker 1: through intuition and experience. So it's kind of neat, and 327 00:21:34,480 --> 00:21:37,080 Speaker 1: also the more you look into it, the more likely 328 00:21:37,119 --> 00:21:39,879 Speaker 1: you might find it's little spooky, because it's weird to 329 00:21:39,880 --> 00:21:43,960 Speaker 1: consider that our approaches to making choices and solving problems 330 00:21:44,240 --> 00:21:50,440 Speaker 1: can be reduced down to mathematical expressions. But let's leave 331 00:21:50,520 --> 00:21:53,840 Speaker 1: the potential existential crises alone for now, shall we. So 332 00:21:53,960 --> 00:21:57,280 Speaker 1: moving on, we have another smarty pants we need to 333 00:21:57,320 --> 00:22:03,240 Speaker 1: talk about Andre Markov, mathematician. In the early twentie century. 334 00:22:03,320 --> 00:22:07,159 Speaker 1: He began studying the nature of certain random processes that 335 00:22:07,240 --> 00:22:10,040 Speaker 1: follow a particular type of rule, which we now call 336 00:22:10,240 --> 00:22:15,400 Speaker 1: the Markov property. That rule says that for this particular process, 337 00:22:15,440 --> 00:22:19,640 Speaker 1: the next stage of the process only depends upon the 338 00:22:19,680 --> 00:22:23,960 Speaker 1: current stage, but not any stages that came before then. 339 00:22:24,400 --> 00:22:28,480 Speaker 1: So let's take my ridiculous T shirt example and let's 340 00:22:28,480 --> 00:22:30,880 Speaker 1: build it out a little bit further. Let's say that 341 00:22:31,000 --> 00:22:33,680 Speaker 1: I've got three T shirts to my name. One of 342 00:22:33,720 --> 00:22:36,320 Speaker 1: them is that they might be Giant's shirt. One is 343 00:22:36,359 --> 00:22:40,040 Speaker 1: a plain blue T shirt, and the third is a 344 00:22:40,119 --> 00:22:43,159 Speaker 1: shirt that has the tech Stuff logo on it. And 345 00:22:43,960 --> 00:22:48,879 Speaker 1: it's based off of long observation that you've determined these 346 00:22:48,920 --> 00:22:53,040 Speaker 1: following facts. If I am wearing that they might be 347 00:22:53,119 --> 00:22:57,639 Speaker 1: Giant's shirt today, I definitely will not wear it tomorrow. 348 00:22:58,040 --> 00:23:01,199 Speaker 1: But there's a fifty fifty shot I'll wear either the 349 00:23:01,200 --> 00:23:05,000 Speaker 1: blue shirt or the tech Stuff shirt. Now, if I'm 350 00:23:05,040 --> 00:23:09,040 Speaker 1: wearing the blue shirt today, there's a ten chance I'm 351 00:23:09,040 --> 00:23:12,520 Speaker 1: going to wear the same blue shirt tomorrow. Don't worry, 352 00:23:12,800 --> 00:23:16,840 Speaker 1: I'll wash it first. There's a sixty chance that I'll 353 00:23:16,880 --> 00:23:19,560 Speaker 1: wear the tech Stuff shirt, and there's a thirty percent 354 00:23:19,640 --> 00:23:22,879 Speaker 1: chance I'll wear the they Might Be Giant shirt. But 355 00:23:23,800 --> 00:23:26,439 Speaker 1: if I'm wearing the tech stuff shirt today, there's a 356 00:23:26,440 --> 00:23:29,639 Speaker 1: seventy chance I'll wear it again tomorrow because I like 357 00:23:29,720 --> 00:23:33,000 Speaker 1: to promote myself. But there's a thirty percent chance I'll 358 00:23:33,000 --> 00:23:35,439 Speaker 1: wear the they Might be Giant shirt, and there is 359 00:23:35,520 --> 00:23:38,160 Speaker 1: no chance that I'm going to wear the blue one 360 00:23:38,520 --> 00:23:42,760 Speaker 1: in this case. So those are our various scenarios. Right 361 00:23:43,080 --> 00:23:47,800 Speaker 1: which shirt I will wear tomorrow depends only upon which 362 00:23:47,880 --> 00:23:51,359 Speaker 1: shirt I am wearing today. What I wore yesterday has 363 00:23:51,400 --> 00:23:55,359 Speaker 1: no bearing on the outcome for tomorrow, So today is 364 00:23:55,400 --> 00:23:59,119 Speaker 1: all that matters. And depending on which shirt I wear, 365 00:23:59,560 --> 00:24:02,879 Speaker 1: you can make some probability predictions for tomorrow. So we 366 00:24:02,920 --> 00:24:05,840 Speaker 1: can actually use this approach to figure out the probability 367 00:24:05,920 --> 00:24:09,080 Speaker 1: that I might wear the tech Stuff shirts, say ten 368 00:24:09,200 --> 00:24:12,359 Speaker 1: days in a row, since there's a better than even 369 00:24:12,480 --> 00:24:16,000 Speaker 1: chance that if I'm wearing tech Stuff today, I'll end 370 00:24:16,080 --> 00:24:19,280 Speaker 1: up wearing it again tomorrow, and if I wear it tomorrow, 371 00:24:19,480 --> 00:24:22,119 Speaker 1: then there's a better than fift chance that I'm going 372 00:24:22,160 --> 00:24:25,840 Speaker 1: to wear it the following day. But at some point 373 00:24:25,960 --> 00:24:29,119 Speaker 1: you're going to see that the odds are starting to 374 00:24:29,200 --> 00:24:33,600 Speaker 1: be against you, for you know, increasingly long strings of 375 00:24:33,640 --> 00:24:37,240 Speaker 1: wearing the tech stuff shirt. Anyway, Markov chains would become 376 00:24:37,320 --> 00:24:40,159 Speaker 1: one of the types of processes that machine learning models 377 00:24:40,200 --> 00:24:43,760 Speaker 1: would incorporate, with some models looking at the current state 378 00:24:43,880 --> 00:24:46,879 Speaker 1: of a given process and then make predictions on what 379 00:24:47,160 --> 00:24:50,679 Speaker 1: the next state will be with no need to look 380 00:24:50,800 --> 00:24:56,720 Speaker 1: back at the previous decisions. The Markov chain is memory less. 381 00:24:57,640 --> 00:25:00,960 Speaker 1: Now that's just a couple of the mathematicians whose work 382 00:25:01,080 --> 00:25:05,399 Speaker 1: underlies elements of machine learning. There's also structure we need 383 00:25:05,440 --> 00:25:09,800 Speaker 1: to talk about. In a man named Donald Hebb wrote 384 00:25:09,800 --> 00:25:13,520 Speaker 1: a book titled The Organization of Behavior, and in that book, 385 00:25:14,080 --> 00:25:18,560 Speaker 1: Hebb gave hypothesis on how neurons, that is, how how 386 00:25:18,640 --> 00:25:22,840 Speaker 1: brain cells interact with one another. His ideas included the 387 00:25:22,840 --> 00:25:27,119 Speaker 1: notion that if two neurons interact with one another regularly, 388 00:25:27,640 --> 00:25:31,000 Speaker 1: that is, if one fires, that the second one is 389 00:25:31,040 --> 00:25:35,280 Speaker 1: also likely to fire. They end up forming a tighter 390 00:25:35,320 --> 00:25:40,399 Speaker 1: communicative relationship with each other. Not long after his expression 391 00:25:40,400 --> 00:25:44,199 Speaker 1: of this hypothesis. Computer scientists began to think of a 392 00:25:44,200 --> 00:25:48,480 Speaker 1: potential way to do this artificially, with machines creating the 393 00:25:48,560 --> 00:25:54,440 Speaker 1: equivalent of artificial neurons. The relative strength in relationship between 394 00:25:54,720 --> 00:25:59,560 Speaker 1: artificial neurons is something we describe by Wait, that's going 395 00:25:59,600 --> 00:26:02,919 Speaker 1: to be an important part of machine learning. WIT. By 396 00:26:02,920 --> 00:26:06,120 Speaker 1: the way, is W E I G H T, as 397 00:26:06,160 --> 00:26:11,439 Speaker 1: in this relationship is weighted more heavily than that relationship. 398 00:26:12,200 --> 00:26:16,080 Speaker 1: In the early nineteen fifties, an IBM researcher named Arthur 399 00:26:16,280 --> 00:26:19,919 Speaker 1: Samuel created a program designed to win at checkers. The 400 00:26:19,960 --> 00:26:22,920 Speaker 1: program would do a quick analysis of where pieces were 401 00:26:23,160 --> 00:26:27,120 Speaker 1: on a checkerboard and whose move it was, and then 402 00:26:27,200 --> 00:26:30,520 Speaker 1: calculate the chances of each side winning the game based 403 00:26:30,560 --> 00:26:33,280 Speaker 1: on those positions. And it did this with a mini 404 00:26:33,320 --> 00:26:38,000 Speaker 1: max approach. Alright, so checkers is a two player turn 405 00:26:38,080 --> 00:26:41,160 Speaker 1: based game. Player one makes a move, then player two 406 00:26:41,160 --> 00:26:43,560 Speaker 1: can make a move. There are a finite number of 407 00:26:43,600 --> 00:26:47,439 Speaker 1: moves that can be made, a finite number of possibilities, 408 00:26:47,480 --> 00:26:51,760 Speaker 1: though admittedly it's a pretty good number of possibilities. But 409 00:26:51,880 --> 00:26:54,159 Speaker 1: let's say a game has been going on for a 410 00:26:54,200 --> 00:26:57,080 Speaker 1: few moves, and you've got your two sides you've got 411 00:26:57,080 --> 00:26:59,639 Speaker 1: the red checkers over on player one side and the 412 00:26:59,720 --> 00:27:02,639 Speaker 1: black checkers for a player to Let's say it's player 413 00:27:02,720 --> 00:27:06,080 Speaker 1: one's move. For the purposes of this example, will say 414 00:27:06,080 --> 00:27:08,880 Speaker 1: that player one really just has one piece that they 415 00:27:09,520 --> 00:27:12,800 Speaker 1: can actually move on this turn, and it can move 416 00:27:12,840 --> 00:27:17,160 Speaker 1: into one of two open spaces. So player one has 417 00:27:17,200 --> 00:27:20,280 Speaker 1: to make a choice. After that choice, it's going to 418 00:27:20,320 --> 00:27:23,720 Speaker 1: be player two's turn, so we can create a decision 419 00:27:23,800 --> 00:27:28,399 Speaker 1: treat illustrating the possible choices and the possible outcomes of 420 00:27:28,440 --> 00:27:32,440 Speaker 1: those choices. These choices are the children of the starting 421 00:27:32,440 --> 00:27:35,880 Speaker 1: position for player one, so player one's starting position has 422 00:27:36,119 --> 00:27:39,960 Speaker 1: two children. Player too will have their own choices to 423 00:27:40,040 --> 00:27:43,760 Speaker 1: make after that decision has been made, but those choices 424 00:27:43,760 --> 00:27:48,400 Speaker 1: are going to depend upon whatever move player one ultimately takes. 425 00:27:48,440 --> 00:27:51,720 Speaker 1: So we can extend out our decision treat showing the 426 00:27:51,800 --> 00:27:56,120 Speaker 1: branching possible moves that player Too might make, And these 427 00:27:56,160 --> 00:28:00,639 Speaker 1: are the children of the two possible outcomes of our choice. 428 00:28:01,160 --> 00:28:04,960 Speaker 1: After player two's turn, it's player ones turn again, which 429 00:28:04,960 --> 00:28:08,760 Speaker 1: means we need to branch those decisions out even further. 430 00:28:09,359 --> 00:28:12,000 Speaker 1: And this is all before player one has even made 431 00:28:12,240 --> 00:28:16,840 Speaker 1: that first choice. We're just evaluating possibilities. At some point, 432 00:28:17,080 --> 00:28:19,560 Speaker 1: either when we have plotted far enough out that we 433 00:28:19,640 --> 00:28:23,760 Speaker 1: know all possible outcomes of the game, or we're just 434 00:28:24,240 --> 00:28:26,919 Speaker 1: reaching a point where it would be unmanageable for us 435 00:28:26,920 --> 00:28:29,879 Speaker 1: to go any further, we need to actually analyze what 436 00:28:29,960 --> 00:28:35,639 Speaker 1: our options are. The endpoints represent either a win, a loss, 437 00:28:35,920 --> 00:28:39,720 Speaker 1: or a draw for player one, or, if we haven't 438 00:28:39,760 --> 00:28:41,959 Speaker 1: extended out the tree all the way to the end 439 00:28:41,960 --> 00:28:45,040 Speaker 1: of the game, at least a change in advantage, whether 440 00:28:45,240 --> 00:28:47,840 Speaker 1: it would be in player one's advantage to make that 441 00:28:47,920 --> 00:28:52,680 Speaker 1: move or disadvantage. We could actually assign numerical values to 442 00:28:52,760 --> 00:28:56,760 Speaker 1: each end point, with positive values representing an advantage for 443 00:28:56,840 --> 00:29:00,120 Speaker 1: player one and a negative value representing an advantage for 444 00:29:00,120 --> 00:29:03,040 Speaker 1: a player too, and once we do that, we can 445 00:29:03,080 --> 00:29:06,840 Speaker 1: see which pathways tend to lead to better outcomes for 446 00:29:07,040 --> 00:29:11,360 Speaker 1: player one. We work backward through the decision tree, so 447 00:29:11,680 --> 00:29:15,120 Speaker 1: on all the decisions that end in an advantage for 448 00:29:15,200 --> 00:29:18,080 Speaker 1: player one, we can say this is the choice that 449 00:29:18,120 --> 00:29:21,640 Speaker 1: player one would take. But then we know that a 450 00:29:21,640 --> 00:29:25,200 Speaker 1: player to player two is always going to choose whichever 451 00:29:25,320 --> 00:29:29,360 Speaker 1: choice has the greatest advantage for that player, so we 452 00:29:29,440 --> 00:29:32,400 Speaker 1: have to actually take that into account as we're working backward, 453 00:29:33,400 --> 00:29:36,840 Speaker 1: and this is how we can finally get to the 454 00:29:36,840 --> 00:29:39,120 Speaker 1: point where we decide which move we're going to make. 455 00:29:39,200 --> 00:29:42,760 Speaker 1: Because these decisions as you go backward up the tree, 456 00:29:43,560 --> 00:29:47,480 Speaker 1: they ultimately inform you which of those two choices is 457 00:29:47,520 --> 00:29:51,280 Speaker 1: going to give you the best result. Those values, well, 458 00:29:51,440 --> 00:29:54,280 Speaker 1: those are weights. So for player one, the goal is 459 00:29:54,320 --> 00:29:57,640 Speaker 1: to pick the path that has the highest positive value. 460 00:29:58,040 --> 00:30:00,680 Speaker 1: For player too, it's to pick the path that has 461 00:30:00,720 --> 00:30:04,320 Speaker 1: the lowest possible value or the highest negative value if 462 00:30:04,360 --> 00:30:06,800 Speaker 1: you prefer so. In other words, player one might be 463 00:30:06,840 --> 00:30:09,960 Speaker 1: thinking something like, if I move to Spot A, my 464 00:30:10,080 --> 00:30:13,160 Speaker 1: chance of winning this game, But if I moved to 465 00:30:13,160 --> 00:30:17,960 Speaker 1: Spot B, it's only so. Of course, those percentages will 466 00:30:18,000 --> 00:30:19,960 Speaker 1: also depend on what player two is going to do 467 00:30:20,000 --> 00:30:22,880 Speaker 1: in response. Some moves that player two might do could 468 00:30:23,000 --> 00:30:26,520 Speaker 1: end up guaranteeing a win for player one. This is 469 00:30:26,560 --> 00:30:30,080 Speaker 1: the mini max approach, and there's an algorithm that guides it. 470 00:30:30,080 --> 00:30:33,800 Speaker 1: It depends upon the current position within a game and 471 00:30:33,920 --> 00:30:36,680 Speaker 1: how many moves or how much depth it has to 472 00:30:36,720 --> 00:30:40,240 Speaker 1: take into account, and for which player is it actually 473 00:30:40,280 --> 00:30:44,440 Speaker 1: helping out. What happens is if player one does this 474 00:30:44,480 --> 00:30:48,720 Speaker 1: evaluation and finds that both options are negative, well, then 475 00:30:49,560 --> 00:30:51,760 Speaker 1: this is something that happens in games, right, Sometimes you 476 00:30:51,840 --> 00:30:54,880 Speaker 1: find out there is no good move, like any move 477 00:30:54,920 --> 00:30:56,880 Speaker 1: you make is going to be a losing move. Well, 478 00:30:56,920 --> 00:30:59,040 Speaker 1: the only option at that point is to choose the 479 00:30:59,160 --> 00:31:01,920 Speaker 1: least bad had one, so it would be whatever the 480 00:31:01,960 --> 00:31:06,360 Speaker 1: smallest negative value choice was. Our Next big development that 481 00:31:06,400 --> 00:31:10,720 Speaker 1: I need to mention is Frank Rosenblatt's artificial neural network 482 00:31:10,840 --> 00:31:15,480 Speaker 1: called Perceptron. Its purpose was to recognize shapes and patterns, 483 00:31:15,840 --> 00:31:18,400 Speaker 1: and it was originally going to be its own machine 484 00:31:18,520 --> 00:31:23,040 Speaker 1: like actual hardware, but the first incarnation of Perceptron would 485 00:31:23,080 --> 00:31:26,000 Speaker 1: actually be in the form of software rather than hardware. 486 00:31:26,320 --> 00:31:29,880 Speaker 1: There was a purpose built Perceptron later, but the original 487 00:31:29,880 --> 00:31:34,360 Speaker 1: one was software. Despite some early excitement, the Perceptron proved 488 00:31:34,400 --> 00:31:37,960 Speaker 1: to be somewhat limited in its capabilities, and interest in 489 00:31:38,040 --> 00:31:41,320 Speaker 1: artificial neural networks died down for a while as a result. 490 00:31:42,320 --> 00:31:45,080 Speaker 1: In a way, you could kind of compare this to 491 00:31:45,280 --> 00:31:48,320 Speaker 1: some other technologies that got a big hype cycle and 492 00:31:48,360 --> 00:31:52,440 Speaker 1: then later deflated. Virtual reality is the one I always 493 00:31:52,480 --> 00:31:54,920 Speaker 1: go with. Back in the nineteen nineties, the world was 494 00:31:55,000 --> 00:32:00,000 Speaker 1: really hyped for virtual reality. People had incredibly unrealistic x 495 00:32:00,000 --> 00:32:03,320 Speaker 1: spectations for what VR actually meant and what it could do, 496 00:32:04,000 --> 00:32:06,720 Speaker 1: and when it turned out the VR wasn't nearly as 497 00:32:06,720 --> 00:32:10,600 Speaker 1: sophisticated as people were imagining, a lot of enthusiasm dropped 498 00:32:10,640 --> 00:32:15,320 Speaker 1: out for the entire field, and with that dropped funding 499 00:32:15,440 --> 00:32:18,480 Speaker 1: and support, and as a result, development and VR hit 500 00:32:18,520 --> 00:32:21,560 Speaker 1: a real wall, with only a fraction of the people 501 00:32:21,600 --> 00:32:24,640 Speaker 1: who had been working in the field sticking around, and 502 00:32:25,200 --> 00:32:27,600 Speaker 1: they had to scramble just to find funding to keep 503 00:32:27,640 --> 00:32:30,680 Speaker 1: their projects going. So VR was effectively put on the 504 00:32:30,720 --> 00:32:34,520 Speaker 1: shelf and wouldn't make much progress for nearly twenty years. Well. 505 00:32:34,640 --> 00:32:39,120 Speaker 1: Artificial neural networks had a very similar issue, but other 506 00:32:39,160 --> 00:32:43,680 Speaker 1: computer scientists eventually found ways to design artificial neural networks. 507 00:32:43,960 --> 00:32:47,240 Speaker 1: They could do some pretty amazing things if they had 508 00:32:47,280 --> 00:32:50,680 Speaker 1: access to enough data. When we come back, i'll talk 509 00:32:50,720 --> 00:32:53,560 Speaker 1: a little bit more about that and what it all means, 510 00:32:53,600 --> 00:33:04,800 Speaker 1: but first let's take another quick break. So we left 511 00:33:04,840 --> 00:33:07,800 Speaker 1: off with the AI field going into hibernation for a 512 00:33:07,840 --> 00:33:11,720 Speaker 1: little bit. Theory and mathematics were bumping up against the 513 00:33:11,760 --> 00:33:15,280 Speaker 1: limitations of technology, which wasn't quite at the level to 514 00:33:15,840 --> 00:33:19,040 Speaker 1: put all that theory to the test. Plus there needed 515 00:33:19,040 --> 00:33:22,000 Speaker 1: to be some tweaks to the approaches, but those came 516 00:33:22,120 --> 00:33:26,200 Speaker 1: with time and more mathematicians found new ways to create 517 00:33:26,280 --> 00:33:30,720 Speaker 1: artificial neural networks capable of stuff like pattern recognition and learning. 518 00:33:31,320 --> 00:33:36,400 Speaker 1: So let's imagine another decision tree. We've got our starting position. 519 00:33:37,160 --> 00:33:40,000 Speaker 1: This is probably where we put some input. We would 520 00:33:40,120 --> 00:33:44,200 Speaker 1: feed data into a system, and let's say from that 521 00:33:44,360 --> 00:33:47,600 Speaker 1: starting position, we have a process that's going to transform 522 00:33:47,720 --> 00:33:52,080 Speaker 1: that input into one of two possible ways. So we've 523 00:33:52,120 --> 00:33:57,240 Speaker 1: got two potential outputs for that first step. Like our 524 00:33:57,320 --> 00:34:00,560 Speaker 1: mini max example, we can go down several layers of 525 00:34:00,640 --> 00:34:04,800 Speaker 1: possible choices, and we can wait the relationships between these 526 00:34:04,800 --> 00:34:08,600 Speaker 1: different choices. So if the incoming value is higher than 527 00:34:08,760 --> 00:34:12,760 Speaker 1: a certain amount, maybe the node sends it down one pathway, 528 00:34:12,800 --> 00:34:15,880 Speaker 1: But if the value is lower than that arbitrary amount, 529 00:34:16,200 --> 00:34:19,399 Speaker 1: the node will send it down a different pathway. This 530 00:34:19,520 --> 00:34:23,480 Speaker 1: is drastically oversimplifying, but I hope you kind of get 531 00:34:23,520 --> 00:34:26,960 Speaker 1: the idea. It's like a big sorting system, and the 532 00:34:27,000 --> 00:34:30,479 Speaker 1: goal is that at the very end whatever comes out 533 00:34:30,600 --> 00:34:35,640 Speaker 1: as output is correct or true. Ideally, you've got a 534 00:34:35,680 --> 00:34:40,840 Speaker 1: system that is self improving. It trains itself to be better. 535 00:34:41,320 --> 00:34:44,560 Speaker 1: But how the heck does that happen? Well, let's consider 536 00:34:44,920 --> 00:34:50,000 Speaker 1: cats for a bit, not the musical and good Heaven's 537 00:34:50,120 --> 00:34:56,000 Speaker 1: definitely not the movie musical. That is a subject that 538 00:34:56,239 --> 00:34:59,000 Speaker 1: deserves its own episode. Maybe one day I'll figure out 539 00:34:59,280 --> 00:35:01,000 Speaker 1: a way to tell a cackled that film with some 540 00:35:01,040 --> 00:35:04,080 Speaker 1: sort of tech capacity, But honestly, I'm just not ready 541 00:35:04,120 --> 00:35:07,480 Speaker 1: to do that yet. From like an emotional standpoint as 542 00:35:07,520 --> 00:35:11,760 Speaker 1: well as a research one. No, Let's say you're teaching 543 00:35:11,800 --> 00:35:16,480 Speaker 1: a computer system to recognize cats pictures of cats, and 544 00:35:16,480 --> 00:35:20,240 Speaker 1: the system has an artificial neural network that accepts input 545 00:35:20,600 --> 00:35:23,920 Speaker 1: pictures of cats and then filters that input through the 546 00:35:23,960 --> 00:35:27,920 Speaker 1: network to make the determination does this picture include a 547 00:35:28,000 --> 00:35:31,320 Speaker 1: cat in it? And you start feeding it lots of images. 548 00:35:31,719 --> 00:35:34,279 Speaker 1: The neural network acts on the data according to the 549 00:35:34,400 --> 00:35:39,640 Speaker 1: weighted relationship between the artificial neurons, and it produces an output. 550 00:35:40,440 --> 00:35:43,759 Speaker 1: Now here's the thing. We already know what we want 551 00:35:43,800 --> 00:35:46,880 Speaker 1: the output to be, because we can recognize if a 552 00:35:46,920 --> 00:35:50,040 Speaker 1: picture has a cat inet or not. Maybe we've got 553 00:35:50,200 --> 00:35:53,560 Speaker 1: one thousand pictures. This is the training data we're going 554 00:35:53,600 --> 00:35:57,040 Speaker 1: to use for this machine learning process. We also know 555 00:35:57,120 --> 00:35:59,759 Speaker 1: that eight hundred of those pictures have a cat in 556 00:35:59,800 --> 00:36:03,399 Speaker 1: the and two don't, so we know what we want 557 00:36:03,400 --> 00:36:06,400 Speaker 1: the results to be. We've got an artificial neural network 558 00:36:06,600 --> 00:36:10,000 Speaker 1: in which some neurons or nodes will accept input and 559 00:36:10,040 --> 00:36:12,680 Speaker 1: perform a function based on that input, and then the 560 00:36:12,719 --> 00:36:16,759 Speaker 1: weighted connections that neuron has to other neurons will determine 561 00:36:16,880 --> 00:36:19,719 Speaker 1: where it passes the information down until we get to 562 00:36:19,760 --> 00:36:23,040 Speaker 1: an output. And this happens until we get that conclusion. 563 00:36:23,680 --> 00:36:27,319 Speaker 1: So what happens if the computer's answer is wrong? What 564 00:36:27,520 --> 00:36:30,400 Speaker 1: if we feed those one thousand photos to it and 565 00:36:30,480 --> 00:36:33,719 Speaker 1: says only three hundred of them have cats in them? 566 00:36:33,719 --> 00:36:37,719 Speaker 1: While we have to go back and adjust those weighted connections, 567 00:36:37,719 --> 00:36:42,080 Speaker 1: because clearly something didn't go right, the connections within the 568 00:36:42,120 --> 00:36:47,080 Speaker 1: network need to be readjusted. We would likely start closest 569 00:36:47,120 --> 00:36:51,120 Speaker 1: to our output and see which neurons seem to contribute 570 00:36:51,120 --> 00:36:55,239 Speaker 1: to the mistake, which which neurons were responsible, In other words, 571 00:36:55,280 --> 00:36:58,080 Speaker 1: for it to say, oh, only three these pictures had 572 00:36:58,440 --> 00:37:01,920 Speaker 1: cats in them, and then we would adjust the weights, 573 00:37:01,960 --> 00:37:06,120 Speaker 1: the incoming weights of connections to those neurons in order 574 00:37:06,160 --> 00:37:10,160 Speaker 1: to try and favor pathways that lead to correct answers. 575 00:37:10,680 --> 00:37:13,640 Speaker 1: Then we feed it the one thousand pictures again and 576 00:37:13,719 --> 00:37:16,720 Speaker 1: we look at those results. Then we do this again 577 00:37:16,920 --> 00:37:20,239 Speaker 1: and again and again, every time, tweaking the network a 578 00:37:20,280 --> 00:37:24,520 Speaker 1: little bit so that it gets a bit better. Eventually, 579 00:37:24,760 --> 00:37:28,239 Speaker 1: when we have trained the system, we can start to 580 00:37:28,400 --> 00:37:32,960 Speaker 1: feed brand new data to the network, not the stuff 581 00:37:33,000 --> 00:37:36,920 Speaker 1: we've trained it on, but pictures that we and the 582 00:37:36,960 --> 00:37:40,440 Speaker 1: system have never seen before. And if our network is 583 00:37:40,440 --> 00:37:42,719 Speaker 1: a good one, if we have trained it well, it 584 00:37:42,760 --> 00:37:46,520 Speaker 1: will sort through these new photos and it will count 585 00:37:46,560 --> 00:37:49,560 Speaker 1: up the ones that have the cat pictures lickety split. 586 00:37:50,040 --> 00:37:54,080 Speaker 1: This approach is called supervised learning because it involves kind 587 00:37:54,120 --> 00:37:58,120 Speaker 1: of grading the network on its homework and then working 588 00:37:58,160 --> 00:38:02,000 Speaker 1: with it to get better. Heck, with the right algorithm, 589 00:38:02,000 --> 00:38:05,759 Speaker 1: a neural network can learn to recognize and differentiate patterns 590 00:38:06,200 --> 00:38:09,759 Speaker 1: even if we never explicitly told the system what it 591 00:38:09,840 --> 00:38:13,960 Speaker 1: was looking for. Google discovered this several years ago when 592 00:38:14,000 --> 00:38:18,280 Speaker 1: it fed several thousand YouTube videos to an enormous artificial 593 00:38:18,320 --> 00:38:22,600 Speaker 1: neural network. The system analyzed the videos that were fed 594 00:38:22,640 --> 00:38:26,800 Speaker 1: to it and gradually recognized patterns that represented different types 595 00:38:26,800 --> 00:38:32,399 Speaker 1: of stuff, like people or like cats, because there are 596 00:38:32,440 --> 00:38:35,760 Speaker 1: a lot of cat videos on YouTube, and the network 597 00:38:36,120 --> 00:38:38,360 Speaker 1: got to the point where it could identify an image 598 00:38:38,360 --> 00:38:42,239 Speaker 1: of a cat fairly reliably better than seventy of the time, 599 00:38:42,680 --> 00:38:46,480 Speaker 1: even though it was never told how to do that, 600 00:38:47,200 --> 00:38:51,080 Speaker 1: or it was never even told what a cat was. So, 601 00:38:51,120 --> 00:38:54,360 Speaker 1: as Google representatives put it, they said, it had to 602 00:38:54,520 --> 00:38:57,960 Speaker 1: invent the concept of a cat. It had to recognize 603 00:38:58,480 --> 00:39:02,960 Speaker 1: that cats are not the same as people, which I 604 00:39:03,000 --> 00:39:07,360 Speaker 1: think is a big slap in the face to some cats. Really, 605 00:39:08,000 --> 00:39:11,800 Speaker 1: what it said was that I recognized this particular pattern 606 00:39:11,840 --> 00:39:16,319 Speaker 1: of features, and I recognized that these other instances of 607 00:39:16,400 --> 00:39:20,080 Speaker 1: creatures that have a similar pattern seemed to match that, 608 00:39:20,320 --> 00:39:24,160 Speaker 1: and so I draw the conclusion that this instance of 609 00:39:24,200 --> 00:39:28,360 Speaker 1: a thing belongs with all these other instances of things 610 00:39:28,440 --> 00:39:32,880 Speaker 1: that are similar in characteristics. So this was more of 611 00:39:32,920 --> 00:39:36,719 Speaker 1: an example of unsupervised learning, and that the system, when 612 00:39:36,719 --> 00:39:39,879 Speaker 1: fed enough data, began to categorize stuff all on its 613 00:39:39,880 --> 00:39:43,920 Speaker 1: own through its own parameters. Now, one neat way that 614 00:39:43,960 --> 00:39:47,120 Speaker 1: computer scientists will train up systems for certain types of 615 00:39:47,160 --> 00:39:53,640 Speaker 1: applications is through a generative adversarial network, which I admit 616 00:39:53,760 --> 00:39:56,440 Speaker 1: sounds kind of sinister, doesn't it, And I mean it 617 00:39:56,520 --> 00:39:59,879 Speaker 1: can be, but it doesn't have to be essentially near 618 00:40:00,120 --> 00:40:04,320 Speaker 1: Using two different artificial neural networks. One of the networks 619 00:40:04,320 --> 00:40:08,240 Speaker 1: has a specific job. It's to fool the other network. 620 00:40:08,520 --> 00:40:11,480 Speaker 1: So the other network's job is to detect attempts to 621 00:40:11,560 --> 00:40:16,240 Speaker 1: fool it versus legitimate data. So let's use an example. 622 00:40:16,440 --> 00:40:18,399 Speaker 1: Let's say you're trying to create a system that can 623 00:40:18,440 --> 00:40:25,400 Speaker 1: make realistic but entirely computer generated, that is, fabricated photographs 624 00:40:25,440 --> 00:40:28,680 Speaker 1: of people. So, in other words, these are computer generated 625 00:40:28,719 --> 00:40:32,040 Speaker 1: images that don't actually represent a real person at all. 626 00:40:32,680 --> 00:40:36,359 Speaker 1: We've got one artificial neural network, the generator, and its 627 00:40:36,440 --> 00:40:41,160 Speaker 1: job is to create images of people that can pass 628 00:40:41,360 --> 00:40:44,640 Speaker 1: as real photographs. Then we've got our other network, which 629 00:40:44,680 --> 00:40:48,360 Speaker 1: is the discriminator. This is trying to sort out real 630 00:40:48,400 --> 00:40:52,960 Speaker 1: photos of actual people from pictures that have been generated 631 00:40:52,960 --> 00:40:57,640 Speaker 1: by the generative system. And we pick these two networks 632 00:40:57,680 --> 00:41:01,880 Speaker 1: against each other. The idea here is that both systems 633 00:41:02,000 --> 00:41:05,759 Speaker 1: get better as they test one another out. If the 634 00:41:05,840 --> 00:41:10,440 Speaker 1: generator network is falling behind because the discriminator can suss 635 00:41:10,480 --> 00:41:13,040 Speaker 1: out the fakes too easily, well, then it's time to 636 00:41:13,040 --> 00:41:17,240 Speaker 1: tweak some weights in that neural network that are leading 637 00:41:17,280 --> 00:41:22,560 Speaker 1: to dissatisfactory computer generated images and try it again. But then, 638 00:41:22,600 --> 00:41:27,799 Speaker 1: if the discriminator is starting to miss fakes while, it's 639 00:41:27,800 --> 00:41:31,480 Speaker 1: time to tweak the discriminator network. So it's better at 640 00:41:31,600 --> 00:41:36,080 Speaker 1: spotting the false pictures. Now along the way, some pretty 641 00:41:36,080 --> 00:41:40,760 Speaker 1: extraordinary stuff can happen. There are photos of computer generated faces, 642 00:41:41,120 --> 00:41:45,400 Speaker 1: not altered pictures, not ones created by a human artist, 643 00:41:45,760 --> 00:41:50,120 Speaker 1: but entirely composed via a computer, and they can look 644 00:41:50,520 --> 00:41:56,000 Speaker 1: absolutely realistic, complete with consistent lighting and shadows. This is 645 00:41:56,080 --> 00:42:00,759 Speaker 1: only after lots of training sessions the networks learn what 646 00:42:00,840 --> 00:42:04,920 Speaker 1: the giveaways are, like, what is it that leads the 647 00:42:04,920 --> 00:42:08,040 Speaker 1: discriminator to say, no, this is a fake photo, and 648 00:42:08,080 --> 00:42:10,600 Speaker 1: how can you fix that? It reminds me a bit 649 00:42:10,640 --> 00:42:14,080 Speaker 1: of how photo experts used to point out really bad 650 00:42:14,160 --> 00:42:18,560 Speaker 1: photoshop jobs and explaining how certain elements like shadows or 651 00:42:18,680 --> 00:42:22,120 Speaker 1: edges or whatever, we're a dead giveaway that someone had 652 00:42:22,160 --> 00:42:26,280 Speaker 1: altered an image. Well, similar rules exist for generated images, 653 00:42:26,640 --> 00:42:30,480 Speaker 1: and through training, the generator gets better at making really 654 00:42:30,560 --> 00:42:34,600 Speaker 1: convincing examples that don't fall into the traps that would 655 00:42:34,600 --> 00:42:39,239 Speaker 1: reveal it as a fake. Over time, generative networks can 656 00:42:39,280 --> 00:42:42,279 Speaker 1: get good enough to produce stuff that would be very 657 00:42:42,320 --> 00:42:44,600 Speaker 1: difficult for a human to tell apart from the quote 658 00:42:44,640 --> 00:42:48,400 Speaker 1: unquote real thing, and discriminators can get good enough to 659 00:42:48,440 --> 00:42:52,680 Speaker 1: detect fakes that would otherwise pass human inspection. So an 660 00:42:52,719 --> 00:42:57,240 Speaker 1: example of This is the current ongoing battle with deep fakes. 661 00:42:57,280 --> 00:43:00,960 Speaker 1: These are computer generated videos that appear to be legit. 662 00:43:01,360 --> 00:43:04,800 Speaker 1: If they're done well enough, they can have famous people 663 00:43:04,880 --> 00:43:07,160 Speaker 1: in them. Doesn't have to be a famous person, but 664 00:43:07,239 --> 00:43:09,680 Speaker 1: it can show a video of someone doing something that 665 00:43:09,719 --> 00:43:13,799 Speaker 1: they absolutely never did, but according to the video, they did, 666 00:43:14,360 --> 00:43:16,840 Speaker 1: and it can be really convincing if it's done well. 667 00:43:17,320 --> 00:43:21,680 Speaker 1: A good deep fake can fool people if you aren't 668 00:43:21,719 --> 00:43:23,879 Speaker 1: paying too much attention. Some of the really good ones 669 00:43:23,920 --> 00:43:29,000 Speaker 1: can pass pretty deep scrutiny. So this requires researchers to 670 00:43:29,000 --> 00:43:32,520 Speaker 1: come up with solutions that are pretty subtle and beyond 671 00:43:32,520 --> 00:43:35,640 Speaker 1: the average person's ability to replicate, like looking at the 672 00:43:35,719 --> 00:43:39,720 Speaker 1: reflections in the person's eyes and whether or not they 673 00:43:39,760 --> 00:43:43,600 Speaker 1: seem realistic or a computer generated. But that really just 674 00:43:43,680 --> 00:43:47,800 Speaker 1: represents another hurdle for the generative side. So in other words, 675 00:43:48,680 --> 00:43:53,799 Speaker 1: this is a seesaw approach, right. It's creating fakes on 676 00:43:53,800 --> 00:43:57,160 Speaker 1: one side and detecting them on the other side. It's 677 00:43:57,200 --> 00:44:00,000 Speaker 1: something we see in artificial intelligence in general. A similar 678 00:44:00,000 --> 00:44:03,520 Speaker 1: our story played out with the old capture systems, where 679 00:44:04,040 --> 00:44:06,440 Speaker 1: you know, we saw back and forth between methods to 680 00:44:06,520 --> 00:44:10,799 Speaker 1: try and weed out bots by using capture images that 681 00:44:10,840 --> 00:44:15,000 Speaker 1: only humans could really parse, and then we saw improved 682 00:44:15,040 --> 00:44:19,040 Speaker 1: bots that could analyze these images and return correct results, 683 00:44:19,520 --> 00:44:22,840 Speaker 1: which meant it was necessary to create more difficult captures. 684 00:44:22,960 --> 00:44:25,600 Speaker 1: Eventually get to a point where the captures are difficult 685 00:44:25,719 --> 00:44:28,239 Speaker 1: enough where the average person can't even pass them, and 686 00:44:28,239 --> 00:44:30,799 Speaker 1: then you have to go to a different method. We 687 00:44:30,880 --> 00:44:33,720 Speaker 1: also see this play out in the cyber security realm, 688 00:44:33,760 --> 00:44:36,960 Speaker 1: where you might say the thieves get better at lock picking, 689 00:44:37,360 --> 00:44:40,800 Speaker 1: and then security experts make better locks, and the cycle 690 00:44:40,880 --> 00:44:46,080 Speaker 1: just repeats endlessly. One thing that has really fueled machine 691 00:44:46,160 --> 00:44:50,040 Speaker 1: learning recently is the era of big data. Being able 692 00:44:50,080 --> 00:44:54,680 Speaker 1: to harvest information on a truly massive scale provides the 693 00:44:54,680 --> 00:44:59,560 Speaker 1: opportunity to feed that data into various machine learning systems 694 00:45:00,200 --> 00:45:04,680 Speaker 1: to search for meaning within that data. These systems might 695 00:45:04,840 --> 00:45:08,560 Speaker 1: scour the information to look for stuff like criminal activity 696 00:45:08,920 --> 00:45:13,120 Speaker 1: like financial crimes or the attempt to move some money 697 00:45:13,160 --> 00:45:17,120 Speaker 1: around from various criminal exploits. Or it could be used 698 00:45:17,160 --> 00:45:20,640 Speaker 1: to look for trends like market trends, or it might 699 00:45:20,640 --> 00:45:24,879 Speaker 1: be used to plot possible spikes in COVID nineteen transmission 700 00:45:25,280 --> 00:45:28,440 Speaker 1: where those might occur where people should really be focusing 701 00:45:28,480 --> 00:45:31,759 Speaker 1: their attention. But now we got to think back on 702 00:45:31,840 --> 00:45:35,080 Speaker 1: what I said earlier about looking up at the sky 703 00:45:35,200 --> 00:45:39,600 Speaker 1: and seeing shapes in the clouds. There's a risk that 704 00:45:39,680 --> 00:45:42,319 Speaker 1: comes along with machine learning. Actually, technically there are a 705 00:45:42,320 --> 00:45:45,120 Speaker 1: lot of risks, but this one is a biggie. It 706 00:45:45,239 --> 00:45:49,680 Speaker 1: is possible for machines like humans, to detect a pattern 707 00:45:49,840 --> 00:45:54,480 Speaker 1: where there really isn't a pattern. Systems might interpret noise 708 00:45:54,760 --> 00:45:57,279 Speaker 1: to be signal, and depending on what you're using the 709 00:45:57,320 --> 00:46:01,240 Speaker 1: system to do, that could lead you to some seriously dangerous, 710 00:46:01,360 --> 00:46:05,799 Speaker 1: incorrect conclusions. In some cases, you could just be inconvenient, 711 00:46:05,840 --> 00:46:09,000 Speaker 1: but depending on what you're working toward, it could be catastrophic. 712 00:46:09,120 --> 00:46:12,000 Speaker 1: And so computer scientists know they have to do a 713 00:46:12,000 --> 00:46:15,600 Speaker 1: lot of analysis to make sure that patterns that are 714 00:46:15,640 --> 00:46:21,440 Speaker 1: identified through machine learning processes are actually real before acting 715 00:46:21,640 --> 00:46:28,320 Speaker 1: on that information. Likewise, bias is something that we humans have, well, 716 00:46:28,440 --> 00:46:31,719 Speaker 1: it's also something that machine learning systems have too. Now, 717 00:46:31,800 --> 00:46:35,319 Speaker 1: sometimes bias is intentional. It can take the form of 718 00:46:35,360 --> 00:46:42,000 Speaker 1: those weighted relationships between artificial neurons. Other times, a systems architects, 719 00:46:42,080 --> 00:46:44,080 Speaker 1: you know, the people who put it together, They might 720 00:46:44,200 --> 00:46:48,879 Speaker 1: have introduced bias, not through conscious effort, but merely through 721 00:46:49,400 --> 00:46:52,480 Speaker 1: the approach they took and that approach might have been 722 00:46:52,560 --> 00:46:56,120 Speaker 1: too narrow. We've seen this pop up a lot again 723 00:46:56,160 --> 00:46:59,840 Speaker 1: with facial recognition technologies, many of which have a sliding 724 00:47:00,200 --> 00:47:04,560 Speaker 1: scale of efficacy. They might be more reliable with certain 725 00:47:04,600 --> 00:47:09,000 Speaker 1: ethnicities like white people, over others. That points that a 726 00:47:09,120 --> 00:47:12,920 Speaker 1: likely problem with the way those systems were trained. This 727 00:47:13,040 --> 00:47:15,600 Speaker 1: is one of the reasons why many companies have made 728 00:47:15,640 --> 00:47:19,760 Speaker 1: a choice to stop supplying certain parties like police forces 729 00:47:19,800 --> 00:47:24,360 Speaker 1: and military branches with facial recognition systems. The systems aren't 730 00:47:24,400 --> 00:47:28,600 Speaker 1: reliable for all demographic groups and thus could cause disproportionate 731 00:47:28,680 --> 00:47:32,360 Speaker 1: harm to certain populations. It would be a technological approach 732 00:47:32,400 --> 00:47:36,040 Speaker 1: to systemic racism, and this stuff is already out there 733 00:47:36,080 --> 00:47:38,959 Speaker 1: in the wild. You might think a computer system can't 734 00:47:38,960 --> 00:47:43,640 Speaker 1: be biased or prejudiced or racist, and sure, we're still 735 00:47:43,800 --> 00:47:46,120 Speaker 1: not at the point where these systems are thinking in 736 00:47:46,160 --> 00:47:49,239 Speaker 1: the way that humans do, but the outcome is still 737 00:47:49,360 --> 00:47:53,920 Speaker 1: disproportionately harmful to some groups. That's not to say that 738 00:47:53,960 --> 00:47:58,040 Speaker 1: machine learning itself is bad. It's not bad. It's a tool, 739 00:47:58,360 --> 00:48:02,520 Speaker 1: just as all technology is a tool used properly with 740 00:48:02,640 --> 00:48:05,960 Speaker 1: a careful hand to make sure that biases understood and 741 00:48:06,040 --> 00:48:10,600 Speaker 1: where needed mitigated and where work can be double or 742 00:48:10,640 --> 00:48:14,840 Speaker 1: triple checked before acted upon. It is a remarkably useful tool, 743 00:48:15,040 --> 00:48:18,759 Speaker 1: one that will power and design and improve elements in 744 00:48:18,800 --> 00:48:23,040 Speaker 1: our lives if it's under the correct stewardship. But it 745 00:48:23,160 --> 00:48:26,560 Speaker 1: does require a bit more hands on work. We can't 746 00:48:27,120 --> 00:48:32,520 Speaker 1: just leave it to the machines just yet. Well, that 747 00:48:32,560 --> 00:48:35,960 Speaker 1: wraps up this look at the concept of machine learning 748 00:48:36,000 --> 00:48:39,720 Speaker 1: and some of the thought that underlies it. This really 749 00:48:39,840 --> 00:48:44,160 Speaker 1: is a very high level treatment of machine learning. There 750 00:48:44,200 --> 00:48:47,080 Speaker 1: are plenty of resources online if you want to dive 751 00:48:47,120 --> 00:48:50,040 Speaker 1: in and learn more. A lot of them get very 752 00:48:50,120 --> 00:48:52,760 Speaker 1: heavy into the math, so if that's not your bag, 753 00:48:53,560 --> 00:48:56,000 Speaker 1: it might be a little challenging to navigate. It certainly 754 00:48:56,080 --> 00:48:59,279 Speaker 1: is for me. I love learning about the stuff, but 755 00:49:00,160 --> 00:49:03,239 Speaker 1: a lot of it requires me to look up a term, 756 00:49:03,560 --> 00:49:06,359 Speaker 1: then look up a term that explains that term, and 757 00:49:06,400 --> 00:49:09,600 Speaker 1: so on, and I go down a rabbit hole. I 758 00:49:09,640 --> 00:49:13,000 Speaker 1: hope you enjoyed that classic episode. I guess not classic, 759 00:49:13,040 --> 00:49:15,759 Speaker 1: that rerun episode of tech stuff. You can't call it 760 00:49:15,800 --> 00:49:19,080 Speaker 1: a classic if it's just a year old, right, So anyway, 761 00:49:19,120 --> 00:49:22,440 Speaker 1: I will be back again tomorrow hopefully, and we will 762 00:49:22,480 --> 00:49:25,000 Speaker 1: have a new episode, y'all If you want to get 763 00:49:25,040 --> 00:49:26,480 Speaker 1: in touch with me and let me know what you 764 00:49:26,480 --> 00:49:28,640 Speaker 1: would like me to cover in future episodes. There are 765 00:49:28,640 --> 00:49:30,320 Speaker 1: a couple of ways of doing that. You can drop 766 00:49:30,400 --> 00:49:32,359 Speaker 1: a note on Twitter. Several of you have been doing 767 00:49:32,360 --> 00:49:35,760 Speaker 1: that recently and I've got I've got a list of topics. 768 00:49:35,800 --> 00:49:39,120 Speaker 1: So thank you so much. That's fantastic. I really appreciate it. 769 00:49:39,520 --> 00:49:43,840 Speaker 1: Keep them coming. The The handle for the podcast Twitter 770 00:49:43,920 --> 00:49:47,759 Speaker 1: feed is text Stuff hs W. If, however, you would 771 00:49:47,760 --> 00:49:50,040 Speaker 1: like to leave me a voice message, you can go 772 00:49:50,120 --> 00:49:52,520 Speaker 1: to the I Heart Radio app go to the tech 773 00:49:52,600 --> 00:49:55,759 Speaker 1: stuff page. There's a little microphone icon you click on 774 00:49:55,840 --> 00:49:58,600 Speaker 1: that you can leave a message of up to thirty 775 00:49:58,680 --> 00:50:01,480 Speaker 1: seconds and if you like me to include that message 776 00:50:01,480 --> 00:50:04,120 Speaker 1: in an upcoming episode, just let me know in the message. 777 00:50:04,120 --> 00:50:06,719 Speaker 1: Because I'm an opt in kind of guy. That's it. 778 00:50:06,960 --> 00:50:09,120 Speaker 1: Hope you all are doing well and I'll talk to 779 00:50:09,120 --> 00:50:17,440 Speaker 1: you again really soon. Y text Stuff is an I 780 00:50:17,560 --> 00:50:21,040 Speaker 1: Heart Radio production. For more podcasts from my Heart Radio, 781 00:50:21,400 --> 00:50:24,560 Speaker 1: visit the i Heart Radio app, Apple Podcasts, or wherever 782 00:50:24,640 --> 00:50:26,160 Speaker 1: you listen to your favorite shows.