1 00:00:04,400 --> 00:00:07,800 Speaker 1: Welcome to tech Stuff, a production from I Heart Radio. 2 00:00:12,119 --> 00:00:15,440 Speaker 1: Hey there, and welcome to tech Stuff. This is your host, 3 00:00:15,640 --> 00:00:19,120 Speaker 1: Jonathan Strickland. I'm an executive producer with I Heart Radio 4 00:00:19,160 --> 00:00:22,360 Speaker 1: and I love all things sex. You know, folks, Back 5 00:00:22,400 --> 00:00:28,000 Speaker 1: in nineteen eighties, six comedy science fiction film that I 6 00:00:28,040 --> 00:00:32,400 Speaker 1: saw in the theater about a robot the games sentience 7 00:00:32,440 --> 00:00:35,640 Speaker 1: and becomes a total goofball what it will? It hit 8 00:00:35,720 --> 00:00:39,199 Speaker 1: theaters in eight six and it was called Short Circuit. 9 00:00:39,880 --> 00:00:44,200 Speaker 1: The movie starred Steve Gutenberg, Ali Sheety, and lamentably a 10 00:00:44,240 --> 00:00:48,080 Speaker 1: white actor named Fisher Stevens playing a non white character, 11 00:00:48,720 --> 00:00:52,640 Speaker 1: someone who is Indian. I should add that's not Steven's fault. 12 00:00:52,880 --> 00:00:55,760 Speaker 1: I mean, he auditioned to be in a movie and 13 00:00:55,840 --> 00:00:58,280 Speaker 1: he got a gig. He didn't cast himself in the film, 14 00:00:58,320 --> 00:01:01,800 Speaker 1: and he has since talked about his experiences realizing the 15 00:01:01,840 --> 00:01:04,280 Speaker 1: problems with a white man playing a non white character, 16 00:01:04,319 --> 00:01:08,880 Speaker 1: but setting aside all the problematic whitewashing, the movie showed 17 00:01:09,000 --> 00:01:11,920 Speaker 1: this robot, who in the course of the film names 18 00:01:11,959 --> 00:01:17,000 Speaker 1: itself Johnny five. Learning. It learns about the world around it, 19 00:01:17,120 --> 00:01:20,840 Speaker 1: It learns about people, It learns about human concepts like 20 00:01:21,000 --> 00:01:25,720 Speaker 1: humor and emotion, and the general idea was pretty cute. 21 00:01:26,800 --> 00:01:31,119 Speaker 1: Now the nifty thing is machines actually can learn. In fact, 22 00:01:31,200 --> 00:01:35,120 Speaker 1: machine learning is a really important field of study these days, 23 00:01:35,480 --> 00:01:38,920 Speaker 1: complete with its own challenges and risks. I've talked about 24 00:01:39,000 --> 00:01:41,399 Speaker 1: machine learning a few times in the past, but I 25 00:01:41,400 --> 00:01:44,240 Speaker 1: figured we could do a deeper dive to understand what 26 00:01:44,400 --> 00:01:48,120 Speaker 1: machine learning is, what it isn't, how people are leveraging 27 00:01:48,160 --> 00:01:51,880 Speaker 1: machine learning, and why. I said that it does come 28 00:01:51,920 --> 00:01:58,280 Speaker 1: with risks, so let's learn about machines learning. It will 29 00:01:58,320 --> 00:02:02,480 Speaker 1: be impossible to talk about machine learning without also talking 30 00:02:02,520 --> 00:02:08,079 Speaker 1: about artificial intelligence or AI. And this term artificial intelligence 31 00:02:08,280 --> 00:02:12,880 Speaker 1: is a real doozy. It trips people up, even people 32 00:02:12,960 --> 00:02:17,880 Speaker 1: who have dedicated their lives to researching and developing artificial intelligence. 33 00:02:18,200 --> 00:02:22,679 Speaker 1: You can get two experts in AI talking about AI 34 00:02:22,800 --> 00:02:25,600 Speaker 1: and find out that because they have slightly different takes 35 00:02:25,639 --> 00:02:31,320 Speaker 1: on what AI is, there are some communication issues. It's 36 00:02:31,320 --> 00:02:34,200 Speaker 1: not as simple as red versus blue would have you think, 37 00:02:34,840 --> 00:02:40,440 Speaker 1: what does the A stand for? So when you really 38 00:02:40,880 --> 00:02:43,200 Speaker 1: boil it down, it comes out as as no big 39 00:02:43,200 --> 00:02:46,239 Speaker 1: surprise that there's a lot of ambiguity here. After all, 40 00:02:46,600 --> 00:02:51,639 Speaker 1: how would you define intelligence just intelligence, not artificial intelligence 41 00:02:52,000 --> 00:02:56,640 Speaker 1: just intelligence? Well, would it be the ability to learn, 42 00:02:57,000 --> 00:03:01,240 Speaker 1: that is, to acquire skills and knowledge? Or is it 43 00:03:01,320 --> 00:03:04,680 Speaker 1: the application of learning? Is it problems solving? Is it 44 00:03:05,160 --> 00:03:08,440 Speaker 1: being able to think ahead and make plans in order 45 00:03:08,480 --> 00:03:12,720 Speaker 1: to achieve a specific goal? Is it the ability to 46 00:03:12,960 --> 00:03:16,560 Speaker 1: examine a problem and deconstructed in order to figure out 47 00:03:16,600 --> 00:03:19,600 Speaker 1: the best solution a more specific version of problem solving? 48 00:03:20,240 --> 00:03:25,560 Speaker 1: Is it the ability to recognize, understand, and navigate emotional scenarios? Now, 49 00:03:25,680 --> 00:03:30,920 Speaker 1: arguably it's all of these things and more. We all 50 00:03:31,000 --> 00:03:35,400 Speaker 1: have kind of an intuitive grasp on what intelligence is, 51 00:03:36,280 --> 00:03:40,960 Speaker 1: but defining it in a simple way tends to feel 52 00:03:41,000 --> 00:03:44,080 Speaker 1: reductive and it leaves out a lot of important details. 53 00:03:44,480 --> 00:03:50,160 Speaker 1: So if defining just general intelligence is hard, it stands 54 00:03:50,200 --> 00:03:55,000 Speaker 1: for reason that defining artificial intelligence is also a tough job. Heck, 55 00:03:55,320 --> 00:03:58,640 Speaker 1: even coming up with a number of different types of 56 00:03:58,680 --> 00:04:02,720 Speaker 1: AI is tricky. And if you don't believe me, just 57 00:04:02,920 --> 00:04:08,920 Speaker 1: google the phrase different types of artificial intelligence. Never mind, 58 00:04:08,960 --> 00:04:10,680 Speaker 1: you don't. You don't really actually have to do that. 59 00:04:10,720 --> 00:04:13,800 Speaker 1: I already did it, though, Feel free to do it 60 00:04:13,840 --> 00:04:16,839 Speaker 1: yourself and check my work if you like. When I 61 00:04:17,080 --> 00:04:20,640 Speaker 1: googled that phrase different types of AI, some of The 62 00:04:20,680 --> 00:04:24,960 Speaker 1: top results included a blog post on BMC Software titled 63 00:04:25,240 --> 00:04:28,840 Speaker 1: four types of Artificial Intelligence. But then there was also 64 00:04:28,880 --> 00:04:31,840 Speaker 1: an article on code bots that was titled what are 65 00:04:31,880 --> 00:04:34,960 Speaker 1: the three types of AI? And then there was an 66 00:04:35,040 --> 00:04:40,080 Speaker 1: article from Forbes titled seven types of Artificial Intelligence. See, 67 00:04:40,279 --> 00:04:43,320 Speaker 1: we can't even agree on how many versions of A 68 00:04:43,480 --> 00:04:48,560 Speaker 1: EITHER are because defining a I is really hard. It 69 00:04:48,680 --> 00:04:52,080 Speaker 1: largely depends upon how you view AI and then how 70 00:04:52,080 --> 00:04:56,039 Speaker 1: you break it down into different realms of intelligence. Now 71 00:04:56,080 --> 00:04:59,839 Speaker 1: we could go super high level because a classic way 72 00:04:59,839 --> 00:05:04,960 Speaker 1: to look at AI is strong versus weak artificial intelligence. 73 00:05:06,560 --> 00:05:12,440 Speaker 1: Strong AI UH sometimes called artificial general intelligence, would be 74 00:05:12,600 --> 00:05:17,640 Speaker 1: a machine that processes information and at least appears to 75 00:05:17,839 --> 00:05:21,599 Speaker 1: have some form of consciousness and self awareness and the 76 00:05:21,640 --> 00:05:26,000 Speaker 1: ability to both have experiences and to be aware that 77 00:05:26,120 --> 00:05:30,239 Speaker 1: it is having experiences. It might even feel emotion, though 78 00:05:30,880 --> 00:05:34,880 Speaker 1: maybe not emotions that we could easily identify or sympathize with. 79 00:05:35,520 --> 00:05:38,839 Speaker 1: So this would be the kind of machine that would 80 00:05:39,080 --> 00:05:42,560 Speaker 1: think in a way similar to humans. It would be 81 00:05:42,600 --> 00:05:45,960 Speaker 1: able to sense its environment and not just react, but 82 00:05:46,120 --> 00:05:49,800 Speaker 1: really process what is going on and build and understanding. 83 00:05:50,000 --> 00:05:51,880 Speaker 1: It's the type of AI that we see a lot 84 00:05:52,160 --> 00:05:55,239 Speaker 1: in science fiction. That's the type of AI of Johnny 85 00:05:55,320 --> 00:05:59,240 Speaker 1: five from Short Circuit, or how from two thousand one, 86 00:05:59,360 --> 00:06:02,440 Speaker 1: or the draw aids in Star Wars. It's also a 87 00:06:02,440 --> 00:06:06,560 Speaker 1: type of artificial intelligence that we have yet to actually 88 00:06:06,600 --> 00:06:11,479 Speaker 1: achieve in the real world. So then what is weak AI. Well, 89 00:06:12,680 --> 00:06:16,120 Speaker 1: you could say it's everything else, or you could say 90 00:06:16,120 --> 00:06:21,080 Speaker 1: it's the building blocks that maybe collectively will lead to 91 00:06:21,200 --> 00:06:26,200 Speaker 1: strong AI. Week AI involves processes that allow machines to 92 00:06:26,279 --> 00:06:31,360 Speaker 1: complete tasks. So, for example, image recognition software could fall 93 00:06:31,400 --> 00:06:34,680 Speaker 1: into this category. Once upon a time, in order to 94 00:06:34,760 --> 00:06:39,560 Speaker 1: search photos effectively, you needed to actually add meta data 95 00:06:39,839 --> 00:06:45,039 Speaker 1: like tags to those photos. So, for example, I might 96 00:06:45,400 --> 00:06:49,240 Speaker 1: tag pictures of my dog with the meta tag dog, 97 00:06:50,040 --> 00:06:52,719 Speaker 1: and then if I wanted to see photos of my pooch, 98 00:06:53,320 --> 00:06:55,440 Speaker 1: then I would pull up my photo app and search 99 00:06:55,760 --> 00:06:58,840 Speaker 1: the term dog and all the photos that I had 100 00:06:58,880 --> 00:07:01,280 Speaker 1: tagged with the word dog would show up. But if 101 00:07:01,320 --> 00:07:04,679 Speaker 1: I had failed to tag some pictures of my dog, 102 00:07:05,279 --> 00:07:07,839 Speaker 1: those pictures wouldn't pop up in search because the computer 103 00:07:07,839 --> 00:07:11,240 Speaker 1: program wasn't actually looking for dogs and my photos, it 104 00:07:11,360 --> 00:07:13,960 Speaker 1: was just looking for photos that had that particular meta 105 00:07:14,000 --> 00:07:18,040 Speaker 1: tag attached to it. But now we've reached a point 106 00:07:18,240 --> 00:07:21,400 Speaker 1: where at least some photo apps are using image recognition 107 00:07:21,480 --> 00:07:25,240 Speaker 1: to analyze photos, and these will return results that the 108 00:07:25,280 --> 00:07:28,679 Speaker 1: algorithm has identified as having a reasonable chance of meeting 109 00:07:28,880 --> 00:07:31,720 Speaker 1: your search query. So if I used an app like 110 00:07:31,760 --> 00:07:35,480 Speaker 1: that and I put in dog as my search term, 111 00:07:35,520 --> 00:07:38,239 Speaker 1: it could pull up photos that had no meta tags 112 00:07:38,240 --> 00:07:41,160 Speaker 1: attached to them at all, because the search is relying 113 00:07:41,200 --> 00:07:44,640 Speaker 1: on image recognition. Now, this also means that if the 114 00:07:44,680 --> 00:07:48,120 Speaker 1: image recognition algorithm isn't very good, I could get some 115 00:07:48,200 --> 00:07:50,360 Speaker 1: images that don't have a dog in them at all, 116 00:07:50,880 --> 00:07:54,040 Speaker 1: or it might miss other images that have my dog 117 00:07:54,120 --> 00:07:56,760 Speaker 1: in them. But my point is that the ability to 118 00:07:56,840 --> 00:07:59,760 Speaker 1: identify whether or not a dog is in a particular 119 00:08:00,080 --> 00:08:05,760 Speaker 1: photo represents a kind of weak artificial intelligence. You wouldn't 120 00:08:05,800 --> 00:08:10,640 Speaker 1: say that the photo search tool possesses humanlike intelligence, because 121 00:08:10,680 --> 00:08:14,200 Speaker 1: really it only does one thing. It's analyzing photos and 122 00:08:14,240 --> 00:08:17,840 Speaker 1: looks for matches to specific search queries, but it can't 123 00:08:17,920 --> 00:08:21,440 Speaker 1: do anything outside of that use case. However, that's just 124 00:08:21,560 --> 00:08:24,520 Speaker 1: one little example. There are all sorts of other ones, 125 00:08:24,560 --> 00:08:30,480 Speaker 1: like voice recognition, environmental sensing, course plotting, that kind of thing, 126 00:08:30,720 --> 00:08:33,880 Speaker 1: and in some circles, as we get better at making machines, 127 00:08:33,920 --> 00:08:39,040 Speaker 1: and systems that can do these things. Those elements seem 128 00:08:39,080 --> 00:08:42,360 Speaker 1: to kind of drift away from the ongoing conversation about 129 00:08:42,440 --> 00:08:46,199 Speaker 1: artificial intelligence. A guy named Larry Tesler, who was a 130 00:08:46,240 --> 00:08:49,160 Speaker 1: computer scientist who worked at lots of really important places 131 00:08:49,240 --> 00:08:54,520 Speaker 1: like Xerox Park and Amazon and Apple. He once observed, 132 00:08:54,640 --> 00:08:59,920 Speaker 1: quote intelligence is whatever machines haven't done yet, end quote. 133 00:09:00,440 --> 00:09:03,480 Speaker 1: So his point was that the reason that AI is 134 00:09:03,559 --> 00:09:06,120 Speaker 1: really hard to talk about is that the goal post 135 00:09:06,320 --> 00:09:12,920 Speaker 1: for what actually is artificial intelligence is constantly moving. Now, 136 00:09:12,920 --> 00:09:16,719 Speaker 1: this pretty much mirrors how we think about things like consciousness. 137 00:09:17,120 --> 00:09:20,640 Speaker 1: Lots of people study consciousness, and the general sense I 138 00:09:20,679 --> 00:09:23,240 Speaker 1: get is that it's a lot easier for people to 139 00:09:23,280 --> 00:09:29,080 Speaker 1: talk about what isn't consciousness rather than what consciousness actually is. 140 00:09:29,760 --> 00:09:33,480 Speaker 1: And it seems like artificial intelligence is in a similar place, 141 00:09:33,559 --> 00:09:36,520 Speaker 1: which really isn't that big of a surprise as we 142 00:09:36,679 --> 00:09:41,200 Speaker 1: closely associate intelligence with consciousness. Now this leads us to 143 00:09:41,720 --> 00:09:45,160 Speaker 1: why there are so many different takes on how many 144 00:09:45,200 --> 00:09:48,680 Speaker 1: types of AI there are. It all depends on how 145 00:09:48,800 --> 00:09:53,199 Speaker 1: you classify different disciplines in artificial intelligence, and over time, 146 00:09:53,679 --> 00:09:57,679 Speaker 1: a lot of disciplines that were previously distinct from AI 147 00:09:57,800 --> 00:10:01,599 Speaker 1: have sort of converged into becoming heart of the AI discussion. 148 00:10:01,880 --> 00:10:04,840 Speaker 1: Machine learning, as it turns out, was part of the 149 00:10:04,920 --> 00:10:09,520 Speaker 1: AI discussion, branched off from it, and then rejoined the 150 00:10:09,559 --> 00:10:12,920 Speaker 1: AI discussion years later. So I am not going to 151 00:10:12,960 --> 00:10:16,240 Speaker 1: go down all the different approaches to classification because I 152 00:10:16,280 --> 00:10:18,719 Speaker 1: don't know that they would be that valuable to us. 153 00:10:19,200 --> 00:10:21,120 Speaker 1: They would really just illustrate that there are a lot 154 00:10:21,160 --> 00:10:26,280 Speaker 1: of different ways to look at the subject. So if 155 00:10:26,360 --> 00:10:30,559 Speaker 1: you ever find yourself in a conversation about AI, it 156 00:10:30,640 --> 00:10:33,720 Speaker 1: might be a good idea to set a few ground 157 00:10:33,840 --> 00:10:37,440 Speaker 1: rules as to what everyone means when they use the 158 00:10:37,520 --> 00:10:42,760 Speaker 1: term artificial intelligence. That can help with expectations and understanding. 159 00:10:43,320 --> 00:10:46,200 Speaker 1: Or you could just run for the nearest exit, which 160 00:10:46,240 --> 00:10:49,920 Speaker 1: is what people tend to do whenever I start talking 161 00:10:49,960 --> 00:10:56,040 Speaker 1: about it anyway. What about machine learning, Well, from one perspective, 162 00:10:56,280 --> 00:10:59,199 Speaker 1: you could say machine learning is a sub discipline of 163 00:10:59,280 --> 00:11:03,080 Speaker 1: artificial and eligence, although like I said, it hasn't always 164 00:11:03,120 --> 00:11:07,080 Speaker 1: been viewed as such. I think most people would say 165 00:11:07,080 --> 00:11:11,000 Speaker 1: that the ability to learn that is to take information 166 00:11:11,160 --> 00:11:15,280 Speaker 1: and experience and then have some form of understanding of 167 00:11:15,320 --> 00:11:19,080 Speaker 1: those things so that you can apply that to future tasks. 168 00:11:19,240 --> 00:11:23,160 Speaker 1: Potentially getting better over time. I would say most people 169 00:11:23,200 --> 00:11:26,720 Speaker 1: would call that part of intelligence, but you could also 170 00:11:26,760 --> 00:11:29,240 Speaker 1: be a bit more wishy washy and say it's related to, 171 00:11:29,880 --> 00:11:33,520 Speaker 1: you know, artificial intelligence, as opposed to being part of AI. 172 00:11:33,640 --> 00:11:37,839 Speaker 1: Since the definition of AI is let's say, fluid, either 173 00:11:37,920 --> 00:11:41,520 Speaker 1: way of classifying. Machine learning works. As far as I'm concerned, 174 00:11:42,600 --> 00:11:46,160 Speaker 1: machine learning boils down to the idea of creating a 175 00:11:46,200 --> 00:11:50,120 Speaker 1: system that can learn as it performs a task. It 176 00:11:50,160 --> 00:11:54,679 Speaker 1: can learn what works and more importantly, what does not work. 177 00:11:55,200 --> 00:11:57,440 Speaker 1: You may have heard that we learn a lot more 178 00:11:57,520 --> 00:12:01,200 Speaker 1: from our mistakes than we do from our successes, which 179 00:12:01,880 --> 00:12:05,320 Speaker 1: there's pretty much true in my experience. When something goes wrong, 180 00:12:05,880 --> 00:12:11,280 Speaker 1: it's usually, but not always, possible to trace the event 181 00:12:11,480 --> 00:12:14,760 Speaker 1: or events that led to the failure. You can identify 182 00:12:14,840 --> 00:12:19,120 Speaker 1: decisions that we're probably the wrong ones or that led 183 00:12:19,200 --> 00:12:22,679 Speaker 1: to a bad outcome, But if you have a success, 184 00:12:23,080 --> 00:12:27,160 Speaker 1: it's hard to figure out which decisions were key to 185 00:12:27,280 --> 00:12:30,960 Speaker 1: that successful outcome. Did your decision at step two set 186 00:12:31,000 --> 00:12:33,560 Speaker 1: you on the right path, or was your choice at 187 00:12:33,559 --> 00:12:36,920 Speaker 1: step three so good that it helped correct a mistake 188 00:12:37,160 --> 00:12:39,920 Speaker 1: that you made it step two. But a good approach 189 00:12:39,960 --> 00:12:43,480 Speaker 1: to machine learning involves a system that can adjust things 190 00:12:43,520 --> 00:12:47,160 Speaker 1: on its own to reduce mistakes and increase the success rate. 191 00:12:47,520 --> 00:12:50,040 Speaker 1: And another way of putting it is that instead of 192 00:12:50,080 --> 00:12:53,720 Speaker 1: programming a system to arrive at a specific outcome, you 193 00:12:53,800 --> 00:12:57,160 Speaker 1: are training the system to learn how to do it 194 00:12:57,240 --> 00:13:00,520 Speaker 1: by itself. And that sounds a bit magical when you 195 00:13:00,559 --> 00:13:03,760 Speaker 1: put it that way, doesn't it. It sounds like someone 196 00:13:03,840 --> 00:13:06,880 Speaker 1: just took a computer and showed it pictures of cats 197 00:13:07,080 --> 00:13:09,640 Speaker 1: and then expected the computer to know what a cat was. 198 00:13:10,440 --> 00:13:13,840 Speaker 1: And this actually does mirror an actual project that really 199 00:13:14,240 --> 00:13:17,880 Speaker 1: did do that, But I'm leaving out some big important 200 00:13:17,880 --> 00:13:22,200 Speaker 1: information in the middle. Now. One big step is that 201 00:13:22,240 --> 00:13:26,520 Speaker 1: computers and machines can't just magically learn by default. People 202 00:13:26,600 --> 00:13:29,840 Speaker 1: first had to come up with a methodology that allows 203 00:13:29,920 --> 00:13:32,560 Speaker 1: machines to go through the process of completing a task, 204 00:13:33,200 --> 00:13:36,960 Speaker 1: then making adjustments to the process of doing that task, 205 00:13:37,360 --> 00:13:40,880 Speaker 1: which would then improve future results. We have to lay 206 00:13:40,880 --> 00:13:45,440 Speaker 1: the groundwork in architecture and theory and algorithms. We have 207 00:13:45,520 --> 00:13:49,600 Speaker 1: to build the logical pathways that computers can follow in 208 00:13:49,720 --> 00:13:52,720 Speaker 1: order for them to learn. A lot of machine learning 209 00:13:53,120 --> 00:13:57,360 Speaker 1: revolves around patterns and pattern recognition. So what do I 210 00:13:57,400 --> 00:14:01,400 Speaker 1: mean by patterns? Well, I mean some form of regularity 211 00:14:01,480 --> 00:14:06,800 Speaker 1: and predictability. Machine learning models analyze patterns and attempt to 212 00:14:06,880 --> 00:14:11,640 Speaker 1: draw conclusions based on those patterns. This in itself is 213 00:14:11,640 --> 00:14:15,720 Speaker 1: tricky stuff. So why is that, Well, it's because sometimes 214 00:14:15,960 --> 00:14:19,960 Speaker 1: we might think there's a pattern, when in reality there 215 00:14:20,080 --> 00:14:25,480 Speaker 1: is not. We humans are pretty good at recognizing patterns, 216 00:14:25,680 --> 00:14:29,480 Speaker 1: which makes sense. It's a survival mechanism. If you were 217 00:14:29,520 --> 00:14:33,160 Speaker 1: to look at tall grass and you see patterns that 218 00:14:33,240 --> 00:14:37,320 Speaker 1: suggest the presence of a predator like a tiger, well 219 00:14:37,440 --> 00:14:40,520 Speaker 1: you would know that danger is nearby, and you would 220 00:14:40,520 --> 00:14:43,560 Speaker 1: have the opportunity to do something about that to help 221 00:14:43,600 --> 00:14:48,960 Speaker 1: your chances of survival. If, however, you remained blissfully unaware 222 00:14:49,080 --> 00:14:51,960 Speaker 1: of the danger, you'd be far more likely to fall 223 00:14:52,000 --> 00:14:55,920 Speaker 1: prey to that hungry tiger. So recognizing patterns is one 224 00:14:55,920 --> 00:14:58,760 Speaker 1: of the abilities that gave humans a chance to live 225 00:14:58,800 --> 00:15:02,440 Speaker 1: another day, and, from an evolutionary standpoint, a chance to 226 00:15:02,800 --> 00:15:07,680 Speaker 1: make more humans. But sometimes we humans will perceive a 227 00:15:07,720 --> 00:15:12,920 Speaker 1: pattern where none actually exists. A simple example of this 228 00:15:13,080 --> 00:15:16,960 Speaker 1: is the fun exercise of laying on your back outside, 229 00:15:17,360 --> 00:15:20,200 Speaker 1: looking up at the clouds and saying, what does that 230 00:15:20,240 --> 00:15:23,960 Speaker 1: cloud remind you of? The shapes of clouds? Which have 231 00:15:24,560 --> 00:15:28,480 Speaker 1: no significance and are the product of environmental factors, can 232 00:15:28,600 --> 00:15:32,600 Speaker 1: seem to suggest patterns to us. We might see a dog, 233 00:15:32,840 --> 00:15:36,120 Speaker 1: or a car or a face, but we know that 234 00:15:36,280 --> 00:15:40,360 Speaker 1: what we're really seeing with just the appearance of a pattern, 235 00:15:40,440 --> 00:15:43,360 Speaker 1: it's it's not evidence of a pattern actually being there. 236 00:15:43,400 --> 00:15:50,040 Speaker 1: It's noise, not signal, but it could be misinterpreted as signal. Well, 237 00:15:50,080 --> 00:15:53,000 Speaker 1: it turns out that in machine learning applications this is 238 00:15:53,080 --> 00:15:55,520 Speaker 1: also an issue. I'll talk about it more towards the 239 00:15:55,600 --> 00:15:59,800 Speaker 1: end of this episode. Computers can sometimes misinterpret data and 240 00:16:00,080 --> 00:16:04,000 Speaker 1: termines something represents a pattern when it really doesn't. When 241 00:16:04,040 --> 00:16:07,000 Speaker 1: that happens, a system relying on machine learning can produce 242 00:16:07,080 --> 00:16:11,480 Speaker 1: false positives, and the consequences can sometimes be funny, like hey, 243 00:16:11,520 --> 00:16:14,320 Speaker 1: this image recognition software thinks this coffee mug is actually 244 00:16:14,360 --> 00:16:17,320 Speaker 1: a kiddie cat, or they can be really serious and 245 00:16:17,360 --> 00:16:22,440 Speaker 1: potentially harmful. Hey, this facial recognition software has misidentified a person, 246 00:16:22,720 --> 00:16:25,640 Speaker 1: marking them as, say, a person of interest in a 247 00:16:25,680 --> 00:16:29,080 Speaker 1: criminal case. And it's all because this facial recognition software 248 00:16:29,120 --> 00:16:32,560 Speaker 1: isn't very good at differentiating people of color. That's a 249 00:16:32,680 --> 00:16:36,520 Speaker 1: real problem that really happens. Now, when we come back 250 00:16:36,800 --> 00:16:40,400 Speaker 1: I'll give a little overview of the evolution of machine learning, 251 00:16:40,880 --> 00:16:44,200 Speaker 1: but before we do that, let's take a quick break 252 00:16:51,840 --> 00:16:55,320 Speaker 1: to talk about the history of machine learning. We first 253 00:16:55,360 --> 00:16:59,120 Speaker 1: have to look back much much earlier, long before the 254 00:16:59,160 --> 00:17:02,880 Speaker 1: era of computers, and talk about how thinkers like Thomas 255 00:17:02,960 --> 00:17:07,600 Speaker 1: Bayes thought about the act of problem solving. Bays was 256 00:17:07,680 --> 00:17:11,240 Speaker 1: born way back in two so quite a bit before 257 00:17:11,280 --> 00:17:14,480 Speaker 1: we were thinking about machine learning, but he was interested 258 00:17:14,600 --> 00:17:19,560 Speaker 1: in problem solving for problems involving probabilities, and specifically the 259 00:17:19,600 --> 00:17:24,000 Speaker 1: relationship between different probabilities. I think it's easier to talk 260 00:17:24,040 --> 00:17:27,520 Speaker 1: about if I give you an example. So let's make 261 00:17:27,560 --> 00:17:30,320 Speaker 1: a silly one, all right, So let's say we got 262 00:17:30,320 --> 00:17:35,440 Speaker 1: ourselves a plucky podcaster. Hey there, everybody, It's Jonathan Strickland, 263 00:17:36,080 --> 00:17:39,600 Speaker 1: and it's Tuesday as I record this, And because of 264 00:17:39,760 --> 00:17:43,199 Speaker 1: who I am, you know who this podcaster is. And 265 00:17:43,280 --> 00:17:47,480 Speaker 1: because it's Tuesday, there is a chance I am wearing 266 00:17:47,640 --> 00:17:51,159 Speaker 1: a they might Be Giants T shirt. And we also 267 00:17:51,240 --> 00:17:55,760 Speaker 1: know that if this podcaster is wearing a they might 268 00:17:55,800 --> 00:17:59,879 Speaker 1: be Giants T shirt on a Tuesday, there's a sixty 269 00:18:00,119 --> 00:18:03,440 Speaker 1: percent chance that I'm going to end up wearing pajamas 270 00:18:03,520 --> 00:18:06,960 Speaker 1: on Wednesday. But we also know that if I did 271 00:18:07,080 --> 00:18:11,280 Speaker 1: not where they Might be Giants shirt on Tuesday, and 272 00:18:11,400 --> 00:18:15,280 Speaker 1: remember there's a sixty chance I didn't, then we know 273 00:18:15,400 --> 00:18:17,920 Speaker 1: there's an eighty percent chance I'm going to be wearing 274 00:18:17,960 --> 00:18:22,240 Speaker 1: pajamas on Wednesday. Well, Bays worked out a way that 275 00:18:22,320 --> 00:18:28,040 Speaker 1: described this sort of probability relationship between different discrete events, 276 00:18:28,200 --> 00:18:32,000 Speaker 1: and using his reasoning, you can work forward or backward 277 00:18:32,000 --> 00:18:35,959 Speaker 1: based on probabilities. Bays would describe wearing a they Might 278 00:18:36,000 --> 00:18:39,320 Speaker 1: be Giant shirt on Tuesday as one event and wearing 279 00:18:39,320 --> 00:18:43,600 Speaker 1: pajamas on Wednesday as a separate event, and then describe 280 00:18:43,640 --> 00:18:46,400 Speaker 1: the two not only determining how likely it is I'll 281 00:18:46,400 --> 00:18:50,720 Speaker 1: wear pajamas on Wednesday, but if we start with the 282 00:18:50,920 --> 00:18:53,320 Speaker 1: later event. In other words, if we start with the 283 00:18:53,359 --> 00:18:57,240 Speaker 1: fact that it's Wednesday and I'm wearing pajamas, we could 284 00:18:57,240 --> 00:19:02,120 Speaker 1: work out how likely it was that yesterday, on Tuesday, 285 00:19:02,200 --> 00:19:05,439 Speaker 1: I was wearing they Might Be Giants shirt. That was 286 00:19:05,560 --> 00:19:08,000 Speaker 1: his his contribution, that you can work this in either 287 00:19:08,119 --> 00:19:11,679 Speaker 1: direction if you know these different variables. Now, Bay has 288 00:19:11,760 --> 00:19:15,240 Speaker 1: never published his thoughts, but rather send an essay explaining 289 00:19:15,280 --> 00:19:18,040 Speaker 1: it to a friend of his, who then made sure 290 00:19:18,080 --> 00:19:20,840 Speaker 1: that The work was published after Bays had passed away, 291 00:19:20,880 --> 00:19:25,040 Speaker 1: and a few decades later Pierre Simon Laplace would take 292 00:19:25,119 --> 00:19:27,560 Speaker 1: this work that Bays had done and flesh it out 293 00:19:27,600 --> 00:19:32,280 Speaker 1: into an actual formal theorem. It's an important example of 294 00:19:32,320 --> 00:19:36,840 Speaker 1: conditional probability, and a lot of what machine learning is 295 00:19:37,640 --> 00:19:42,800 Speaker 1: really boiled down to is dealing with different probabilities, not certainties, which, 296 00:19:42,800 --> 00:19:44,119 Speaker 1: when you get down to it, is what most of 297 00:19:44,200 --> 00:19:46,120 Speaker 1: us are doing most of the time. Right We make 298 00:19:46,160 --> 00:19:51,480 Speaker 1: decisions based on at least perceived probabilities. Sometimes these decisions 299 00:19:51,520 --> 00:19:54,960 Speaker 1: might feel like they're a coin flip situation that any 300 00:19:55,040 --> 00:19:58,399 Speaker 1: choice is equally likely to precipitate a good outcome or 301 00:19:58,440 --> 00:20:01,399 Speaker 1: a bad outcome. Other times we might make a choice 302 00:20:01,400 --> 00:20:04,960 Speaker 1: because we feel the probabilities are stacked favorably one way 303 00:20:05,080 --> 00:20:08,840 Speaker 1: over another. Sometimes we will make a choice to back 304 00:20:08,960 --> 00:20:13,679 Speaker 1: the least probable outcome because well, humans are not always 305 00:20:13,720 --> 00:20:17,399 Speaker 1: superrational and hex sometimes the long shot does pay off, 306 00:20:17,560 --> 00:20:22,560 Speaker 1: so that keeps Vegas in business. Bayes' theorem is just 307 00:20:22,680 --> 00:20:26,000 Speaker 1: one example of ways that mathematicians and philosophers figured out 308 00:20:26,040 --> 00:20:31,280 Speaker 1: ways to mathematically express problem solving and decision making, and 309 00:20:31,320 --> 00:20:33,440 Speaker 1: a lot of this was figuring out if there were 310 00:20:33,520 --> 00:20:36,119 Speaker 1: a way to boil down things that most of us 311 00:20:36,119 --> 00:20:40,280 Speaker 1: approached through intuition and experience. So it's kind of neat, 312 00:20:40,880 --> 00:20:43,480 Speaker 1: and also the more you look into it, the more 313 00:20:43,560 --> 00:20:46,240 Speaker 1: likely you might find it's a little spooky, because it's 314 00:20:46,240 --> 00:20:49,639 Speaker 1: weird to consider that our approaches to making choices and 315 00:20:49,720 --> 00:20:55,399 Speaker 1: solving problems can be reduced down to mathematical expressions. But 316 00:20:56,440 --> 00:21:00,359 Speaker 1: let's leave the potential existential crises alone for now, shall we. 317 00:21:00,480 --> 00:21:03,920 Speaker 1: So moving on, we have another smarty pants we need 318 00:21:03,960 --> 00:21:08,479 Speaker 1: to talk about Andre Markov, a Russian mathematician. In the 319 00:21:08,560 --> 00:21:12,120 Speaker 1: early twentieth century. He began studying the nature of certain 320 00:21:12,240 --> 00:21:16,160 Speaker 1: random processes that follow a particular type of rule, which 321 00:21:16,160 --> 00:21:20,000 Speaker 1: we now call the Markov property. That rule says that 322 00:21:20,400 --> 00:21:24,600 Speaker 1: for this particular process, the next stage of the process 323 00:21:24,720 --> 00:21:29,120 Speaker 1: only depends upon the current stage, but not any stages 324 00:21:29,160 --> 00:21:33,520 Speaker 1: that came before then. So let's take my ridiculous T 325 00:21:33,720 --> 00:21:36,600 Speaker 1: shirt example, and let's build it out a little bit further. 326 00:21:37,000 --> 00:21:39,800 Speaker 1: Let's say that I've got three T shirts to my name. 327 00:21:40,200 --> 00:21:42,119 Speaker 1: One of them is that they might be Giant's shirt, 328 00:21:42,680 --> 00:21:46,160 Speaker 1: one is a plain blue T shirt, and the third 329 00:21:46,480 --> 00:21:49,240 Speaker 1: is a shirt that has the tech stuff logo on it, 330 00:21:49,800 --> 00:21:54,959 Speaker 1: and it's based off of long observation that you've determined 331 00:21:55,280 --> 00:21:59,680 Speaker 1: these following facts. If I am wearing that they Might 332 00:21:59,680 --> 00:22:04,399 Speaker 1: Be shirt today, I definitely will not wear it tomorrow. 333 00:22:04,800 --> 00:22:08,280 Speaker 1: But there's a fifty shot I'll wear either the blue 334 00:22:08,280 --> 00:22:12,080 Speaker 1: shirt or the tech Stuff shirt. Now, if I'm wearing 335 00:22:12,280 --> 00:22:15,800 Speaker 1: the blue shirt today, there's a ten percent chance I'm 336 00:22:15,800 --> 00:22:19,280 Speaker 1: going to wear the same blue shirt tomorrow. Don't worry, 337 00:22:19,520 --> 00:22:23,600 Speaker 1: I'll wash it first. There's a sixty chance that I'll 338 00:22:23,600 --> 00:22:26,320 Speaker 1: wear the tech Stuff shirt, and there's a thirty percent 339 00:22:26,400 --> 00:22:29,600 Speaker 1: chance I'll wear the they Might Be Giant shirt. But 340 00:22:30,520 --> 00:22:33,159 Speaker 1: if I'm wearing the tech Stuff shirt today, there's a 341 00:22:33,200 --> 00:22:36,399 Speaker 1: seventy chance I'll wear it again tomorrow because I like 342 00:22:36,440 --> 00:22:39,760 Speaker 1: to promote myself. But there's a thirty percent chance I'll 343 00:22:39,760 --> 00:22:42,159 Speaker 1: wear the they Might Be Giant shirt, and there is 344 00:22:42,280 --> 00:22:44,920 Speaker 1: no chance that I'm going to wear the blue one 345 00:22:45,240 --> 00:22:49,520 Speaker 1: in this case. So those are our various scenarios. Right 346 00:22:49,800 --> 00:22:54,560 Speaker 1: which shirt I will wear tomorrow depends only upon which 347 00:22:54,640 --> 00:22:58,120 Speaker 1: shirt I am wearing today. What I wore yesterday has 348 00:22:58,119 --> 00:23:02,119 Speaker 1: no bearing on the outcome for tomorrow, So today is 349 00:23:02,160 --> 00:23:05,879 Speaker 1: all that matters. And depending on which shirt I wear, 350 00:23:06,320 --> 00:23:09,639 Speaker 1: you can make some probability predictions for tomorrow. So we 351 00:23:09,640 --> 00:23:12,600 Speaker 1: can actually use this approach to figure out the probability 352 00:23:12,640 --> 00:23:15,840 Speaker 1: that I might wear the tech Stuff shirts, say ten 353 00:23:15,920 --> 00:23:19,119 Speaker 1: days in a row, since there's a better than even 354 00:23:19,240 --> 00:23:22,760 Speaker 1: chance that if I'm wearing tech stuff today, I'll end 355 00:23:22,840 --> 00:23:26,000 Speaker 1: up wearing it again tomorrow. And if I wear it tomorrow, 356 00:23:26,240 --> 00:23:28,879 Speaker 1: then there's a better than fifty chance that I'm going 357 00:23:28,920 --> 00:23:32,639 Speaker 1: to wear it the following day. But at some point 358 00:23:32,720 --> 00:23:35,880 Speaker 1: you're going to see that the odds are starting to 359 00:23:35,960 --> 00:23:40,320 Speaker 1: be against you, for you know, increasingly long strings of 360 00:23:40,400 --> 00:23:44,000 Speaker 1: wearing the tech Stuff shirt. Anyway, Markov chains would become 361 00:23:44,040 --> 00:23:46,920 Speaker 1: one of the types of processes that machine learning models 362 00:23:46,960 --> 00:23:50,520 Speaker 1: would incorporate, with some models looking at the current state 363 00:23:50,600 --> 00:23:53,639 Speaker 1: of a given process and then make predictions on what 364 00:23:53,920 --> 00:23:57,399 Speaker 1: the next state will be with no need to look 365 00:23:57,560 --> 00:24:03,320 Speaker 1: back at the previous decision. The Markov chain is memory less. 366 00:24:04,400 --> 00:24:07,680 Speaker 1: Now that's just a couple of the mathematicians whose work 367 00:24:07,840 --> 00:24:12,159 Speaker 1: underlies elements of machine learning. There's also structure we need 368 00:24:12,200 --> 00:24:15,880 Speaker 1: to talk about. In ninety nine, a man named Donald 369 00:24:15,920 --> 00:24:19,520 Speaker 1: Hebb wrote a book titled The Organization of Behavior, and 370 00:24:19,600 --> 00:24:24,200 Speaker 1: in that book, Hebb gave a hypothesis on how neurons, 371 00:24:24,480 --> 00:24:27,879 Speaker 1: that is, how how brain cells interact with one another. 372 00:24:28,440 --> 00:24:32,480 Speaker 1: His ideas included the notion that if two neurons interact 373 00:24:32,520 --> 00:24:36,760 Speaker 1: with one another regularly, that is, if one fires, that 374 00:24:36,880 --> 00:24:40,440 Speaker 1: the second one is also likely to fire. They end 375 00:24:40,520 --> 00:24:44,959 Speaker 1: up forming a tighter communicative relationship with each other. Not 376 00:24:45,160 --> 00:24:50,320 Speaker 1: long after his expression of this hypothesis, computer scientists began 377 00:24:50,359 --> 00:24:53,000 Speaker 1: to think of a potential way to do this artificially, 378 00:24:53,400 --> 00:24:59,120 Speaker 1: with machines creating the equivalent of artificial neurons. The relative 379 00:24:59,280 --> 00:25:04,080 Speaker 1: strength and relationship between artificial neurons is something we described 380 00:25:04,119 --> 00:25:07,520 Speaker 1: by wait that's going to be an important part of 381 00:25:07,560 --> 00:25:11,439 Speaker 1: machine learning. WIT. By the way, is W E I 382 00:25:11,720 --> 00:25:15,640 Speaker 1: G H T, as in this relationship is weighted more 383 00:25:15,720 --> 00:25:21,040 Speaker 1: heavily than that relationship. In the early nineteen fifties, and 384 00:25:21,200 --> 00:25:25,080 Speaker 1: IBM researcher named Arthur Samuel created a program designed to 385 00:25:25,119 --> 00:25:28,399 Speaker 1: win at checkers. The program would do a quick analysis 386 00:25:28,440 --> 00:25:32,680 Speaker 1: of where pieces were on a checkerboard and whose move 387 00:25:32,720 --> 00:25:36,080 Speaker 1: it was, and then calculate the chances of each side 388 00:25:36,080 --> 00:25:38,960 Speaker 1: winning the game based on those positions, and it did 389 00:25:39,000 --> 00:25:43,119 Speaker 1: this with a mini max approach. Alright, so checkers is 390 00:25:43,160 --> 00:25:46,840 Speaker 1: a two player turn based game. Player one makes a move, 391 00:25:47,160 --> 00:25:49,399 Speaker 1: then player two can make a move. There are a 392 00:25:49,440 --> 00:25:52,840 Speaker 1: finite number of moves that can be made, a finite 393 00:25:52,960 --> 00:25:57,159 Speaker 1: number of possibilities, though admittedly it's a pretty good number 394 00:25:57,200 --> 00:26:00,520 Speaker 1: of possibilities. But let's say a game has been going 395 00:26:00,520 --> 00:26:03,439 Speaker 1: on for a few moves, and you've got your two sides. 396 00:26:03,480 --> 00:26:06,040 Speaker 1: You've got the red checkers over on player one side 397 00:26:06,160 --> 00:26:08,880 Speaker 1: and the black checkers for a player to Let's say 398 00:26:08,880 --> 00:26:12,080 Speaker 1: it's player one's move. For the purposes of this example, 399 00:26:12,400 --> 00:26:15,040 Speaker 1: will say that player one really just has one piece 400 00:26:15,200 --> 00:26:19,119 Speaker 1: that they can actually move on this turn, and it 401 00:26:19,160 --> 00:26:23,480 Speaker 1: can move into one of two open spaces. So player 402 00:26:23,520 --> 00:26:26,760 Speaker 1: one has to make a choice. After that choice, it's 403 00:26:26,760 --> 00:26:29,800 Speaker 1: going to be player two's turn. So we can create 404 00:26:29,840 --> 00:26:34,360 Speaker 1: a decision tree illustrating the possible choices and the possible 405 00:26:34,400 --> 00:26:38,639 Speaker 1: outcomes of those choices. These choices are the children of 406 00:26:38,680 --> 00:26:42,040 Speaker 1: the starting position for player one, so player one's starting 407 00:26:42,040 --> 00:26:46,120 Speaker 1: position has two children. Player too will have their own 408 00:26:46,200 --> 00:26:49,520 Speaker 1: choices to make after that decision has been made, but 409 00:26:49,840 --> 00:26:53,240 Speaker 1: those choices are going to depend upon whatever move player 410 00:26:53,280 --> 00:26:57,360 Speaker 1: one ultimately takes. So we can extend out our decision 411 00:26:57,440 --> 00:27:01,920 Speaker 1: tree showing the branching possible move that Player Too might make, 412 00:27:02,480 --> 00:27:05,879 Speaker 1: and these are the children of the two possible outcomes 413 00:27:05,920 --> 00:27:10,160 Speaker 1: of our first choice. After player two's turn, it's player 414 00:27:10,240 --> 00:27:14,000 Speaker 1: ones turn again, which means we need to branch those 415 00:27:14,040 --> 00:27:17,720 Speaker 1: decisions out even further. And this is all before player 416 00:27:17,800 --> 00:27:22,560 Speaker 1: one has even made that first choice. We're just evaluating possibilities. 417 00:27:22,840 --> 00:27:25,560 Speaker 1: At some point, either when we have plotted far enough 418 00:27:25,600 --> 00:27:28,840 Speaker 1: out that we know all possible outcomes of the game, 419 00:27:29,520 --> 00:27:32,399 Speaker 1: or we're just reaching a point where it would be 420 00:27:32,520 --> 00:27:35,399 Speaker 1: unmanageable for us to go any further. We need to 421 00:27:35,440 --> 00:27:40,160 Speaker 1: actually analyze what our options are. The endpoints represent either 422 00:27:40,720 --> 00:27:45,119 Speaker 1: a win, a loss, or a draw for player one, or, 423 00:27:45,160 --> 00:27:48,320 Speaker 1: if we haven't extended out the tree all the way 424 00:27:48,359 --> 00:27:50,600 Speaker 1: to the end of the game, at least a change 425 00:27:50,600 --> 00:27:54,160 Speaker 1: in advantage, whether it would be in player one's advantage 426 00:27:54,200 --> 00:27:58,119 Speaker 1: to make that move or disadvantage. We could actually assign 427 00:27:58,240 --> 00:28:01,760 Speaker 1: numerical values to each in the point, with positive values 428 00:28:01,840 --> 00:28:05,600 Speaker 1: representing an advantage for player one and a negative value 429 00:28:05,640 --> 00:28:09,080 Speaker 1: representing an advantage for Player Too. And once we do that, 430 00:28:09,480 --> 00:28:12,600 Speaker 1: we can see which pathways tend to lead to better 431 00:28:12,640 --> 00:28:17,399 Speaker 1: outcomes for Player one. We work backward through the decision tree. 432 00:28:17,840 --> 00:28:21,800 Speaker 1: So on all the decisions that end in an advantage 433 00:28:21,800 --> 00:28:24,760 Speaker 1: for player one, we can say this is the choice 434 00:28:24,760 --> 00:28:28,119 Speaker 1: that player one would take. But then we know that 435 00:28:28,240 --> 00:28:31,200 Speaker 1: for player to player two is always going to choose 436 00:28:31,520 --> 00:28:35,640 Speaker 1: whichever choice has the grace advantage for that player. So 437 00:28:36,040 --> 00:28:38,160 Speaker 1: we have to actually take that into account as we're 438 00:28:38,200 --> 00:28:43,400 Speaker 1: working backward, and this is how we can finally get 439 00:28:43,440 --> 00:28:45,520 Speaker 1: to the point where we decide which move we're going 440 00:28:45,560 --> 00:28:48,920 Speaker 1: to make, because these decisions, as you go backward up 441 00:28:48,960 --> 00:28:53,040 Speaker 1: the tree, they ultimately inform you which of those two 442 00:28:53,160 --> 00:28:58,040 Speaker 1: choices is going to give you the best result. Those values, well, 443 00:28:58,160 --> 00:29:01,040 Speaker 1: those are weights. So for player one, the goal is 444 00:29:01,080 --> 00:29:04,680 Speaker 1: to pick the path that has the highest positive value. 445 00:29:04,800 --> 00:29:07,440 Speaker 1: For player too, it's to pick the path that has 446 00:29:07,480 --> 00:29:11,080 Speaker 1: the lowest possible value or the highest negative value, if 447 00:29:11,080 --> 00:29:13,560 Speaker 1: you prefer. So. In other words, player one might be 448 00:29:13,600 --> 00:29:16,720 Speaker 1: thinking something like, if I moved to Spot A, my 449 00:29:16,840 --> 00:29:19,800 Speaker 1: chance of winning this game is but if I moved 450 00:29:19,800 --> 00:29:24,480 Speaker 1: to Spot B, it's only so. Of course, those percentages 451 00:29:24,520 --> 00:29:26,600 Speaker 1: will also depend on what player two is going to 452 00:29:26,640 --> 00:29:29,240 Speaker 1: do in response. Some moves that Player Too might do 453 00:29:29,440 --> 00:29:33,000 Speaker 1: could end up guaranteeing a win for player one. This 454 00:29:33,160 --> 00:29:35,920 Speaker 1: is the mini max approach, and there's an algorithm that 455 00:29:36,000 --> 00:29:39,320 Speaker 1: guides it. It depends upon the current position within a 456 00:29:39,400 --> 00:29:43,120 Speaker 1: game and how many moves or how much depth it 457 00:29:43,160 --> 00:29:46,200 Speaker 1: has to take into account, and for which player is 458 00:29:46,240 --> 00:29:50,760 Speaker 1: it actually helping out. What happens is if player one 459 00:29:50,880 --> 00:29:55,040 Speaker 1: does this evaluation and finds that both options are negative, well, 460 00:29:55,240 --> 00:29:58,360 Speaker 1: then this is something that happens in games. Right. Sometimes 461 00:29:58,440 --> 00:30:01,360 Speaker 1: you find out there is no good move, like any 462 00:30:01,440 --> 00:30:03,640 Speaker 1: move you make is going to be a losing move. Well, 463 00:30:03,640 --> 00:30:05,800 Speaker 1: the only option at that point is to choose the 464 00:30:05,920 --> 00:30:09,200 Speaker 1: least bad one, so it would be whatever the smallest 465 00:30:09,240 --> 00:30:13,240 Speaker 1: negative value choice was. Our next big development that I 466 00:30:13,280 --> 00:30:18,880 Speaker 1: need to mention is Frank Rosenblatt's artificial neural network called Perceptron. 467 00:30:19,560 --> 00:30:22,880 Speaker 1: Its purpose was to recognize shapes and patterns, and it 468 00:30:22,920 --> 00:30:26,480 Speaker 1: was originally going to be its own machine like actual hardware, 469 00:30:27,000 --> 00:30:30,480 Speaker 1: but the first incarnation of Perceptron would actually be in 470 00:30:30,520 --> 00:30:33,840 Speaker 1: the form of software rather than hardware. There was a 471 00:30:33,880 --> 00:30:37,480 Speaker 1: purpose built Perceptron later, but the original one was software. 472 00:30:37,960 --> 00:30:41,960 Speaker 1: Despite some early excitement, the Perceptron proved to be somewhat 473 00:30:42,080 --> 00:30:46,000 Speaker 1: limited in its capabilities and interest in artificial neural networks 474 00:30:46,040 --> 00:30:49,560 Speaker 1: died down for a while as a result. In a way, 475 00:30:50,200 --> 00:30:53,200 Speaker 1: you could kind of compare this to some other technologies 476 00:30:53,200 --> 00:30:56,640 Speaker 1: that got a big hype cycle and then later deflated. 477 00:30:57,040 --> 00:31:00,120 Speaker 1: Virtual reality is the one I always go with. Back 478 00:31:00,120 --> 00:31:02,720 Speaker 1: in the nineteen nineties, the world was really hyped for 479 00:31:02,800 --> 00:31:08,120 Speaker 1: virtual reality. People had incredibly unrealistic expectations for what VR 480 00:31:08,320 --> 00:31:11,200 Speaker 1: actually meant and what it could do. And when it 481 00:31:11,200 --> 00:31:14,600 Speaker 1: turned out the VR wasn't nearly as sophisticated as people 482 00:31:14,600 --> 00:31:18,400 Speaker 1: were imagining, a lot of enthusiasm dropped out for the 483 00:31:18,600 --> 00:31:23,040 Speaker 1: entire field, and with that dropped funding and support, and 484 00:31:23,080 --> 00:31:26,120 Speaker 1: as a result, development and VR hit a real wall, 485 00:31:26,200 --> 00:31:29,040 Speaker 1: with only a fraction of the people who had been 486 00:31:29,080 --> 00:31:32,400 Speaker 1: working in the field sticking around, and they had to 487 00:31:32,440 --> 00:31:35,320 Speaker 1: scramble just to find funding to keep their projects going. 488 00:31:35,720 --> 00:31:38,280 Speaker 1: So VR was effectively put on the shelf and wouldn't 489 00:31:38,320 --> 00:31:42,160 Speaker 1: make much progress for nearly twenty years. Well. Artificial neural 490 00:31:42,200 --> 00:31:46,959 Speaker 1: networks had a very similar issue, but other computer scientists 491 00:31:47,280 --> 00:31:51,000 Speaker 1: eventually found ways to design artificial neural networks. They could 492 00:31:51,040 --> 00:31:54,520 Speaker 1: do some pretty amazing things if they had access to 493 00:31:54,680 --> 00:31:57,960 Speaker 1: enough data. When we come back, i'll talk a little 494 00:31:57,960 --> 00:32:00,440 Speaker 1: bit more about that and what it all means. But 495 00:32:00,560 --> 00:32:11,800 Speaker 1: first let's take another quick break. So we left off 496 00:32:11,880 --> 00:32:15,080 Speaker 1: with the AI field going into hibernation for a little bit. 497 00:32:15,640 --> 00:32:20,040 Speaker 1: Theory and mathematics were bumping up against the limitations of technology, 498 00:32:20,200 --> 00:32:23,160 Speaker 1: which wasn't quite at the level to put all that 499 00:32:23,280 --> 00:32:26,160 Speaker 1: theory to the test. Plus there needed to be some 500 00:32:26,200 --> 00:32:30,120 Speaker 1: tweaks to the approaches, but those came with time and 501 00:32:30,400 --> 00:32:34,440 Speaker 1: more mathematicians found new ways to create artificial neural networks 502 00:32:34,480 --> 00:32:38,880 Speaker 1: capable of stuff like pattern recognition and learning. So let's 503 00:32:39,040 --> 00:32:44,040 Speaker 1: imagine another decision tree. We've got our starting position. This 504 00:32:44,080 --> 00:32:47,200 Speaker 1: is probably where we put some input. We would feed 505 00:32:47,640 --> 00:32:51,920 Speaker 1: data into a system, and let's say from that starting position, 506 00:32:51,960 --> 00:32:55,160 Speaker 1: we have a process that's going to transform that input 507 00:32:55,760 --> 00:32:59,520 Speaker 1: into one of two possible ways. So we've got two 508 00:33:00,040 --> 00:33:05,120 Speaker 1: potential outputs for that first step. Like our mini max example, 509 00:33:05,440 --> 00:33:08,800 Speaker 1: we can go down several layers of possible choices, and 510 00:33:08,840 --> 00:33:12,640 Speaker 1: we can wait the relationships between these different choices. So 511 00:33:13,080 --> 00:33:16,280 Speaker 1: if the incoming value is higher than a certain amount, 512 00:33:16,640 --> 00:33:19,800 Speaker 1: maybe the node sends it down one pathway, but if 513 00:33:19,880 --> 00:33:23,479 Speaker 1: the value is lower than that arbitrary amount, the node 514 00:33:23,600 --> 00:33:28,760 Speaker 1: will send it down a different pathway. This is drastically oversimplifying, 515 00:33:28,920 --> 00:33:31,240 Speaker 1: but I hope you kind of get the idea. It's 516 00:33:31,240 --> 00:33:34,400 Speaker 1: like a big sorting system and the goal is that 517 00:33:34,800 --> 00:33:38,760 Speaker 1: at the very end, whatever comes out as output is 518 00:33:38,880 --> 00:33:43,320 Speaker 1: correct or true. Ideally, you've got a system that is 519 00:33:43,480 --> 00:33:48,760 Speaker 1: self improving. It trains itself to be better. But how 520 00:33:48,800 --> 00:33:52,440 Speaker 1: the heck does that happen? Well, let's consider cats for 521 00:33:52,480 --> 00:33:57,760 Speaker 1: a bit, not the musical and could heavens definitely not 522 00:33:58,200 --> 00:34:02,760 Speaker 1: the movie music a coal that is a subject that 523 00:34:02,960 --> 00:34:05,760 Speaker 1: deserves its own episode. Maybe one day I'll figure out 524 00:34:06,000 --> 00:34:08,120 Speaker 1: a way to tackle that film with some sort of 525 00:34:08,160 --> 00:34:11,120 Speaker 1: tech capacity, but honestly, I'm just not ready to do 526 00:34:11,200 --> 00:34:14,560 Speaker 1: that yet, from like an emotional standpoint as well as 527 00:34:14,560 --> 00:34:19,040 Speaker 1: a research one. No, let's say you're teaching a computer 528 00:34:19,080 --> 00:34:23,720 Speaker 1: system to recognize cats pictures of cats, and the system 529 00:34:23,760 --> 00:34:27,840 Speaker 1: has an artificial neural network that accepts input pictures of 530 00:34:27,880 --> 00:34:31,400 Speaker 1: cats and then filters that input through the network to 531 00:34:31,520 --> 00:34:35,399 Speaker 1: make the determination does this picture include a cat in it? 532 00:34:35,920 --> 00:34:38,880 Speaker 1: And you start feeding it lots of images. The neural 533 00:34:38,920 --> 00:34:42,359 Speaker 1: network acts on the data according to the weighted relationship 534 00:34:42,520 --> 00:34:47,480 Speaker 1: between the artificial neurons, and it produces an output. Now 535 00:34:47,800 --> 00:34:50,640 Speaker 1: here's the thing we already know what we want the 536 00:34:50,640 --> 00:34:54,000 Speaker 1: output to be because we can recognize of a picture 537 00:34:54,000 --> 00:34:57,160 Speaker 1: has a cat inet or not. Maybe we've got one 538 00:34:57,239 --> 00:35:00,440 Speaker 1: thousand pictures. This is the training data we're going to 539 00:35:00,600 --> 00:35:03,960 Speaker 1: use for this machine learning process. We also know that 540 00:35:04,080 --> 00:35:06,719 Speaker 1: eight hundred of those pictures have a cat in them 541 00:35:06,760 --> 00:35:10,160 Speaker 1: and two hundred don't, so we know what we want 542 00:35:10,160 --> 00:35:13,160 Speaker 1: the results to be. We've got an artificial neural network 543 00:35:13,360 --> 00:35:16,759 Speaker 1: in which some neurons or nodes will accept input and 544 00:35:16,760 --> 00:35:19,440 Speaker 1: perform a function based on that input, and then the 545 00:35:19,480 --> 00:35:23,360 Speaker 1: weighted connections that neuron has to other neurons will determine 546 00:35:23,640 --> 00:35:26,480 Speaker 1: where it passes the information down until we get to 547 00:35:26,480 --> 00:35:29,759 Speaker 1: an output. And this happens until we get that conclusion. 548 00:35:30,440 --> 00:35:34,160 Speaker 1: So what happens if the computer's answer is wrong. One 549 00:35:34,239 --> 00:35:37,160 Speaker 1: if we feed those one thousand photos to it and 550 00:35:37,239 --> 00:35:40,239 Speaker 1: says only three hundred of them have cats in them. 551 00:35:40,480 --> 00:35:43,720 Speaker 1: While we have to go back and adjust those weighted 552 00:35:43,840 --> 00:35:48,719 Speaker 1: connections because clearly something didn't go right, the connections within 553 00:35:48,760 --> 00:35:53,000 Speaker 1: the network need to be readjusted. We would likely start 554 00:35:53,320 --> 00:35:57,200 Speaker 1: closest to our output and see which neurons seem to 555 00:35:57,239 --> 00:36:01,560 Speaker 1: contribute to the mistake, which which neurons were responsible, in 556 00:36:01,600 --> 00:36:04,160 Speaker 1: other words, for it to say, oh, only three these 557 00:36:04,200 --> 00:36:07,960 Speaker 1: pictures had cats in them. And then we would adjust 558 00:36:08,040 --> 00:36:11,719 Speaker 1: the weights, the incoming weights of connections to those neurons 559 00:36:12,360 --> 00:36:15,880 Speaker 1: in order to try and favor pathways that lead to 560 00:36:16,040 --> 00:36:19,480 Speaker 1: correct answers. Then we feed it the one thousand pictures 561 00:36:19,560 --> 00:36:22,719 Speaker 1: again and we look at those results. Then we do 562 00:36:22,840 --> 00:36:26,480 Speaker 1: this again and again and again, every time tweaking the 563 00:36:26,520 --> 00:36:31,239 Speaker 1: network a little bit so that it gets a bit better. Eventually, 564 00:36:31,520 --> 00:36:34,960 Speaker 1: when we have trained the system, we can start to 565 00:36:35,120 --> 00:36:39,720 Speaker 1: feed brand new data to the network, not the stuff 566 00:36:39,760 --> 00:36:43,640 Speaker 1: we've trained it on, but pictures that we and the 567 00:36:43,680 --> 00:36:47,160 Speaker 1: system have never seen before. And if our network is 568 00:36:47,200 --> 00:36:49,480 Speaker 1: a good one, if we have trained it well, it 569 00:36:49,520 --> 00:36:53,239 Speaker 1: will sort through these new photos and it will count 570 00:36:53,320 --> 00:36:56,320 Speaker 1: up the ones that have the cat pictures lickety split. 571 00:36:56,760 --> 00:37:00,839 Speaker 1: This approach is called supervised learning because it involves kind 572 00:37:00,840 --> 00:37:04,880 Speaker 1: of grading the network on its homework and then working 573 00:37:04,920 --> 00:37:08,759 Speaker 1: with it to get better. Heck, with the right algorithm, 574 00:37:08,760 --> 00:37:12,480 Speaker 1: a neural network can learn to recognize and differentiate patterns 575 00:37:12,960 --> 00:37:16,520 Speaker 1: even if we never explicitly told the system what it 576 00:37:16,600 --> 00:37:20,719 Speaker 1: was looking for. Google discovered this several years ago when 577 00:37:20,760 --> 00:37:25,040 Speaker 1: it fed several thousand YouTube videos to an enormous artificial 578 00:37:25,080 --> 00:37:29,359 Speaker 1: neural network. The system analyzed the videos that were fed 579 00:37:29,400 --> 00:37:33,520 Speaker 1: to it and gradually recognized patterns that represented different types 580 00:37:33,560 --> 00:37:39,160 Speaker 1: of stuff, like people or like cats, because there are 581 00:37:39,200 --> 00:37:42,520 Speaker 1: a lot of cat videos on YouTube, and the network 582 00:37:42,880 --> 00:37:45,120 Speaker 1: got to the point where it could identify an image 583 00:37:45,120 --> 00:37:48,960 Speaker 1: of a cat fairly reliably better than seventy of the time, 584 00:37:49,440 --> 00:37:53,239 Speaker 1: even though it was never told how to do that, 585 00:37:53,920 --> 00:37:57,839 Speaker 1: or it was never even told what a cat was. So, 586 00:37:57,880 --> 00:38:01,120 Speaker 1: as Google representatives put it, they said, it had to 587 00:38:01,239 --> 00:38:04,719 Speaker 1: invent the concept of a cat. It had to recognize 588 00:38:05,200 --> 00:38:09,719 Speaker 1: that cats are not the same as people, which I 589 00:38:09,719 --> 00:38:14,080 Speaker 1: think is a big slap in the face to some cats. Really, 590 00:38:14,760 --> 00:38:18,560 Speaker 1: what it said was that I recognized this particular pattern 591 00:38:18,600 --> 00:38:23,040 Speaker 1: of features, and I recognized that these other instances of 592 00:38:23,120 --> 00:38:26,839 Speaker 1: creatures that have a similar pattern seemed to match that, 593 00:38:27,040 --> 00:38:30,919 Speaker 1: and so I draw the conclusion that this instance of 594 00:38:30,960 --> 00:38:35,080 Speaker 1: a thing belongs with all these other instances of things 595 00:38:35,160 --> 00:38:39,640 Speaker 1: that are similar in characteristics. So this was more of 596 00:38:39,640 --> 00:38:43,439 Speaker 1: an example of unsupervised learning, and that the system, when 597 00:38:43,480 --> 00:38:46,640 Speaker 1: fed enough data, began to categorize stuff all on its 598 00:38:46,640 --> 00:38:50,680 Speaker 1: own through its own parameters. Now, one neat way that 599 00:38:50,719 --> 00:38:54,720 Speaker 1: computer scientists will train up systems for certain types of applications. 600 00:38:55,280 --> 00:39:00,840 Speaker 1: Is through a generative adversarial network, which I admit sounds 601 00:39:00,920 --> 00:39:03,719 Speaker 1: kind of sinister, doesn't it, And I mean it can be, 602 00:39:03,920 --> 00:39:07,719 Speaker 1: but it doesn't have to be. Essentially, you're using two 603 00:39:07,760 --> 00:39:11,480 Speaker 1: different artificial neural networks. One of the networks has a 604 00:39:11,520 --> 00:39:15,520 Speaker 1: specific job, it's to fool the other network. So the 605 00:39:15,520 --> 00:39:18,719 Speaker 1: other network's job is to detect attempts to fool it 606 00:39:19,080 --> 00:39:23,520 Speaker 1: versus legitimate data. So let's use an example. Let's say 607 00:39:23,560 --> 00:39:26,200 Speaker 1: you're trying to create a system that can make realistic 608 00:39:26,600 --> 00:39:33,040 Speaker 1: but entirely computer generated, that is, fabricated photographs of people. So, 609 00:39:33,080 --> 00:39:36,440 Speaker 1: in other words, these are computer generated images that don't 610 00:39:36,560 --> 00:39:40,239 Speaker 1: actually represent a real person at all. We've got one 611 00:39:40,320 --> 00:39:43,600 Speaker 1: artificial neural network, the generator, and its job is to 612 00:39:43,680 --> 00:39:49,399 Speaker 1: create images of people that can pass as real photographs. 613 00:39:49,760 --> 00:39:52,520 Speaker 1: Then we've got our other network, which is the discriminator. 614 00:39:52,840 --> 00:39:56,520 Speaker 1: This is trying to sort out real photos of actual 615 00:39:56,600 --> 00:40:02,600 Speaker 1: people from pictures that have been generated by the generative system. 616 00:40:02,640 --> 00:40:06,400 Speaker 1: And we put these two networks against each other. The 617 00:40:06,480 --> 00:40:10,279 Speaker 1: idea here is that both systems get better as they 618 00:40:10,320 --> 00:40:14,880 Speaker 1: test one another out. If the generator network is falling 619 00:40:14,920 --> 00:40:19,040 Speaker 1: behind because the discriminator can suss out the fakes too easily. Well, 620 00:40:19,080 --> 00:40:22,280 Speaker 1: then it's time to tweak some weights in that neural 621 00:40:22,320 --> 00:40:27,680 Speaker 1: network that are leading to dissatisfactory computer generated images and 622 00:40:27,719 --> 00:40:31,960 Speaker 1: try it again. But then, if the discriminator is starting 623 00:40:32,000 --> 00:40:36,160 Speaker 1: to miss fakes while, it's time to tweak the discriminator 624 00:40:36,200 --> 00:40:41,080 Speaker 1: network so it's better at spotting the false pictures. Not 625 00:40:41,200 --> 00:40:44,879 Speaker 1: Along the way, some pretty extraordinary stuff can happen. There 626 00:40:44,880 --> 00:40:50,120 Speaker 1: are photos of computer generated faces, not altered pictures, not 627 00:40:50,280 --> 00:40:54,920 Speaker 1: ones created by a human artist, but entirely composed by 628 00:40:54,960 --> 00:40:59,440 Speaker 1: a computer, and they can look absolutely realistic, complete with 629 00:40:59,480 --> 00:41:04,680 Speaker 1: consistent lighting and shadows. This is only after lots of 630 00:41:04,760 --> 00:41:09,000 Speaker 1: training sessions the networks learn what the giveaways are, like, 631 00:41:09,520 --> 00:41:12,920 Speaker 1: what is it that leads the discriminator to say, no, 632 00:41:13,480 --> 00:41:15,920 Speaker 1: this is a fake photo, and how can you fix that? 633 00:41:16,400 --> 00:41:19,399 Speaker 1: It reminds me a bit of how photo experts used 634 00:41:19,440 --> 00:41:22,919 Speaker 1: to point out really bad photoshop jobs and explaining how 635 00:41:23,160 --> 00:41:27,120 Speaker 1: certain elements like shadows or edges or whatever, we're a 636 00:41:27,239 --> 00:41:30,880 Speaker 1: dead giveaway that someone had altered an image. Well, similar 637 00:41:31,000 --> 00:41:35,160 Speaker 1: rules exist for generated images, and through training, the generator 638 00:41:35,200 --> 00:41:39,640 Speaker 1: gets better at making really convincing examples that don't fall 639 00:41:39,680 --> 00:41:42,560 Speaker 1: into the traps that would reveal it as a fake. 640 00:41:43,520 --> 00:41:47,360 Speaker 1: Over time, generative networks can get good enough to produce 641 00:41:47,360 --> 00:41:50,080 Speaker 1: stuff that would be very difficult for a human to 642 00:41:50,160 --> 00:41:54,000 Speaker 1: tell apart from the quote unquote real thing, and discriminators 643 00:41:54,040 --> 00:41:57,280 Speaker 1: can get good enough to detect fakes that would otherwise 644 00:41:57,360 --> 00:42:01,000 Speaker 1: pass human inspection. So an example of this is the 645 00:42:01,080 --> 00:42:05,520 Speaker 1: current ongoing battle with deep fakes. These are computer generated 646 00:42:05,640 --> 00:42:09,600 Speaker 1: videos that appear to be legit. If they're done well enough, 647 00:42:10,080 --> 00:42:12,680 Speaker 1: they can have famous people in them. Doesn't have to 648 00:42:12,680 --> 00:42:15,200 Speaker 1: be a famous person, but it can show a video 649 00:42:15,200 --> 00:42:18,799 Speaker 1: of someone doing something that they absolutely never did, but 650 00:42:19,160 --> 00:42:21,799 Speaker 1: according to the video, they did, and it can be 651 00:42:21,840 --> 00:42:25,200 Speaker 1: really convincing if it's done well. A good deep fake 652 00:42:25,520 --> 00:42:29,520 Speaker 1: can fool people if you aren't paying too much attention. 653 00:42:29,600 --> 00:42:33,680 Speaker 1: Some of the really good ones can pass pretty deep scrutiny. 654 00:42:33,760 --> 00:42:37,319 Speaker 1: So this requires researchers to come up with solutions that 655 00:42:37,360 --> 00:42:41,040 Speaker 1: are pretty subtle and beyond the average person's ability to replicate, 656 00:42:41,120 --> 00:42:44,880 Speaker 1: like looking at the reflections in the person's eyes and 657 00:42:44,920 --> 00:42:48,680 Speaker 1: whether or not they seem realistic or a computer generated. 658 00:42:48,760 --> 00:42:53,520 Speaker 1: But that really just represents another hurdle for the generative side. 659 00:42:53,800 --> 00:42:57,960 Speaker 1: So in other words, this is a seesaw approach, right. 660 00:42:58,840 --> 00:43:02,600 Speaker 1: It's creating face as on one side and detecting them 661 00:43:02,600 --> 00:43:05,080 Speaker 1: on the other side. It's something we see an artificial 662 00:43:05,120 --> 00:43:08,040 Speaker 1: intelligence in general. A similar story played out with the 663 00:43:08,080 --> 00:43:11,880 Speaker 1: old capture systems, where you know, we saw back and 664 00:43:11,920 --> 00:43:15,440 Speaker 1: forth between methods to try and weed out bots by 665 00:43:15,560 --> 00:43:19,960 Speaker 1: using capture images that only humans could really parse, and 666 00:43:20,000 --> 00:43:24,280 Speaker 1: then we saw improved bots that could analyze these images 667 00:43:24,320 --> 00:43:27,920 Speaker 1: and return correct results, which men it was necessary to 668 00:43:27,960 --> 00:43:31,080 Speaker 1: create more difficult captures. Eventually get to a point where 669 00:43:31,200 --> 00:43:34,200 Speaker 1: the captures are difficult enough where the average person can't 670 00:43:34,200 --> 00:43:36,040 Speaker 1: even pass them, and then you have to go to 671 00:43:36,080 --> 00:43:39,080 Speaker 1: a different method. We also see this play out in 672 00:43:39,120 --> 00:43:42,320 Speaker 1: the cyber security realm, where you might say, the thieves 673 00:43:42,360 --> 00:43:45,600 Speaker 1: get better at lock picking, and then security experts make 674 00:43:45,719 --> 00:43:50,719 Speaker 1: better locks, and the cycle just repeats endlessly. One thing 675 00:43:50,840 --> 00:43:54,720 Speaker 1: that has really fueled machine learning recently is the era 676 00:43:54,880 --> 00:43:58,480 Speaker 1: of big data. Being able to harvest information on a 677 00:43:58,680 --> 00:44:04,000 Speaker 1: truly massive scale provides the opportunity to feed that data 678 00:44:04,120 --> 00:44:09,120 Speaker 1: into various machine learning systems to search for meaning within 679 00:44:09,200 --> 00:44:13,560 Speaker 1: that data. These systems might scour the information to look 680 00:44:13,560 --> 00:44:18,120 Speaker 1: for stuff like criminal activity like financial crimes or the 681 00:44:18,160 --> 00:44:22,520 Speaker 1: attempt to move some money around from various criminal exploits. 682 00:44:22,760 --> 00:44:25,400 Speaker 1: Or it could be used to look for trends like 683 00:44:25,480 --> 00:44:29,279 Speaker 1: market trends, or it might be used to plot possible 684 00:44:29,360 --> 00:44:33,560 Speaker 1: spikes in COVID nineteen transmission where those might occur where 685 00:44:33,680 --> 00:44:37,560 Speaker 1: people should really be focusing their attention. But now we 686 00:44:37,680 --> 00:44:40,360 Speaker 1: got to think back on what I said earlier about 687 00:44:40,440 --> 00:44:44,000 Speaker 1: looking up at the sky and seeing shapes in the clouds. 688 00:44:45,000 --> 00:44:48,240 Speaker 1: There's a risk that comes along with machine learning. Actually, 689 00:44:48,280 --> 00:44:50,560 Speaker 1: technically there are a lot of risks, but this one 690 00:44:50,680 --> 00:44:54,280 Speaker 1: is a biggie. It is possible for machines like humans 691 00:44:54,760 --> 00:44:58,600 Speaker 1: to detect a pattern where there really isn't a pattern. 692 00:44:59,080 --> 00:45:03,160 Speaker 1: Systems might interpret noise to be signal, and depending on 693 00:45:03,239 --> 00:45:06,160 Speaker 1: what you're using the system to do, that could lead 694 00:45:06,200 --> 00:45:10,680 Speaker 1: you to some seriously dangerous incorrect conclusions. In some cases, 695 00:45:11,280 --> 00:45:13,640 Speaker 1: you could just be inconvenient, but depending on what you're 696 00:45:13,760 --> 00:45:17,319 Speaker 1: working toward, it could be catastrophic. And so computer scientists 697 00:45:17,400 --> 00:45:19,799 Speaker 1: know they have to do a lot of analysis to 698 00:45:19,840 --> 00:45:24,320 Speaker 1: make sure that patterns that are identified through machine learning 699 00:45:24,360 --> 00:45:30,960 Speaker 1: processes are actually real before acting on that information. Likewise, 700 00:45:31,440 --> 00:45:35,640 Speaker 1: bias is something that we humans have. Well, it's also 701 00:45:35,680 --> 00:45:39,560 Speaker 1: something that machine learning systems have too. Now, sometimes bias 702 00:45:39,800 --> 00:45:42,800 Speaker 1: is intentional. It can take the form of those weighted 703 00:45:42,920 --> 00:45:48,960 Speaker 1: relationships between artificial neurons. Other times, a systems architects, you know, 704 00:45:49,040 --> 00:45:52,280 Speaker 1: the people who put it together. They might have introduced bias, 705 00:45:52,400 --> 00:45:57,400 Speaker 1: not through conscious effort, but merely through the approach they took, 706 00:45:57,760 --> 00:46:01,120 Speaker 1: and that approach might have been too narrow. Ow we've 707 00:46:01,160 --> 00:46:04,719 Speaker 1: seen this pop up a lot again with facial recognition technologies, 708 00:46:04,760 --> 00:46:08,880 Speaker 1: many of which have a sliding scale of efficacy. They 709 00:46:08,960 --> 00:46:13,000 Speaker 1: might be more reliable with certain ethnicities like white people, 710 00:46:13,320 --> 00:46:16,960 Speaker 1: over others. That points a a likely problem with the 711 00:46:16,960 --> 00:46:20,239 Speaker 1: way those systems were trained. This is one of the 712 00:46:20,320 --> 00:46:23,560 Speaker 1: reasons why many companies have made a choice to stop 713 00:46:23,600 --> 00:46:28,080 Speaker 1: supplying certain parties like police forces and military branches with 714 00:46:28,160 --> 00:46:32,799 Speaker 1: facial recognition systems. The systems aren't reliable for all demographic 715 00:46:32,840 --> 00:46:37,120 Speaker 1: groups and thus could cause disproportionate harm to certain populations. 716 00:46:37,400 --> 00:46:40,440 Speaker 1: It would be a technological approach to systemic racism, and 717 00:46:40,480 --> 00:46:44,000 Speaker 1: this stuff is already out there in the wild. You 718 00:46:44,080 --> 00:46:47,480 Speaker 1: might think a computer system can't be biased or prejudiced 719 00:46:47,760 --> 00:46:51,160 Speaker 1: or racist, and sure, we're still not. At the point 720 00:46:51,200 --> 00:46:53,920 Speaker 1: where these systems are thinking in the way that humans do, 721 00:46:54,280 --> 00:46:59,160 Speaker 1: but the outcome is still disproportionately harmful to some groups. 722 00:46:59,640 --> 00:47:02,640 Speaker 1: Now that not to say that machine learning itself is bad. 723 00:47:03,120 --> 00:47:06,880 Speaker 1: It's not bad. It's a tool, just as all technology 724 00:47:06,920 --> 00:47:10,319 Speaker 1: is a tool used properly with a careful hand to 725 00:47:10,360 --> 00:47:15,000 Speaker 1: make sure that biases understood and where needed mitigated, and 726 00:47:15,200 --> 00:47:19,080 Speaker 1: where work can be double or triple checked before acted upon. 727 00:47:19,520 --> 00:47:22,840 Speaker 1: It is a remarkably useful tool, one that will power 728 00:47:22,960 --> 00:47:27,600 Speaker 1: and design and improve elements in our lives if it's 729 00:47:27,719 --> 00:47:31,040 Speaker 1: under the correct stewardship. But it does require a bit 730 00:47:31,040 --> 00:47:34,920 Speaker 1: more hands on work. We can't just leave it to 731 00:47:34,960 --> 00:47:40,320 Speaker 1: the machines just yet. Well, that wraps up this look 732 00:47:40,520 --> 00:47:43,400 Speaker 1: at the concept of machine learning and some of the 733 00:47:43,920 --> 00:47:48,040 Speaker 1: thought that underlies it. This really is a very high 734 00:47:48,120 --> 00:47:52,440 Speaker 1: level treatment of machine learning. There are plenty of resources 735 00:47:52,480 --> 00:47:54,719 Speaker 1: online if you want to dive in and learn more. 736 00:47:55,080 --> 00:47:58,040 Speaker 1: A lot of them get very heavy into the math. 737 00:47:58,280 --> 00:48:00,719 Speaker 1: So if that's not your bag, Uh, it might be 738 00:48:00,719 --> 00:48:03,200 Speaker 1: a little challenging to navigate. It certainly is for me. 739 00:48:03,840 --> 00:48:07,160 Speaker 1: I love learning about the stuff, but um, a lot 740 00:48:07,200 --> 00:48:10,480 Speaker 1: of it requires me to look up a term, then 741 00:48:10,560 --> 00:48:13,560 Speaker 1: look up a term that explains that term, and so on, 742 00:48:13,760 --> 00:48:16,920 Speaker 1: and I go down a rabbit hole. But hopefully you 743 00:48:17,000 --> 00:48:19,879 Speaker 1: have a better appreciation for what machine learning is at 744 00:48:19,880 --> 00:48:22,479 Speaker 1: this point. If you have suggestions for topics I should 745 00:48:22,560 --> 00:48:26,560 Speaker 1: cover in future text Stuff episodes, let me know. The 746 00:48:26,600 --> 00:48:28,560 Speaker 1: best way to get in touch with me is through 747 00:48:28,600 --> 00:48:32,120 Speaker 1: Twitter and the handle is text stuff H s W, 748 00:48:32,880 --> 00:48:41,320 Speaker 1: and I'll talk to you again really soon. Text Stuff 749 00:48:41,400 --> 00:48:44,600 Speaker 1: is an I Heart Radio production. For more podcasts from 750 00:48:44,600 --> 00:48:48,359 Speaker 1: my heart Radio, visit the i heart Radio app, Apple Podcasts, 751 00:48:48,480 --> 00:48:50,480 Speaker 1: or wherever you listen to your favorite shows.