1 00:00:02,920 --> 00:00:11,080 Speaker 1: Welcome to tech Stuff, a production from iHeartRadio. Hey there, 2 00:00:11,119 --> 00:00:14,560 Speaker 1: and welcome to tech Stuff. I'm your host, Jonathan Strickland. 3 00:00:14,640 --> 00:00:18,040 Speaker 1: I'm an executive producer with iHeart Podcasts. And how the 4 00:00:18,120 --> 00:00:21,960 Speaker 1: tech are you well. I just got back from celebrating 5 00:00:22,079 --> 00:00:25,000 Speaker 1: my birthday. Thank y'all for all of you who are 6 00:00:25,000 --> 00:00:28,040 Speaker 1: wishing me a happy birthday. And here in the United States, 7 00:00:28,120 --> 00:00:32,640 Speaker 1: we're about to have our national holiday celebrating the fourth 8 00:00:32,680 --> 00:00:36,320 Speaker 1: of July. I realize Fourth of July happens everywhere, not 9 00:00:36,520 --> 00:00:38,879 Speaker 1: just in the US, but we celebrate it here in 10 00:00:38,920 --> 00:00:42,680 Speaker 1: the US, and as such, there's very limited time to 11 00:00:42,680 --> 00:00:45,600 Speaker 1: get everything done, and I really wasn't able to pull 12 00:00:45,640 --> 00:00:48,320 Speaker 1: an episode together in time, and I apologize for that, 13 00:00:48,720 --> 00:00:52,239 Speaker 1: but I thought I would bring an older episode to 14 00:00:52,320 --> 00:00:55,480 Speaker 1: y'all so that we can still have an episode to 15 00:00:55,560 --> 00:00:59,800 Speaker 1: listen to today. And typically I would have one of my 16 00:01:00,080 --> 00:01:04,120 Speaker 1: Fireworks episodes play on this day, because Fireworks has a 17 00:01:04,240 --> 00:01:06,240 Speaker 1: very close association with the Fourth of July here in 18 00:01:06,240 --> 00:01:09,400 Speaker 1: the United States. But I've done that for several years 19 00:01:09,440 --> 00:01:11,560 Speaker 1: in a row, and I've thought it might be nice 20 00:01:11,600 --> 00:01:14,760 Speaker 1: to have a break from Fireworks instead. I thought I 21 00:01:14,760 --> 00:01:17,240 Speaker 1: would focus on something that continues to be a very 22 00:01:17,280 --> 00:01:21,120 Speaker 1: important topic in tech, and that is artificial intelligence. And 23 00:01:21,600 --> 00:01:26,280 Speaker 1: AI is incredibly impressive, but there are also lots of 24 00:01:26,480 --> 00:01:31,400 Speaker 1: challenges with AI, and those are ranging from the technological 25 00:01:31,480 --> 00:01:35,800 Speaker 1: side to the social side right and how we implement AI. 26 00:01:36,240 --> 00:01:38,560 Speaker 1: One thing I thought that we don't really get to 27 00:01:38,600 --> 00:01:43,959 Speaker 1: talk about very much is the concept of forgetting with AI. 28 00:01:44,240 --> 00:01:46,440 Speaker 1: We have a lot of generative AI out there that 29 00:01:46,959 --> 00:01:51,120 Speaker 1: is drawing upon huge resources of information, but AI can 30 00:01:51,240 --> 00:01:56,440 Speaker 1: also quote unquote forget. So this episode originally published on 31 00:01:56,520 --> 00:01:59,520 Speaker 1: July thirty first of twenty twenty three. It is called 32 00:01:59,600 --> 00:02:03,640 Speaker 1: Machine Learning and Catastrophic Forgetting. And I think it's a 33 00:02:03,760 --> 00:02:07,080 Speaker 1: useful thing to reflect upon as we see more and 34 00:02:07,160 --> 00:02:14,720 Speaker 1: more headlines about tech companies and their investment increasingly astronomical 35 00:02:14,840 --> 00:02:21,520 Speaker 1: investment in artificial intelligence. I hope you enjoy so. Over 36 00:02:21,560 --> 00:02:24,920 Speaker 1: this past weekend, I was listening to the podcast The 37 00:02:24,919 --> 00:02:27,920 Speaker 1: Skeptics Guide to the Universe, which I have no connection to. 38 00:02:28,200 --> 00:02:31,480 Speaker 1: I just listened to it, and it included a section 39 00:02:31,720 --> 00:02:36,280 Speaker 1: on AI that referenced something I don't think I had 40 00:02:36,400 --> 00:02:39,560 Speaker 1: heard of before, which is really talking more about my 41 00:02:39,680 --> 00:02:43,440 Speaker 1: oversight than anything else. Maybe I did hear about it 42 00:02:43,600 --> 00:02:47,160 Speaker 1: but then I forgot about it, you know, catastrophically. So 43 00:02:47,560 --> 00:02:52,280 Speaker 1: the thing they talked about was catastrophic forgetting in artificial intelligence, 44 00:02:52,280 --> 00:02:57,200 Speaker 1: specifically in machine learning systems built on artificial neural networks. Now, 45 00:02:57,200 --> 00:03:01,760 Speaker 1: before we talk about catastrophic forgetting, which as I mentioned, 46 00:03:01,800 --> 00:03:04,960 Speaker 1: is related to neural networks and machine learning, we really 47 00:03:05,000 --> 00:03:07,360 Speaker 1: need to do a quick reminder, not a quick reminder. 48 00:03:07,360 --> 00:03:09,280 Speaker 1: We need to do a full reminder on how all 49 00:03:09,360 --> 00:03:12,040 Speaker 1: this works. And that's going to require us to do 50 00:03:12,240 --> 00:03:15,560 Speaker 1: a whole lot of remembering. Not a catastrophic amount, but 51 00:03:15,639 --> 00:03:19,280 Speaker 1: a lot. So the history of artificial intelligence as a 52 00:03:19,320 --> 00:03:25,120 Speaker 1: discipline is one of intense and important debates in fields 53 00:03:25,160 --> 00:03:28,040 Speaker 1: like computer science. Now, I have often talked about how 54 00:03:28,120 --> 00:03:31,480 Speaker 1: AI can be seen as the convergence of several other 55 00:03:31,600 --> 00:03:35,600 Speaker 1: disciplines into its own field. And there's more than one 56 00:03:35,600 --> 00:03:40,680 Speaker 1: way to approach the challenge of artificial intelligence. And in 57 00:03:40,760 --> 00:03:43,440 Speaker 1: the history of AI, we actually saw that play out, 58 00:03:44,080 --> 00:03:47,680 Speaker 1: and some would argue the way it played out means 59 00:03:47,720 --> 00:03:51,200 Speaker 1: that we're actually just now playing catch up. So different 60 00:03:51,240 --> 00:03:56,200 Speaker 1: schools of thought pushed these different approaches forward as this 61 00:03:56,400 --> 00:04:01,920 Speaker 1: should be the prevailing methodology we use to develop artificial intelligence. 62 00:04:02,360 --> 00:04:05,440 Speaker 1: This is important because the development of AI does not 63 00:04:05,560 --> 00:04:09,680 Speaker 1: exist in a vacuum, right. It exists in our real world. 64 00:04:10,320 --> 00:04:16,760 Speaker 1: Research requires funding, and when you've got different sides arguing 65 00:04:16,800 --> 00:04:21,160 Speaker 1: that their approach to artificial intelligence is superior and that 66 00:04:21,200 --> 00:04:25,400 Speaker 1: the alternatives are not just inferior, but potentially limited to 67 00:04:25,440 --> 00:04:28,360 Speaker 1: the point of being useless, well you've got a metaphorical 68 00:04:28,440 --> 00:04:31,760 Speaker 1: wrestling match going on. The winner takes home the big 69 00:04:31,800 --> 00:04:36,000 Speaker 1: prize of getting funding for their research, and the loser 70 00:04:36,120 --> 00:04:38,839 Speaker 1: has to scrabble for whatever they can find, and often 71 00:04:39,080 --> 00:04:42,840 Speaker 1: they will see their work languish as a result. By 72 00:04:42,880 --> 00:04:45,960 Speaker 1: the way, this is why I often bring stuff up 73 00:04:46,000 --> 00:04:49,520 Speaker 1: in this podcast that is outside the realm of tech. 74 00:04:50,480 --> 00:04:52,720 Speaker 1: I've received a lot of messages over the years from 75 00:04:52,720 --> 00:04:55,400 Speaker 1: folks saying that I should leave out stuff like money 76 00:04:55,880 --> 00:04:58,640 Speaker 1: or politics. Politics is the big one. But to me, 77 00:04:58,760 --> 00:05:04,720 Speaker 1: that doesn't make sense because tech exists within our world, 78 00:05:04,839 --> 00:05:08,640 Speaker 1: a world that is largely shaped by money and politics. 79 00:05:09,040 --> 00:05:12,000 Speaker 1: I don't think we can separate the tech from all 80 00:05:12,040 --> 00:05:14,440 Speaker 1: of that because I believe that if you were to 81 00:05:14,480 --> 00:05:18,839 Speaker 1: somehow magically remove those influences, If somehow money and politics 82 00:05:18,880 --> 00:05:22,600 Speaker 1: never played a part in the development of technology, our 83 00:05:22,640 --> 00:05:25,479 Speaker 1: tech would look very different from what it does today. 84 00:05:25,960 --> 00:05:29,960 Speaker 1: Not necessarily better or worse, but different. I mean, think 85 00:05:29,960 --> 00:05:36,040 Speaker 1: about Thomas Edison. He was very much driven by financial success, 86 00:05:36,120 --> 00:05:40,200 Speaker 1: like his work in tech was really mostly about making 87 00:05:40,320 --> 00:05:43,520 Speaker 1: lots of money. And without the making lots of money part, 88 00:05:43,920 --> 00:05:47,480 Speaker 1: you don't really have his drive to really bring together 89 00:05:47,560 --> 00:05:50,800 Speaker 1: the brightest minds of his generation and set them to 90 00:05:50,880 --> 00:05:55,080 Speaker 1: work on creating incredible technology. So I think we have 91 00:05:55,240 --> 00:05:58,440 Speaker 1: to take all these things into consideration. Anyway, that's a 92 00:05:58,480 --> 00:06:00,720 Speaker 1: total rabbit trail, and I apology. Let's get back to 93 00:06:00,760 --> 00:06:05,200 Speaker 1: our story. It really begins around nineteen forty three when 94 00:06:05,200 --> 00:06:08,360 Speaker 1: a pair of researchers at the University of Chicago first 95 00:06:08,640 --> 00:06:13,080 Speaker 1: proposed the concept of the basic unit of a neural network. 96 00:06:13,400 --> 00:06:18,279 Speaker 1: Those researchers were Warren McCullough and Walter Pets, And in fact, 97 00:06:18,320 --> 00:06:22,839 Speaker 1: they demonstrate their idea by showing a simple electrical circuit 98 00:06:23,040 --> 00:06:25,839 Speaker 1: the very basis for what would become a neural network. 99 00:06:26,320 --> 00:06:29,679 Speaker 1: So their proposal was a system that would use those 100 00:06:29,720 --> 00:06:33,880 Speaker 1: simple circuits to mimic the neurons that we have in 101 00:06:33,880 --> 00:06:37,720 Speaker 1: our noggins. So our brain consists of a bunch of 102 00:06:37,760 --> 00:06:40,719 Speaker 1: these neurons, and you might wonder how much is a bunch. Well, 103 00:06:41,600 --> 00:06:45,159 Speaker 1: we're talking about on average, around one hundred billion neurons 104 00:06:45,320 --> 00:06:48,920 Speaker 1: in the human brain. These neurons interconnect with each other. 105 00:06:49,040 --> 00:06:51,640 Speaker 1: It's not just a one to one, right, You've got 106 00:06:51,640 --> 00:06:55,839 Speaker 1: these interconnections between all these different neurons, not with every 107 00:06:55,839 --> 00:06:58,880 Speaker 1: neuron connected to every other neuron, but lots of interconnections. 108 00:06:58,880 --> 00:07:01,680 Speaker 1: And if we're looking at just the connections, you would 109 00:07:01,720 --> 00:07:04,839 Speaker 1: count more than one hundred trillion of them in the 110 00:07:04,880 --> 00:07:08,560 Speaker 1: typical human brain. And these connections in our brains make 111 00:07:08,640 --> 00:07:13,320 Speaker 1: up neural circuits. Those circuits light up, and that represents 112 00:07:13,400 --> 00:07:16,640 Speaker 1: us doing lots of different stuff, from experiencing the world 113 00:07:16,680 --> 00:07:20,840 Speaker 1: around us so perception to thinking about a past memory. 114 00:07:21,000 --> 00:07:24,000 Speaker 1: You know that typically is like recreating the same pathway 115 00:07:24,080 --> 00:07:28,440 Speaker 1: over and over, and sometimes we don't recreate it exactly correctly, 116 00:07:28,920 --> 00:07:32,920 Speaker 1: and our memory ends up not being a perfect representation 117 00:07:33,080 --> 00:07:35,880 Speaker 1: of the thing that we actually experienced. This is why 118 00:07:36,120 --> 00:07:39,360 Speaker 1: things like eyewitness testimony is not always very reliable, because 119 00:07:39,400 --> 00:07:44,520 Speaker 1: our memories aren't infallible. They can trick us and we 120 00:07:44,560 --> 00:07:47,040 Speaker 1: can have all those pathways light up. When we learn 121 00:07:47,080 --> 00:07:50,200 Speaker 1: a new skill, we start forming new pathways, and then 122 00:07:50,360 --> 00:07:54,800 Speaker 1: as we practice this skill, we start to reinforce those pathways. 123 00:07:55,160 --> 00:07:58,800 Speaker 1: So McCulla and Pitts propose that we create machines capable 124 00:07:58,880 --> 00:08:03,320 Speaker 1: of doing essentially a similar thing that our brains do, 125 00:08:03,440 --> 00:08:08,680 Speaker 1: so kind of a neuromimicry, not exactly one to one 126 00:08:08,720 --> 00:08:12,600 Speaker 1: the way our brains work, but inspired by the way 127 00:08:12,760 --> 00:08:17,080 Speaker 1: our brains work. Now, we would be limited by what 128 00:08:17,360 --> 00:08:19,920 Speaker 1: the technology of the day would be able to do, 129 00:08:20,360 --> 00:08:23,640 Speaker 1: because there's no feasible way we could create a massive 130 00:08:24,160 --> 00:08:29,640 Speaker 1: electrical system with one hundred billion individual simple circuits with 131 00:08:29,760 --> 00:08:33,240 Speaker 1: more than one hundred trillion connections between them. That would 132 00:08:33,240 --> 00:08:37,199 Speaker 1: be beyond our capability. It would be beyond our resources. 133 00:08:37,559 --> 00:08:40,840 Speaker 1: We could, however, create systems that used interconnected circuits to 134 00:08:40,920 --> 00:08:45,480 Speaker 1: process information and to teach such a system to do 135 00:08:45,559 --> 00:08:50,920 Speaker 1: specific tasks. Now, in nineteen forty nine, Donald Hebb wrote 136 00:08:50,960 --> 00:08:55,080 Speaker 1: a book about biological neurons, and he titled this book 137 00:08:55,320 --> 00:08:59,960 Speaker 1: the Organization of Behavior and suggested neural pathways get stronger 138 00:09:00,520 --> 00:09:03,320 Speaker 1: with additional use, kind of like you know, if you 139 00:09:03,559 --> 00:09:06,520 Speaker 1: exercise your muscles, you build strength over time, while so 140 00:09:06,720 --> 00:09:10,640 Speaker 1: is the same with neural pathways, and if you don't 141 00:09:10,720 --> 00:09:13,240 Speaker 1: use those muscles, well, then your muscles get weaker. Well, 142 00:09:13,320 --> 00:09:16,760 Speaker 1: same with neural pathways. If you end up learning a skill, 143 00:09:17,480 --> 00:09:21,600 Speaker 1: but then over a great amount of time you no 144 00:09:21,640 --> 00:09:24,560 Speaker 1: longer practice that skill, you're going to lose some of 145 00:09:24,600 --> 00:09:27,400 Speaker 1: your ability, maybe not all of it, but at least 146 00:09:27,400 --> 00:09:29,240 Speaker 1: some of it. And you have to you know, like 147 00:09:29,559 --> 00:09:33,240 Speaker 1: I think about wrestlers who come back from from retirement, 148 00:09:33,360 --> 00:09:36,520 Speaker 1: professional wrestlers, they call it ring rust. You got to 149 00:09:36,600 --> 00:09:39,120 Speaker 1: knock off the ring rust and get back into step 150 00:09:39,200 --> 00:09:41,320 Speaker 1: and kind of get back into your groove. And it 151 00:09:41,360 --> 00:09:45,880 Speaker 1: takes a little time. Typically sometimes you know, you can 152 00:09:46,000 --> 00:09:48,280 Speaker 1: get back into the game faster than others, but you 153 00:09:48,400 --> 00:09:53,040 Speaker 1: get the idea. And also heb ended up proposing the 154 00:09:53,080 --> 00:09:58,080 Speaker 1: concept of cells that fire together wire together, meaning that 155 00:09:58,800 --> 00:10:02,800 Speaker 1: neurons that fire at the same time end up strengthening 156 00:10:02,880 --> 00:10:08,160 Speaker 1: faster than other neurons do. So when you get into 157 00:10:08,240 --> 00:10:14,040 Speaker 1: that system, you can actually reinforce those pathways. And for 158 00:10:14,160 --> 00:10:17,120 Speaker 1: AI this would be really important. And it wasn't very 159 00:10:17,160 --> 00:10:20,599 Speaker 1: long after Donald Habb had published this work that researchers 160 00:10:20,600 --> 00:10:23,679 Speaker 1: in the field of AI tried to apply that concept 161 00:10:23,760 --> 00:10:28,480 Speaker 1: that philosophy to computer science. By the mid nineteen fifties, 162 00:10:28,520 --> 00:10:32,040 Speaker 1: the burgeoning computer science lab and AI lab at MIT 163 00:10:32,880 --> 00:10:38,400 Speaker 1: was building out neural networks based on Hebb's ideas. Meanwhile, 164 00:10:38,840 --> 00:10:43,680 Speaker 1: another computer scientist named Frank Rosenblatt was looking at primitive 165 00:10:43,679 --> 00:10:48,079 Speaker 1: neural systems and he started with flies like house flies. 166 00:10:49,040 --> 00:10:52,160 Speaker 1: He wanted to explore systems that were involved when a 167 00:10:52,200 --> 00:10:56,560 Speaker 1: fly would quickly move away after detecting a possible threat, 168 00:10:57,000 --> 00:11:01,439 Speaker 1: like instantly, or at least appear to us to instantly 169 00:11:01,520 --> 00:11:05,480 Speaker 1: react to something. So, for example, a fly swatter coming 170 00:11:05,480 --> 00:11:07,640 Speaker 1: at it, like you might be moving the fly swater 171 00:11:07,720 --> 00:11:09,840 Speaker 1: very quickly, and yet the fly is able to move 172 00:11:10,400 --> 00:11:15,640 Speaker 1: super fast with no perceivable delay. Right, we know that 173 00:11:15,679 --> 00:11:18,200 Speaker 1: we have a delay from when we perceive something to 174 00:11:18,240 --> 00:11:20,520 Speaker 1: when we can act on something. Like if you've ever 175 00:11:20,559 --> 00:11:23,000 Speaker 1: been in a fender bender in a car accident, you 176 00:11:23,040 --> 00:11:25,920 Speaker 1: know that that there's a delay between when you see 177 00:11:25,920 --> 00:11:28,680 Speaker 1: the issue when you can hit the brake, and that 178 00:11:28,920 --> 00:11:32,240 Speaker 1: can lead to accidents. Well, with flies, that delay seems 179 00:11:32,280 --> 00:11:36,600 Speaker 1: to be super super small. So Rosenblatt was really interested 180 00:11:36,960 --> 00:11:40,960 Speaker 1: in exploring the neurological reasons for that. How can that happen? 181 00:11:41,000 --> 00:11:43,520 Speaker 1: It has to be really simple, right, There has to 182 00:11:43,559 --> 00:11:48,199 Speaker 1: be a simple and more or less direct pathway that 183 00:11:48,360 --> 00:11:52,800 Speaker 1: exists to allow a fly to react to detecting a 184 00:11:52,800 --> 00:11:57,160 Speaker 1: potential threat like that, and if you could replicate that 185 00:11:57,920 --> 00:12:02,040 Speaker 1: with electronics, you could have a very simple but potentially 186 00:12:02,200 --> 00:12:07,240 Speaker 1: powerful artificial intelligence system. So he came up with this 187 00:12:07,440 --> 00:12:10,160 Speaker 1: system that would be based off that very simple direct 188 00:12:10,200 --> 00:12:12,240 Speaker 1: pathway that you would see in something like a fly, 189 00:12:12,760 --> 00:12:16,120 Speaker 1: and he called it the perceptron. So he went back 190 00:12:16,200 --> 00:12:18,680 Speaker 1: to the simple circuit design that was proposed by Pitts 191 00:12:18,679 --> 00:12:22,520 Speaker 1: and McCullough and he built out the Mark one perceptron 192 00:12:23,480 --> 00:12:25,920 Speaker 1: or perceptron. I guess I should say, so let's talk 193 00:12:25,920 --> 00:12:28,920 Speaker 1: about a perceptron, like not big P, but a little 194 00:12:29,040 --> 00:12:31,840 Speaker 1: P perceptron. This is probably what we would call a 195 00:12:31,920 --> 00:12:35,680 Speaker 1: neural node in a modern neural network. So the purpose 196 00:12:35,800 --> 00:12:40,000 Speaker 1: of the perceptron was to accept inputs and produce an 197 00:12:40,040 --> 00:12:44,679 Speaker 1: output based on some threshold, Like if the inputs meet 198 00:12:44,720 --> 00:12:47,640 Speaker 1: a certain threshold, one output would be produced. If they 199 00:12:47,720 --> 00:12:49,880 Speaker 1: failed to do so, a different output would be produced. 200 00:12:50,720 --> 00:12:54,880 Speaker 1: The inputs, in turn would be assigned weights, which would 201 00:12:54,880 --> 00:12:58,240 Speaker 1: factor into the output the perceptron would generate. So when 202 00:12:58,240 --> 00:13:04,760 Speaker 1: we're talking weights, I mean weights as in like how 203 00:13:04,840 --> 00:13:08,079 Speaker 1: heavy something is or in this case, how much impact 204 00:13:08,520 --> 00:13:12,200 Speaker 1: that thing has, So we're talking about how much impact 205 00:13:12,280 --> 00:13:15,920 Speaker 1: one input has relative to other inputs. Let me use 206 00:13:15,960 --> 00:13:19,440 Speaker 1: a really mundane human example to kind of explain what 207 00:13:19,520 --> 00:13:22,640 Speaker 1: this means. Let's say that your friend asks you to 208 00:13:22,679 --> 00:13:24,760 Speaker 1: go see a movie with them, and it's going to 209 00:13:24,800 --> 00:13:27,760 Speaker 1: be playing tonight at nine pm. But you've had a 210 00:13:27,880 --> 00:13:30,880 Speaker 1: really busy day and you might not be able to 211 00:13:30,920 --> 00:13:34,320 Speaker 1: even eat dinner until around nine pm. And if you 212 00:13:34,360 --> 00:13:36,280 Speaker 1: go see this movie, it might mean having to skip 213 00:13:36,320 --> 00:13:40,000 Speaker 1: dinner or to try and eat something really fast and 214 00:13:40,120 --> 00:13:43,599 Speaker 1: unhealthy before you go to the movie. What's more, you 215 00:13:43,679 --> 00:13:46,680 Speaker 1: got a really big day tomorrow and you feel like 216 00:13:46,720 --> 00:13:49,480 Speaker 1: you really need to be well rested for it. However, 217 00:13:49,600 --> 00:13:53,320 Speaker 1: at the same time, you haven't seen this friend in ages, 218 00:13:53,360 --> 00:13:55,800 Speaker 1: and you really like this person and you've wanted to 219 00:13:55,800 --> 00:13:59,000 Speaker 1: hang with them for a really long time. Plus the 220 00:13:59,040 --> 00:14:01,600 Speaker 1: movie they're suggesting is one you've really wanted to see 221 00:14:01,600 --> 00:14:04,840 Speaker 1: and you haven't gone yet. Well, you would likely assign 222 00:14:04,960 --> 00:14:09,360 Speaker 1: at least unconsciously weights to each of these factors before 223 00:14:09,360 --> 00:14:11,439 Speaker 1: you make your decision. You know, if getting some dinner 224 00:14:11,480 --> 00:14:14,440 Speaker 1: without having to rush, and also to be really well 225 00:14:14,480 --> 00:14:17,720 Speaker 1: rested for tomorrow are really important to you, you'll probably 226 00:14:18,000 --> 00:14:21,880 Speaker 1: reluctantly decline the offer. But if you really crave some 227 00:14:21,960 --> 00:14:24,000 Speaker 1: time with your friend and you really want to see 228 00:14:24,000 --> 00:14:26,360 Speaker 1: that movie before all the spoilers come out on Facebook 229 00:14:26,400 --> 00:14:30,440 Speaker 1: or whatever, maybe you'll say yes. Your decision depends upon 230 00:14:30,480 --> 00:14:34,520 Speaker 1: the weights you assign those factors, those inputs, even if 231 00:14:34,520 --> 00:14:38,000 Speaker 1: you don't consciously think about it that way. Well, the 232 00:14:38,040 --> 00:14:41,920 Speaker 1: Perceptron system worked in a similar way, produced outputs by 233 00:14:41,920 --> 00:14:46,800 Speaker 1: taking the inputs into consideration, including each input's weight. Moreover, 234 00:14:47,080 --> 00:14:49,560 Speaker 1: the more you submitted inputs, the more the system would 235 00:14:49,640 --> 00:14:53,280 Speaker 1: quote unquote learn how to weight each of those inputs, 236 00:14:53,560 --> 00:14:56,600 Speaker 1: all with the goal of bringing the actual output that 237 00:14:56,640 --> 00:15:00,360 Speaker 1: the process or you know, generates closer to the one 238 00:15:00,560 --> 00:15:04,920 Speaker 1: you want it to generate. Okay, I just said a 239 00:15:04,920 --> 00:15:07,280 Speaker 1: lot there. We've got some more to get through. But 240 00:15:07,320 --> 00:15:09,320 Speaker 1: before we get to that, let's take a quick break, 241 00:15:18,520 --> 00:15:20,920 Speaker 1: all right. Before the break, we were talking about inputs 242 00:15:21,080 --> 00:15:25,240 Speaker 1: and weights and the idea of getting an output that 243 00:15:25,520 --> 00:15:28,240 Speaker 1: is close to what you want the system to do. 244 00:15:28,960 --> 00:15:31,720 Speaker 1: That's not a guarantee, right, The system could generate an 245 00:15:31,720 --> 00:15:35,800 Speaker 1: output that's quote unquote wrong, you know, depending on whatever 246 00:15:35,880 --> 00:15:41,080 Speaker 1: task you've set this machine learning system to learn, and 247 00:15:41,160 --> 00:15:43,280 Speaker 1: that gets a bit conceptual. So let's talk about a 248 00:15:43,320 --> 00:15:45,840 Speaker 1: simple example that I love to use. If you've been 249 00:15:45,840 --> 00:15:48,400 Speaker 1: listening to texta for a while, you've heard this before, 250 00:15:49,400 --> 00:15:53,000 Speaker 1: and that's talking about pictures of cats. Because cats ruled 251 00:15:53,160 --> 00:15:55,440 Speaker 1: the Internet. I don't know if they still do. They 252 00:15:55,480 --> 00:15:58,960 Speaker 1: won't talk to me, so just knock things off shelves. Anyway. 253 00:15:58,960 --> 00:16:01,320 Speaker 1: If your goal is to tea each a computer system 254 00:16:01,720 --> 00:16:06,360 Speaker 1: to differentiate photos that include a cat from photos that 255 00:16:06,440 --> 00:16:10,000 Speaker 1: do not include a cat, well, you would need to 256 00:16:10,040 --> 00:16:13,400 Speaker 1: train the system, and part of that includes feeding the 257 00:16:13,480 --> 00:16:18,200 Speaker 1: system a whole bunch of photographs. Some of those would 258 00:16:18,240 --> 00:16:21,960 Speaker 1: have cats in them, some would not, and chances are 259 00:16:22,040 --> 00:16:25,840 Speaker 1: the system would misidentify photos. Maybe a significant number of 260 00:16:25,840 --> 00:16:28,680 Speaker 1: those photos. You would probably have false positives where the 261 00:16:28,720 --> 00:16:31,560 Speaker 1: system thinks there's a cat there and there's not, and 262 00:16:31,600 --> 00:16:34,280 Speaker 1: false negatives where it doesn't think there's a cat there 263 00:16:34,560 --> 00:16:37,680 Speaker 1: but there is. At that point, your goal is to 264 00:16:37,680 --> 00:16:41,120 Speaker 1: try and teach the system to close the gap between 265 00:16:41,360 --> 00:16:44,800 Speaker 1: the actual results it produces and what you want it 266 00:16:44,920 --> 00:16:47,760 Speaker 1: to produce. In some systems, that means you might have 267 00:16:47,840 --> 00:16:51,320 Speaker 1: to go in manually to adjust the input weights to 268 00:16:51,440 --> 00:16:53,880 Speaker 1: increase the weight of one input versus another in an 269 00:16:53,920 --> 00:16:59,360 Speaker 1: effort to cut down on mistakes. So the perceptron was interesting, 270 00:16:59,760 --> 00:17:03,080 Speaker 1: but it was very limited in complexity. It was essentially 271 00:17:03,160 --> 00:17:05,560 Speaker 1: a single layer where you'd feed a bunch of inputs 272 00:17:05,560 --> 00:17:07,879 Speaker 1: in and you would get an output. So it was 273 00:17:07,920 --> 00:17:11,959 Speaker 1: suitable for a subset of computational challenges, but anything beyond 274 00:17:12,000 --> 00:17:16,119 Speaker 1: that was well beyond its own reach as a single 275 00:17:16,200 --> 00:17:19,719 Speaker 1: layer network. By the late nineteen fifties, other researchers had 276 00:17:19,760 --> 00:17:23,879 Speaker 1: created new neural networks that were multi layered. So a 277 00:17:23,960 --> 00:17:28,160 Speaker 1: node or neuron didn't just accept inputs, it would generate 278 00:17:28,200 --> 00:17:32,600 Speaker 1: outputs that then would become inputs for another layer down. 279 00:17:33,000 --> 00:17:36,399 Speaker 1: So instead of just having one layer of nodes, you 280 00:17:36,400 --> 00:17:38,840 Speaker 1: would have multiple layers of nodes. Typically you would have 281 00:17:39,280 --> 00:17:43,119 Speaker 1: one at the quote unquote top of the network, and 282 00:17:43,160 --> 00:17:44,880 Speaker 1: you would have outputs at the bottom, and the ones 283 00:17:44,880 --> 00:17:47,920 Speaker 1: in between would be often referred to as hidden layers, 284 00:17:48,400 --> 00:17:51,640 Speaker 1: and who knows how many there would be. So anyway 285 00:17:52,040 --> 00:17:54,840 Speaker 1: you would feed data to the system, the initial nodes 286 00:17:54,880 --> 00:17:58,879 Speaker 1: would generate information as outputs that would become inputs for 287 00:17:58,960 --> 00:18:03,680 Speaker 1: the next layer down, which would then continue the process 288 00:18:03,720 --> 00:18:05,679 Speaker 1: and so on and so forth until you get to 289 00:18:05,720 --> 00:18:08,760 Speaker 1: the output. So now you had artificial neural networks that 290 00:18:08,800 --> 00:18:13,199 Speaker 1: could tackle more complex challenges, and you would have multiple 291 00:18:13,200 --> 00:18:17,120 Speaker 1: steps in the process. Didn't necessarily mean they were automatically 292 00:18:17,200 --> 00:18:21,280 Speaker 1: better than the perceptron, was just that they were able 293 00:18:21,320 --> 00:18:27,119 Speaker 1: to tackle more complicated tasks. What followed is something that 294 00:18:27,160 --> 00:18:30,680 Speaker 1: will probably sound really familiar to you if you ever 295 00:18:30,840 --> 00:18:35,919 Speaker 1: follow technology or fads, the hype around machine learning and 296 00:18:36,000 --> 00:18:38,800 Speaker 1: artificial intelligence, and keep in mind this is like the 297 00:18:38,920 --> 00:18:43,920 Speaker 1: nineteen sixties. It grew beyond the technology's actual capabilities. At 298 00:18:43,920 --> 00:18:47,840 Speaker 1: that time. People started to project what this technology would 299 00:18:47,880 --> 00:18:50,239 Speaker 1: be able to do, and they did so thinking it 300 00:18:50,280 --> 00:18:53,520 Speaker 1: was going to be in a very short turnaround, like 301 00:18:53,560 --> 00:18:58,080 Speaker 1: we're right on the very precipice of a monstrous breakthrough 302 00:18:58,119 --> 00:19:00,960 Speaker 1: that will bring the science fiction future into the present. 303 00:19:01,880 --> 00:19:06,719 Speaker 1: So when it was realized that we weren't at that, like, 304 00:19:06,800 --> 00:19:10,639 Speaker 1: that's not how progress typically works. It's usually much more 305 00:19:11,119 --> 00:19:16,200 Speaker 1: gradual and humble than that, well, then enthusiasm around AI 306 00:19:16,280 --> 00:19:18,800 Speaker 1: began to take a hit. And as I mentioned already, 307 00:19:18,840 --> 00:19:22,440 Speaker 1: a big part of AI research really comes down to funding, 308 00:19:23,000 --> 00:19:26,360 Speaker 1: and it gets really challenging to secure funding when public 309 00:19:26,480 --> 00:19:31,200 Speaker 1: opinion dims on a technology. We've seen this happen lots 310 00:19:31,200 --> 00:19:35,000 Speaker 1: of times, right, like three D television was a fad 311 00:19:35,080 --> 00:19:37,720 Speaker 1: that was pushed. Now, granted, that one, you could argue 312 00:19:37,800 --> 00:19:41,120 Speaker 1: was more of an example of manufacturing companies that make 313 00:19:41,200 --> 00:19:44,800 Speaker 1: televisions trying to push a technology on consumers and the 314 00:19:44,800 --> 00:19:47,520 Speaker 1: consumers just weren't interested. You could argue that was the 315 00:19:47,560 --> 00:19:51,000 Speaker 1: case there. But virtual reality in the nineteen nineties definitely 316 00:19:51,040 --> 00:19:54,639 Speaker 1: followed this pathway. There was this excitement around virtual reality. 317 00:19:55,640 --> 00:19:59,480 Speaker 1: Then that excitement faded to almost nothing when people realized 318 00:19:59,480 --> 00:20:02,800 Speaker 1: that the actual state of the art of the technology 319 00:20:03,000 --> 00:20:06,480 Speaker 1: was far below where they expected it to be. And 320 00:20:06,560 --> 00:20:10,040 Speaker 1: suddenly people who are working in VR couldn't get funding 321 00:20:10,200 --> 00:20:12,400 Speaker 1: for their work and they kind of had to scrounge 322 00:20:12,440 --> 00:20:16,359 Speaker 1: around in order to keep the development going at all. 323 00:20:17,040 --> 00:20:19,879 Speaker 1: And then eventually we would see that come back around again. 324 00:20:20,480 --> 00:20:24,040 Speaker 1: You could argue that NFTs recently went through this too, 325 00:20:24,080 --> 00:20:27,560 Speaker 1: where the hype went well beyond what NFTs could actually do. 326 00:20:28,640 --> 00:20:31,920 Speaker 1: I've been really down on NFTs in general. I do 327 00:20:31,960 --> 00:20:37,080 Speaker 1: think that there are potential legitimate uses for NFTs, but 328 00:20:37,160 --> 00:20:43,399 Speaker 1: I think the early examples were frivolous and almost solely 329 00:20:43,480 --> 00:20:49,400 Speaker 1: centered around speculation, as in like financial speculation and as 330 00:20:49,400 --> 00:20:51,320 Speaker 1: a result, there was nothing for it to do other 331 00:20:51,400 --> 00:20:54,520 Speaker 1: than to create a bubble that would ultimately burst, which 332 00:20:54,560 --> 00:20:58,199 Speaker 1: is what happened. And maybe NFTs will recover from that 333 00:20:58,320 --> 00:21:02,440 Speaker 1: and become something that's more fundamentally useful in the Internet 334 00:21:02,520 --> 00:21:05,560 Speaker 1: in the future or in digital commerce in the future. 335 00:21:06,920 --> 00:21:10,879 Speaker 1: But it's going to have to get over the catastrophe 336 00:21:10,920 --> 00:21:13,680 Speaker 1: that happened when the rug was pulled out from underneath 337 00:21:13,760 --> 00:21:19,520 Speaker 1: n FTS. And that was all predictable and preventable. But 338 00:21:21,000 --> 00:21:23,919 Speaker 1: like I've said before, like I've lifted the joke from 339 00:21:23,960 --> 00:21:26,440 Speaker 1: Peter Cook, we've learned from our mistakes. We can repeat 340 00:21:26,480 --> 00:21:31,040 Speaker 1: them almost exactly. Anyway, This same sort of hype cycle 341 00:21:31,119 --> 00:21:35,800 Speaker 1: activity happened with neural networks and machine learning in the 342 00:21:35,880 --> 00:21:41,639 Speaker 1: nineteen sixties. Then enter Marvin Minsky and Seymour Pappart of 343 00:21:41,800 --> 00:21:44,920 Speaker 1: MIT's AI lab. They were leading that lab at the time. 344 00:21:45,280 --> 00:21:49,800 Speaker 1: In nineteen sixty nine, they co authored a book titled Perceptrons. 345 00:21:50,720 --> 00:21:55,040 Speaker 1: They were actually critical of that artificial neural network approach 346 00:21:55,080 --> 00:21:58,080 Speaker 1: to AI and machine learning. They were concerned that the 347 00:21:58,119 --> 00:22:01,040 Speaker 1: limitations of the technology meant that you would need an 348 00:22:01,160 --> 00:22:06,399 Speaker 1: unrealistically huge system of artificial neurons. Perhaps then using that 349 00:22:06,400 --> 00:22:10,639 Speaker 1: system to compute an infinite number of variations of the 350 00:22:10,680 --> 00:22:14,399 Speaker 1: same process or task if you wanted to train the 351 00:22:14,400 --> 00:22:18,879 Speaker 1: weights so that they were of the optimal value. So, 352 00:22:18,920 --> 00:22:22,920 Speaker 1: in other words, they thought, it's too impractical and it's 353 00:22:22,960 --> 00:22:24,960 Speaker 1: going to take too much compute time, and you're never 354 00:22:25,040 --> 00:22:27,360 Speaker 1: going to achieve the result you want. You're never going 355 00:22:27,400 --> 00:22:32,600 Speaker 1: to get to that most perfect system. And they believed 356 00:22:33,119 --> 00:22:37,760 Speaker 1: it just had fundamental inescapable flaws. They had different systems 357 00:22:37,800 --> 00:22:42,120 Speaker 1: in mind. Now Minski and Separate tried to push their 358 00:22:42,160 --> 00:22:44,680 Speaker 1: systems forward, and I could do a full episode about 359 00:22:44,720 --> 00:22:48,800 Speaker 1: them too, and their ideas were not bad. They were different. 360 00:22:49,160 --> 00:22:51,520 Speaker 1: It was a different approach. But this also meant that 361 00:22:51,600 --> 00:22:54,520 Speaker 1: researchers who had been pushing the development of our artificial 362 00:22:54,560 --> 00:22:58,919 Speaker 1: neural networks felt forced to move on to different projects 363 00:22:59,000 --> 00:23:03,600 Speaker 1: because financial support for anything connected to the concept of 364 00:23:03,640 --> 00:23:09,120 Speaker 1: neural networks effectively disappeared, right like funding just dropped for that. 365 00:23:09,200 --> 00:23:13,359 Speaker 1: Because here you had these experts in computer science saying, yeah, 366 00:23:13,560 --> 00:23:19,159 Speaker 1: this approach, while interesting, has already hit an insurmountable obstacle 367 00:23:19,200 --> 00:23:20,960 Speaker 1: and it's not going to go any further. It's gone 368 00:23:21,000 --> 00:23:23,880 Speaker 1: as far as it can go. And so a lot 369 00:23:23,920 --> 00:23:29,640 Speaker 1: of computer scientists blamed Minsky and Separate for essentially demolishing 370 00:23:29,720 --> 00:23:33,680 Speaker 1: funding for neural networks for more than a decade, and 371 00:23:33,680 --> 00:23:37,320 Speaker 1: in fact, this would become an era that retrospectively, computer 372 00:23:37,400 --> 00:23:41,680 Speaker 1: scientists would reference as the AI Winter got all Game 373 00:23:41,720 --> 00:23:44,800 Speaker 1: of Thrones up in here. Now. In nineteen eighty two, 374 00:23:45,240 --> 00:23:49,200 Speaker 1: there was a hint of spring thawing out that AI 375 00:23:49,240 --> 00:23:54,120 Speaker 1: Winter researchers in Japan were starting to resurrect work on 376 00:23:54,280 --> 00:23:58,640 Speaker 1: neural network projects, and meanwhile, a scientist named John Hopfield 377 00:23:59,080 --> 00:24:02,080 Speaker 1: submitted a research paper to the National Academy of Sciences 378 00:24:02,560 --> 00:24:05,280 Speaker 1: that brought neural networks back into discussion here in the 379 00:24:05,359 --> 00:24:10,800 Speaker 1: United States. And because Japan was actively investing in developing 380 00:24:10,800 --> 00:24:15,000 Speaker 1: that technology, institutions in the United States began to open 381 00:24:15,119 --> 00:24:17,359 Speaker 1: up the purse strings a bit because there was a 382 00:24:17,400 --> 00:24:21,280 Speaker 1: concern that if there were something to this artificial neural 383 00:24:21,320 --> 00:24:25,920 Speaker 1: network concept, if in fact those obstacles weren't insurmountable, as 384 00:24:25,960 --> 00:24:30,480 Speaker 1: min Skin Separate had suggested, the US could potentially fall 385 00:24:30,720 --> 00:24:35,320 Speaker 1: behind another country because it would fail to fund its development. So, 386 00:24:35,920 --> 00:24:38,760 Speaker 1: in a desire not to have Japan take the ball 387 00:24:38,800 --> 00:24:41,439 Speaker 1: and run with it, the United States began to invest 388 00:24:41,680 --> 00:24:45,479 Speaker 1: again in artificial neural network research and development. In the 389 00:24:45,480 --> 00:24:50,920 Speaker 1: mid nineteen eighties, computer scientists essentially rediscovered the usefulness of 390 00:24:51,480 --> 00:24:55,639 Speaker 1: a process called back propagation. And I've already talked about 391 00:24:56,160 --> 00:24:58,159 Speaker 1: nodes and weights and stuff, but this is going to 392 00:24:58,160 --> 00:25:00,479 Speaker 1: require a little bit more explanation to under stand what 393 00:25:00,560 --> 00:25:03,760 Speaker 1: back propagation is all about. So let's kind of try 394 00:25:03,800 --> 00:25:07,560 Speaker 1: to visualize a neural network. So you've got your input nodes. 395 00:25:07,920 --> 00:25:10,240 Speaker 1: Just think of a bunch of circles. If you were 396 00:25:10,359 --> 00:25:12,160 Speaker 1: drawing it from top to bottom, this would be your 397 00:25:12,200 --> 00:25:15,679 Speaker 1: top layer. This is like the funnels where you're going 398 00:25:15,760 --> 00:25:19,639 Speaker 1: to feed data into the system. Now you've got a 399 00:25:19,640 --> 00:25:21,400 Speaker 1: whole bunch of these at the top and they can 400 00:25:21,440 --> 00:25:25,240 Speaker 1: accept the data that you're feeding in. They process that data, 401 00:25:25,640 --> 00:25:30,480 Speaker 1: and then based upon some operation, they will then send 402 00:25:30,760 --> 00:25:35,240 Speaker 1: an output to a node one layer down. So there's 403 00:25:35,280 --> 00:25:38,440 Speaker 1: lots of other nodes in the layers below, or maybe 404 00:25:38,480 --> 00:25:40,600 Speaker 1: not as many as you have initial layers. You might 405 00:25:40,600 --> 00:25:45,919 Speaker 1: actually have fewer, and the layers above will send to 406 00:25:46,760 --> 00:25:48,800 Speaker 1: you know, data to a specific node depending upon what 407 00:25:48,960 --> 00:25:54,280 Speaker 1: the outcome is. Whatever the output is, so these nodes 408 00:25:54,400 --> 00:25:58,199 Speaker 1: accept the input. These inputs have a bias and a 409 00:25:58,240 --> 00:26:01,520 Speaker 1: weight to them, and this is one of the hidden layers. 410 00:26:01,560 --> 00:26:04,320 Speaker 1: They will then create an output and send that on 411 00:26:04,480 --> 00:26:09,399 Speaker 1: to nodes another layer down. So this goes on until 412 00:26:09,440 --> 00:26:11,840 Speaker 1: you get to your output layer, where you get your 413 00:26:11,880 --> 00:26:15,399 Speaker 1: final result, and then you can determine whether or not 414 00:26:15,400 --> 00:26:18,800 Speaker 1: the final result matches what you were hoping for. So 415 00:26:18,840 --> 00:26:21,920 Speaker 1: did your system properly identify which photos do and don't 416 00:26:21,960 --> 00:26:24,800 Speaker 1: have cats in them? Now, as I mentioned earlier, you 417 00:26:24,840 --> 00:26:28,000 Speaker 1: typically get results that aren't perfect, but we want to 418 00:26:28,080 --> 00:26:32,760 Speaker 1: train the system to improve with every test. Back propagation 419 00:26:33,400 --> 00:26:36,800 Speaker 1: is one way to do this. So with that propagation, 420 00:26:37,359 --> 00:26:40,040 Speaker 1: you actually start with the final output. You've already done 421 00:26:40,040 --> 00:26:43,480 Speaker 1: a test run, right, and you've got your output, and 422 00:26:44,080 --> 00:26:49,160 Speaker 1: maybe your test has five possible final outcomes, but only 423 00:26:49,200 --> 00:26:52,119 Speaker 1: one of those is the outcome you actually want. Okay, 424 00:26:52,160 --> 00:26:55,359 Speaker 1: we'll say it's outcome number one. We're saying I want 425 00:26:55,359 --> 00:26:59,159 Speaker 1: this system to more often than not come to the 426 00:26:59,160 --> 00:27:01,840 Speaker 1: conclusion that's outcome number one. But you run your test. 427 00:27:02,200 --> 00:27:08,320 Speaker 1: It's got you one thousand little tasks in it, and 428 00:27:08,359 --> 00:27:11,840 Speaker 1: you run your test. You find out that it only 429 00:27:11,920 --> 00:27:14,399 Speaker 1: arrives at outcome number one five percent of the time, 430 00:27:14,640 --> 00:27:17,080 Speaker 1: which is actually worse than random chance. Right, it should 431 00:27:17,080 --> 00:27:19,359 Speaker 1: be twenty percent for random chance, but it's only getting 432 00:27:19,359 --> 00:27:22,359 Speaker 1: there five percent of the time. Something is going really 433 00:27:22,400 --> 00:27:26,199 Speaker 1: wrong with your system for it to mistakenly go to 434 00:27:26,240 --> 00:27:29,760 Speaker 1: one of the other options and very rarely go to 435 00:27:29,800 --> 00:27:32,960 Speaker 1: the correct one. So let's say you also noticed the 436 00:27:32,960 --> 00:27:36,080 Speaker 1: outcome number three. It goes to that one forty percent 437 00:27:36,080 --> 00:27:38,359 Speaker 1: of the time. So it's making this mistake forty percent 438 00:27:38,359 --> 00:27:40,400 Speaker 1: of the time and only getting it right five percent 439 00:27:40,440 --> 00:27:43,080 Speaker 1: of the time. So things are seriously out of whack. 440 00:27:43,160 --> 00:27:47,240 Speaker 1: You need to find which connections which would involve the 441 00:27:47,280 --> 00:27:51,159 Speaker 1: biases and the weights that are within your system that 442 00:27:51,200 --> 00:27:55,359 Speaker 1: are leading it to mistakenly arrive at the wrong outcome, 443 00:27:55,560 --> 00:27:59,960 Speaker 1: so frequently you want to reduce those factors, and simultaneously 444 00:28:00,119 --> 00:28:03,119 Speaker 1: you need to boost the ones that lead the system 445 00:28:03,280 --> 00:28:05,919 Speaker 1: to arrive at outcome number one, because that's the answer 446 00:28:06,000 --> 00:28:09,560 Speaker 1: you actually want the system to get to. All Right, 447 00:28:10,640 --> 00:28:12,720 Speaker 1: I've been droning on for a bit. Let's take another 448 00:28:12,800 --> 00:28:15,320 Speaker 1: quick break. When we come back, I'll finish up explaining 449 00:28:15,359 --> 00:28:27,840 Speaker 1: this and then we'll move on to catastrophic forgetting. Okay, 450 00:28:28,280 --> 00:28:31,280 Speaker 1: so we were talking about how you are looking at 451 00:28:31,320 --> 00:28:35,439 Speaker 1: a system that is coming to the wrong conclusion ninety 452 00:28:35,520 --> 00:28:38,560 Speaker 1: five percent of the time. It is a broken system. 453 00:28:38,880 --> 00:28:43,120 Speaker 1: You have to then figure out what factors are causing 454 00:28:43,120 --> 00:28:46,960 Speaker 1: this to happen, and they are numerous, right, They extend 455 00:28:47,000 --> 00:28:49,480 Speaker 1: all the way up to the very top of your 456 00:28:49,520 --> 00:28:52,160 Speaker 1: neural network, the other end where the input comes in. 457 00:28:52,520 --> 00:28:55,120 Speaker 1: But you can't just change everything all at once. You've 458 00:28:55,120 --> 00:28:58,480 Speaker 1: got to figure this out systematically, and that's what backpropagation 459 00:28:58,600 --> 00:29:03,240 Speaker 1: is really all about. Which links one layer up from 460 00:29:03,280 --> 00:29:07,200 Speaker 1: the output have the greatest impact on the outcome. Right, 461 00:29:07,880 --> 00:29:10,720 Speaker 1: changing everything would be tedious, it would be impractical. You 462 00:29:10,800 --> 00:29:14,120 Speaker 1: might even make things worse. Some of these neural networks 463 00:29:14,160 --> 00:29:19,320 Speaker 1: are confoundingly complicated, so it's not really a feasible solution. 464 00:29:19,680 --> 00:29:22,480 Speaker 1: So instead you look at the connections that are having 465 00:29:22,560 --> 00:29:25,760 Speaker 1: the biggest impact on your outcome. So you want things 466 00:29:25,800 --> 00:29:28,160 Speaker 1: where if you make a small change in either the 467 00:29:28,160 --> 00:29:31,080 Speaker 1: bias or the weight, or maybe both, you'll see a 468 00:29:31,200 --> 00:29:35,040 Speaker 1: larger end effect on the outcome. All the connections are 469 00:29:35,160 --> 00:29:39,040 Speaker 1: arguably important, but some are more important than others. Backpropagation 470 00:29:39,160 --> 00:29:41,880 Speaker 1: works backwards from the result toward the other end of 471 00:29:41,920 --> 00:29:44,720 Speaker 1: the network to tweak those connections. It boosts ones that 472 00:29:44,840 --> 00:29:48,840 Speaker 1: lead to the correct or desired response, and it reduces 473 00:29:48,880 --> 00:29:53,360 Speaker 1: the values of those that lead to incorrect or undesired responses. 474 00:29:53,720 --> 00:29:55,520 Speaker 1: If we were to think of this like the classic 475 00:29:55,600 --> 00:30:00,000 Speaker 1: example and chaos theory, this could potentially involve us studying 476 00:30:00,200 --> 00:30:02,840 Speaker 1: hurricane as it hits land and tracing its history back 477 00:30:02,880 --> 00:30:06,280 Speaker 1: as it moved through the ocean, and we would eventually 478 00:30:06,320 --> 00:30:09,240 Speaker 1: arrive at the point where it was a tropical storm, 479 00:30:09,320 --> 00:30:12,040 Speaker 1: and then we would go further back and see the 480 00:30:12,040 --> 00:30:14,960 Speaker 1: factors that led to the creation of that storm. And 481 00:30:15,000 --> 00:30:16,720 Speaker 1: maybe if we tracked it all the way back, we 482 00:30:16,760 --> 00:30:20,200 Speaker 1: would even find that one of a billion factors that 483 00:30:20,280 --> 00:30:23,480 Speaker 1: made the storm was in fact, a butterfly was flapping 484 00:30:23,560 --> 00:30:25,080 Speaker 1: its wings on the other side of the world and 485 00:30:25,120 --> 00:30:28,400 Speaker 1: that contributed to it. Maybe we find out that butterfly 486 00:30:28,400 --> 00:30:32,200 Speaker 1: flap of its wings had an impact, but it was negligible, 487 00:30:32,240 --> 00:30:33,960 Speaker 1: and that if the butterfly hadn't flapped its wings, the 488 00:30:34,040 --> 00:30:36,640 Speaker 1: hurricane still would have happened. That would be an example 489 00:30:36,680 --> 00:30:40,080 Speaker 1: of well, we don't bother adjusting the weight of the 490 00:30:40,400 --> 00:30:43,360 Speaker 1: of the impact of that butterfly flapping its wings because 491 00:30:43,360 --> 00:30:46,440 Speaker 1: it doesn't matter for the end result. But what if 492 00:30:46,480 --> 00:30:48,880 Speaker 1: we were to discover that that butterfly flap of its 493 00:30:48,920 --> 00:30:53,600 Speaker 1: wings is the only reason the hurricane happened that, or 494 00:30:53,640 --> 00:30:56,040 Speaker 1: at least was the primary reason that all the other 495 00:30:56,120 --> 00:30:59,120 Speaker 1: factors pale in comparison. Well, then we'd want to make 496 00:30:59,120 --> 00:31:04,040 Speaker 1: sure we boost the weight of that input, because clearly 497 00:31:04,080 --> 00:31:09,040 Speaker 1: that butterfly is fundamental for hurricanes. I think hurricanes are 498 00:31:09,080 --> 00:31:12,080 Speaker 1: really dangerous, and I would ask butterflies to kind of chill, 499 00:31:12,600 --> 00:31:16,000 Speaker 1: all right. I mean, I don't want butterflies to go away, 500 00:31:16,760 --> 00:31:20,560 Speaker 1: just you know, maybe stop flapping so much. Anyway, the 501 00:31:20,600 --> 00:31:24,760 Speaker 1: formula for backpropagation gets into some calculus that is well 502 00:31:24,760 --> 00:31:27,920 Speaker 1: beyond my knowledge and skill. So rather than attempt to 503 00:31:28,040 --> 00:31:32,040 Speaker 1: stumble my way through an explanation that I don't actually understand, 504 00:31:33,000 --> 00:31:34,760 Speaker 1: I think it's best to leave the concept at the 505 00:31:34,840 --> 00:31:37,760 Speaker 1: high level that I have described right now. So just 506 00:31:37,840 --> 00:31:39,960 Speaker 1: know that it gets way more granular than what I've 507 00:31:40,000 --> 00:31:44,160 Speaker 1: talked about. But essentially, you're looking at those factors that 508 00:31:44,320 --> 00:31:47,920 Speaker 1: led to the ultimate decision and saying which ones of 509 00:31:47,960 --> 00:31:51,440 Speaker 1: these had the greatest impact, and how can I tweak 510 00:31:51,520 --> 00:31:54,880 Speaker 1: them so that I can shape the outcome to one 511 00:31:54,920 --> 00:31:57,040 Speaker 1: I wanted. If we were thinking about that example I 512 00:31:57,080 --> 00:31:59,600 Speaker 1: gave about whether or not you go to the movies. 513 00:32:00,640 --> 00:32:06,280 Speaker 1: Maybe in present day you starts thinking about past experiences 514 00:32:06,320 --> 00:32:08,520 Speaker 1: where you made a decision to go out when you 515 00:32:08,560 --> 00:32:11,120 Speaker 1: had a big day in the following day, and how 516 00:32:11,680 --> 00:32:14,560 Speaker 1: that impacted you, perhaps negatively. Maybe you're like, man, I 517 00:32:14,560 --> 00:32:17,720 Speaker 1: should have gotten a promotion by now, and then you think, well, 518 00:32:17,760 --> 00:32:20,440 Speaker 1: I do go to the movies an awful lot. You 519 00:32:20,520 --> 00:32:23,200 Speaker 1: might say, I need to adjust some of the factors 520 00:32:23,240 --> 00:32:27,959 Speaker 1: that affect my decision making process and perhaps prioritize my career. 521 00:32:28,480 --> 00:32:33,280 Speaker 1: Or if you've decided that late stage capitalism is terrible 522 00:32:33,360 --> 00:32:35,960 Speaker 1: evil and that you're going to try and live a 523 00:32:36,000 --> 00:32:40,840 Speaker 1: hedonistic lifestyle of a wandering soul, maybe you say, I'm 524 00:32:40,880 --> 00:32:42,680 Speaker 1: going to go and see my movie with my friend, 525 00:32:43,080 --> 00:32:45,640 Speaker 1: and yeah, that's just how it is, because that's the 526 00:32:45,640 --> 00:32:47,560 Speaker 1: most important thing to me. You only go around this 527 00:32:47,640 --> 00:32:50,560 Speaker 1: crazy world once. After all, I'm not telling you which 528 00:32:50,560 --> 00:32:54,440 Speaker 1: way to go. I'm still finding my own way. But yeah, 529 00:32:54,480 --> 00:32:57,640 Speaker 1: back propagation would be how you would go back and say, 530 00:32:57,680 --> 00:33:00,880 Speaker 1: all right, well, because I don't like the outcome that happened, 531 00:33:01,360 --> 00:33:05,440 Speaker 1: I need to change the way. These factors weigh in 532 00:33:05,760 --> 00:33:09,400 Speaker 1: on the decision making process that goes through the whole system. Now, 533 00:33:09,440 --> 00:33:13,080 Speaker 1: the advancements in the science of neural networks proved that 534 00:33:13,120 --> 00:33:16,600 Speaker 1: the technology no longer operated under the constraints that concern 535 00:33:16,720 --> 00:33:19,920 Speaker 1: Minski and support in the late sixties, so once again 536 00:33:20,320 --> 00:33:24,520 Speaker 1: funding found its way to neural network research and development projects. 537 00:33:25,280 --> 00:33:29,840 Speaker 1: Now let's finally talk about forgetting and what makes it catastrophic. 538 00:33:30,640 --> 00:33:34,360 Speaker 1: So you could, in theory, develop an artificial neural network 539 00:33:34,720 --> 00:33:38,480 Speaker 1: and have a library of training data, and the only 540 00:33:38,560 --> 00:33:41,240 Speaker 1: thing you ever do with this network is you feed 541 00:33:41,320 --> 00:33:45,960 Speaker 1: that same set of training data to that same neural 542 00:33:46,040 --> 00:33:50,440 Speaker 1: network over and over in an effort to get performance 543 00:33:50,480 --> 00:33:53,720 Speaker 1: as close to perfect as you possibly can. Just you know, 544 00:33:53,840 --> 00:33:55,640 Speaker 1: it's kind of like if you have a car and 545 00:33:55,680 --> 00:33:59,400 Speaker 1: you're constantly tweaking it so it will perform better, and 546 00:33:59,480 --> 00:34:02,320 Speaker 1: maybe you chase one thing and it boosts performance in 547 00:34:02,360 --> 00:34:06,120 Speaker 1: one area, but it kind of negatively impacts performance in 548 00:34:06,160 --> 00:34:09,080 Speaker 1: another area, so then you got to tweak something else. 549 00:34:09,440 --> 00:34:11,480 Speaker 1: You could be doing that with an artificial neural network 550 00:34:11,520 --> 00:34:13,879 Speaker 1: forever and just be using the same set of training data. 551 00:34:14,320 --> 00:34:16,320 Speaker 1: And all you're trying to do is make a system 552 00:34:16,640 --> 00:34:19,399 Speaker 1: that could handle that training data better than any other 553 00:34:19,440 --> 00:34:21,920 Speaker 1: system in the world, and that would be interesting, but 554 00:34:22,000 --> 00:34:25,560 Speaker 1: it would be useless from a practical standpoint. You could say, like, hey, 555 00:34:25,600 --> 00:34:27,359 Speaker 1: you want to see my machine that can sort through 556 00:34:27,560 --> 00:34:31,360 Speaker 1: only this collection of photographs and pick out the ones 557 00:34:31,400 --> 00:34:33,200 Speaker 1: that have cats in them and the ones that don't. 558 00:34:33,520 --> 00:34:37,799 Speaker 1: Pretty pretty darn effectively, but not perfectly. It's not really 559 00:34:37,800 --> 00:34:41,560 Speaker 1: an interesting value proposition, right, So more likely you are 560 00:34:41,600 --> 00:34:44,400 Speaker 1: eventually going to start feeding lots of different kinds of 561 00:34:44,480 --> 00:34:48,759 Speaker 1: data to this neural network. And know, yeah, you train 562 00:34:48,840 --> 00:34:51,759 Speaker 1: the network on certain data sets, but your goal is 563 00:34:51,800 --> 00:34:54,440 Speaker 1: to feed new sets of data data the system has 564 00:34:54,480 --> 00:34:57,440 Speaker 1: never encountered before and rely on the system's ability to 565 00:34:57,560 --> 00:35:00,760 Speaker 1: process this information correctly to get the result you want. 566 00:35:01,320 --> 00:35:04,040 Speaker 1: And we might even be talking about stuff the human 567 00:35:04,080 --> 00:35:07,880 Speaker 1: beings can't easily do, right, But see, the training data 568 00:35:08,480 --> 00:35:10,200 Speaker 1: is going to mean that the network will start to 569 00:35:10,239 --> 00:35:15,000 Speaker 1: create and reinforce certain pathways, and those pathways will over 570 00:35:15,080 --> 00:35:17,359 Speaker 1: time get stronger and stronger, just as we said at 571 00:35:17,360 --> 00:35:20,520 Speaker 1: the beginning of this episode. But new data is going 572 00:35:20,560 --> 00:35:25,120 Speaker 1: to necessitate new pathways. Sometimes when the system begins to 573 00:35:25,160 --> 00:35:29,960 Speaker 1: form these new pathways, it forgets the old pathways. So 574 00:35:30,000 --> 00:35:32,880 Speaker 1: it's possible for a neural network to actually get worse 575 00:35:33,120 --> 00:35:36,080 Speaker 1: at the task it had previously been trained to do 576 00:35:37,080 --> 00:35:41,120 Speaker 1: with the actual training material. In fact, in a true catastrophe, 577 00:35:41,160 --> 00:35:45,440 Speaker 1: the system might forget the objective and doesn't recognize what 578 00:35:45,480 --> 00:35:48,400 Speaker 1: the desired outcome is meant to be, so the results 579 00:35:48,440 --> 00:35:51,480 Speaker 1: can appear random and meaningless. It's as if the system 580 00:35:51,520 --> 00:35:55,440 Speaker 1: has developed some form of amnesia. So this is prevalent, 581 00:35:56,000 --> 00:36:00,600 Speaker 1: most prevalent anyway, in systems that rely on unguided learning. 582 00:36:01,200 --> 00:36:06,120 Speaker 1: With guided learning, you have engineers who are carefully selecting 583 00:36:06,160 --> 00:36:10,839 Speaker 1: the data that gets fed into a system. An unguided 584 00:36:10,840 --> 00:36:15,160 Speaker 1: system would collect raw data from wherever and attempt to 585 00:36:15,200 --> 00:36:18,560 Speaker 1: deliver desired results, and that those are the kinds of 586 00:36:19,280 --> 00:36:23,000 Speaker 1: neural networks that are more prone to catastrophic forgetting. But 587 00:36:23,080 --> 00:36:27,359 Speaker 1: as I said, machine learning systems tackle new data, maybe 588 00:36:27,360 --> 00:36:31,239 Speaker 1: even new tasks, and then you get the risk of 589 00:36:31,280 --> 00:36:34,080 Speaker 1: the system forgetting stuff. So I jokingly say, it's kind 590 00:36:34,080 --> 00:36:36,200 Speaker 1: of like when I learned something new, it has to 591 00:36:36,200 --> 00:36:39,680 Speaker 1: push out something old, like you know, my friend's phone 592 00:36:39,760 --> 00:36:42,160 Speaker 1: number or something. Suddenly I can no longer remember it 593 00:36:42,239 --> 00:36:45,640 Speaker 1: because I learned some new interesting fact, as if I 594 00:36:45,719 --> 00:36:49,000 Speaker 1: have met my capacity for being able to know things. 595 00:36:49,200 --> 00:36:52,759 Speaker 1: So learning anything new necessitates having to forget something I 596 00:36:52,880 --> 00:36:56,160 Speaker 1: used to know, like gat Ye, because now gat Ye 597 00:36:56,320 --> 00:36:59,439 Speaker 1: is just somebody that I used to know. But wait, 598 00:36:59,680 --> 00:37:04,680 Speaker 1: there's more. Just as a system can experience catastrophic forgetting, 599 00:37:05,400 --> 00:37:10,920 Speaker 1: it can also experience catastrophic remembering. This is when a 600 00:37:10,960 --> 00:37:15,200 Speaker 1: system mistakenly believes it is doing one process, a task 601 00:37:15,320 --> 00:37:19,160 Speaker 1: it had previously been trained to do, rather than the 602 00:37:19,200 --> 00:37:23,040 Speaker 1: one it's actually trying to do. So let's say we've 603 00:37:23,040 --> 00:37:26,359 Speaker 1: got an artificial neural network, and originally we taught it 604 00:37:26,400 --> 00:37:28,920 Speaker 1: to recognize the photos that have cats in them versus 605 00:37:28,920 --> 00:37:31,960 Speaker 1: the ones that don't. But now we have retrained the 606 00:37:32,080 --> 00:37:36,480 Speaker 1: same artificial neural network to try and recognize handwritten text. 607 00:37:37,239 --> 00:37:41,080 Speaker 1: Except when we feed handwritten text to the system, suddenly 608 00:37:41,120 --> 00:37:44,560 Speaker 1: the system believes it's trying to determine where the cats are. 609 00:37:45,160 --> 00:37:47,719 Speaker 1: This is something that can happen with machine learning systems too, 610 00:37:47,719 --> 00:37:50,560 Speaker 1: and you still get bad results out of it. So 611 00:37:50,680 --> 00:37:55,120 Speaker 1: this is a real problem. Now, these are not insurmountable problems. 612 00:37:55,680 --> 00:37:59,520 Speaker 1: There are some solutions that are actually intuitive. For example, 613 00:38:00,120 --> 00:38:03,480 Speaker 1: any gamer out there knows that it's best to save 614 00:38:03,560 --> 00:38:06,400 Speaker 1: your game just before you head into a big boss battle, 615 00:38:06,680 --> 00:38:09,879 Speaker 1: just in case things don't go the way you planned well. 616 00:38:09,880 --> 00:38:13,279 Speaker 1: With artificial neural networks, it's maybe not a bad idea 617 00:38:13,360 --> 00:38:16,640 Speaker 1: to make a copy of a network before you retrain 618 00:38:16,719 --> 00:38:19,200 Speaker 1: it to do something new. Then you still have the 619 00:38:19,200 --> 00:38:23,080 Speaker 1: backup if things do go pair shape. There are other 620 00:38:23,120 --> 00:38:27,799 Speaker 1: approaches to decreasing the risk of catastrophic forgetting or catastrophic remembering. 621 00:38:28,280 --> 00:38:32,400 Speaker 1: An article in Applied Mathematics titled Overcoming Catastrophic forgetting a 622 00:38:32,440 --> 00:38:36,560 Speaker 1: neural networks describes a system in which the researchers purposefully 623 00:38:36,600 --> 00:38:41,640 Speaker 1: slowed down the network's ability to change the weights involved 624 00:38:41,719 --> 00:38:47,359 Speaker 1: in important tasks from previous training cycles. So this makes 625 00:38:47,400 --> 00:38:50,239 Speaker 1: teaching the system to do new tasks a little more 626 00:38:50,320 --> 00:38:57,360 Speaker 1: challenging because it's protecting these weights. It's preventing the system's 627 00:38:57,360 --> 00:39:02,600 Speaker 1: ability to be completely plasid, which means the system has 628 00:39:02,640 --> 00:39:05,160 Speaker 1: to work around these constraints and still learn how to 629 00:39:05,160 --> 00:39:08,279 Speaker 1: do the new task, but in the process it means 630 00:39:08,320 --> 00:39:12,120 Speaker 1: it doesn't forget how to do the previous tasks. This 631 00:39:12,239 --> 00:39:15,920 Speaker 1: article is interesting because the tasks the researchers actually used 632 00:39:16,040 --> 00:39:18,600 Speaker 1: the purposes of training, Like, what were they teaching the 633 00:39:18,680 --> 00:39:21,440 Speaker 1: artificial neural network to do well? They were teaching it 634 00:39:21,640 --> 00:39:24,680 Speaker 1: how to play atari twenty six hundred games. So they 635 00:39:24,680 --> 00:39:27,879 Speaker 1: would start with one game and train the system on 636 00:39:27,960 --> 00:39:31,640 Speaker 1: how to play the game. Then they would give the 637 00:39:31,680 --> 00:39:36,160 Speaker 1: system a new game with different game mechanics, and the 638 00:39:36,200 --> 00:39:38,840 Speaker 1: system would have to learn how to play this new game, 639 00:39:39,440 --> 00:39:41,520 Speaker 1: but they wanted to see if it could still remember 640 00:39:41,520 --> 00:39:43,840 Speaker 1: how to play the original game. That was kind of 641 00:39:43,880 --> 00:39:46,320 Speaker 1: the system they were working on. They were tweaking things 642 00:39:46,920 --> 00:39:51,360 Speaker 1: so that the machine learning artificial neural network as a 643 00:39:51,360 --> 00:39:54,400 Speaker 1: whole could learn how to play multiple Atari twenty six 644 00:39:54,480 --> 00:39:57,400 Speaker 1: hundred games without forgetting how to do the previous ones. 645 00:39:57,840 --> 00:40:00,000 Speaker 1: This is a non trivial task. I mean, it takes 646 00:40:00,040 --> 00:40:03,120 Speaker 1: a lot of work to see exactly how to preserve 647 00:40:03,200 --> 00:40:05,960 Speaker 1: things so that you're not slowing down the learning process 648 00:40:05,960 --> 00:40:08,640 Speaker 1: too much, but you're also not inviting the possibility of 649 00:40:08,640 --> 00:40:13,480 Speaker 1: catastrophic forgetting. Now, that's just one example of how researchers 650 00:40:13,480 --> 00:40:17,080 Speaker 1: are looking to mitigate the problem of catastrophic forgetting in 651 00:40:17,080 --> 00:40:20,719 Speaker 1: catastrophic remembering. There are other methods as well, and maybe 652 00:40:20,760 --> 00:40:23,719 Speaker 1: I'll do another episode where I'll go into more detail 653 00:40:23,960 --> 00:40:27,480 Speaker 1: on some of those. They do get pretty complicated, and 654 00:40:27,480 --> 00:40:31,359 Speaker 1: in fact, eventually Rerilli and I even eventually pretty early 655 00:40:31,400 --> 00:40:35,520 Speaker 1: on I hit my limit for as far as I 656 00:40:35,560 --> 00:40:39,319 Speaker 1: can understand the actual mechanics of the system. So rather 657 00:40:39,400 --> 00:40:43,880 Speaker 1: than you know, try and punch above my weight, I 658 00:40:43,920 --> 00:40:46,640 Speaker 1: think it's best to kind of be a little more general, 659 00:40:47,360 --> 00:40:49,400 Speaker 1: but just to have that understanding to kind of get 660 00:40:49,400 --> 00:40:52,920 Speaker 1: a better appreciation of some of the challenges relating to 661 00:40:53,120 --> 00:40:58,240 Speaker 1: artificial intelligence in general and machine learning in particular. And again, 662 00:40:58,400 --> 00:41:01,560 Speaker 1: like this machine learning issue, you it's really a bigger 663 00:41:01,640 --> 00:41:06,120 Speaker 1: problem with more sophisticated systems that are meant to do 664 00:41:06,400 --> 00:41:09,840 Speaker 1: unsupervised and unguided learning, right, those are the ones that 665 00:41:09,880 --> 00:41:12,520 Speaker 1: are going to be more prone to these issues. If 666 00:41:12,520 --> 00:41:17,640 Speaker 1: we're talking about supervised and guided learning, where engineers are 667 00:41:18,239 --> 00:41:20,880 Speaker 1: being very careful with the data being fed to a system, 668 00:41:21,400 --> 00:41:25,560 Speaker 1: it's less likely to happen. But the whole promise, or 669 00:41:26,200 --> 00:41:29,080 Speaker 1: at least the you know, not the promise of the 670 00:41:29,080 --> 00:41:31,120 Speaker 1: technology itself, but the promise of the people who are 671 00:41:31,520 --> 00:41:34,680 Speaker 1: funding it, is that this technology is going to reach 672 00:41:34,680 --> 00:41:38,160 Speaker 1: a point where it's able to learn on its own 673 00:41:38,239 --> 00:41:41,640 Speaker 1: and be able to do things better than people can do, 674 00:41:41,760 --> 00:41:44,000 Speaker 1: to free us up to doing, you know, stuff we 675 00:41:44,040 --> 00:41:45,799 Speaker 1: want to do instead of stuff we have to do. 676 00:41:46,560 --> 00:41:49,919 Speaker 1: That's like the science fiction dream version of AI. As 677 00:41:49,960 --> 00:41:53,759 Speaker 1: we all know, getting there is much more painful. It's 678 00:41:53,800 --> 00:41:57,600 Speaker 1: not like a simple process of Hey, we've made everything 679 00:41:57,920 --> 00:42:00,759 Speaker 1: easy to do now and you don't have to work 680 00:42:00,800 --> 00:42:03,160 Speaker 1: all day. You can enjoy your life and pursue your 681 00:42:03,239 --> 00:42:07,279 Speaker 1: dreams and develop your hobbies and your interests, and you 682 00:42:07,320 --> 00:42:10,880 Speaker 1: can have fulfillment and somehow money isn't important anymore. Like 683 00:42:10,960 --> 00:42:13,000 Speaker 1: that seems to be the Star Trek version of the 684 00:42:13,000 --> 00:42:15,040 Speaker 1: future that people want it to go in. But as 685 00:42:15,040 --> 00:42:17,839 Speaker 1: we have seen, the process of getting there is way 686 00:42:17,840 --> 00:42:20,719 Speaker 1: more painful. As you know, people face a reality of 687 00:42:21,320 --> 00:42:25,800 Speaker 1: potentially being out of work because of AI, or maybe 688 00:42:25,840 --> 00:42:30,000 Speaker 1: being paid way less to do work because the AI 689 00:42:30,160 --> 00:42:33,319 Speaker 1: is doing most of it. These are not that's not 690 00:42:33,400 --> 00:42:36,800 Speaker 1: Star Trek feature. That's getting like into Blade Runner future, 691 00:42:37,239 --> 00:42:41,520 Speaker 1: So we don't want that one. By the way, the 692 00:42:41,600 --> 00:42:44,120 Speaker 1: tears in the rain speech is fantastic, but you do 693 00:42:44,120 --> 00:42:46,200 Speaker 1: not want to live in the Blade Runner world. Trust me. 694 00:42:47,440 --> 00:42:48,799 Speaker 1: You might not want to live in the Star Trek 695 00:42:48,800 --> 00:42:51,680 Speaker 1: world either, because those outfits don't look that comfortable anyway. 696 00:42:52,520 --> 00:42:57,080 Speaker 1: That's my little discussion about AI, machine learning and cast 697 00:42:57,120 --> 00:43:02,120 Speaker 1: trophic forgetting in castrophic. Remembering this is just one of 698 00:43:02,200 --> 00:43:05,439 Speaker 1: the challenges associated with AI and machine learning. I don't 699 00:43:05,520 --> 00:43:08,160 Speaker 1: mean to suggest it's the one and only, or even 700 00:43:08,160 --> 00:43:11,239 Speaker 1: that it's the most important one, but it is one 701 00:43:11,280 --> 00:43:13,560 Speaker 1: that I had not really heard of until I listened 702 00:43:13,560 --> 00:43:16,480 Speaker 1: to that Skeptics Guide to the Universe episode over the weekend, 703 00:43:17,040 --> 00:43:20,720 Speaker 1: and it was really interesting to dive into the material 704 00:43:20,800 --> 00:43:22,600 Speaker 1: and read up about it and to get a better 705 00:43:22,680 --> 00:43:25,960 Speaker 1: understanding of what it means and how it works. I 706 00:43:26,000 --> 00:43:28,600 Speaker 1: hope you liked that episode from last year, twenty twenty three, 707 00:43:28,719 --> 00:43:32,600 Speaker 1: machine Learning and Catastrophic Forgetting. I am working on other 708 00:43:32,640 --> 00:43:35,440 Speaker 1: episodes that relate to AI. I also want to do 709 00:43:35,480 --> 00:43:40,000 Speaker 1: an episode about companies that claim to be part of 710 00:43:40,040 --> 00:43:43,439 Speaker 1: the artificial intelligence space but in fact use little if 711 00:43:43,560 --> 00:43:47,919 Speaker 1: any AI technology, because that has become a thing. As 712 00:43:47,920 --> 00:43:51,680 Speaker 1: we all know, when there is the combination of huge 713 00:43:51,680 --> 00:43:55,640 Speaker 1: amounts of money and low amounts of understanding, you have 714 00:43:56,120 --> 00:44:00,399 Speaker 1: the perfect breeding ground for scams and con artists and 715 00:44:00,400 --> 00:44:03,640 Speaker 1: that kind of thing. So I do plan on doing 716 00:44:03,680 --> 00:44:08,560 Speaker 1: an episode about various startups that claim at some level 717 00:44:08,680 --> 00:44:12,680 Speaker 1: to be part of artificial intelligence, but when you really 718 00:44:12,880 --> 00:44:16,160 Speaker 1: start to examine them, have little to no connection to 719 00:44:16,200 --> 00:44:18,879 Speaker 1: that world. So being on the lookout for that, It's 720 00:44:19,040 --> 00:44:21,040 Speaker 1: going to take me some time to do some research 721 00:44:21,040 --> 00:44:23,839 Speaker 1: because there's lots of different sources to go through on 722 00:44:23,880 --> 00:44:26,840 Speaker 1: that one. But that's what I'm working on for probably 723 00:44:26,920 --> 00:44:29,920 Speaker 1: next week. I'm hoping next week. In the meantime, for 724 00:44:30,040 --> 00:44:32,359 Speaker 1: those of you here in the United States, I hope 725 00:44:32,360 --> 00:44:36,680 Speaker 1: you have a safe Fourth of July celebration. Make sure 726 00:44:37,000 --> 00:44:39,839 Speaker 1: that you spend time with friends and loved ones, and 727 00:44:40,280 --> 00:44:42,440 Speaker 1: you know, be very careful if you're going to be 728 00:44:42,480 --> 00:44:45,840 Speaker 1: around fireworks. Those things are very dangerous for everyone else 729 00:44:45,880 --> 00:44:48,520 Speaker 1: out there who's not celebrating a holiday and fourth of July. 730 00:44:48,719 --> 00:44:53,360 Speaker 1: I hope you have an excellent Fourth of July wherever 731 00:44:53,400 --> 00:44:56,759 Speaker 1: you are, and that whatever you enjoy doing, you get 732 00:44:56,760 --> 00:44:59,520 Speaker 1: to do a lot of it on the Fourth of July. 733 00:45:00,000 --> 00:45:02,080 Speaker 1: As long as it's you know, not hurting yourself or 734 00:45:02,120 --> 00:45:05,360 Speaker 1: other people. That's it for me. I will talk to 735 00:45:05,400 --> 00:45:16,120 Speaker 1: you again really soon. Tech Stuff is an iHeartRadio production. 736 00:45:16,440 --> 00:45:21,480 Speaker 1: For more podcasts from iHeartRadio, visit the iHeartRadio app, Apple Podcasts, 737 00:45:21,600 --> 00:45:23,600 Speaker 1: or wherever you listen to your favorite shows.