1 00:00:00,000 --> 00:00:02,840 All right, let's dive into neuromorphic programming. 2 00:00:02,840 --> 00:00:05,840 Emerging directions for brain-inspired hardware. 3 00:00:06,720 --> 00:00:09,840 You know, it's all about figuring out how to program 4 00:00:09,840 --> 00:00:11,480 those brain-inspired computers. 5 00:00:11,480 --> 00:00:13,080 Yeah, those brain-inspired computers 6 00:00:13,080 --> 00:00:15,640 everyone's talking about really are fascinating. 7 00:00:15,640 --> 00:00:17,720 They're designed to process information more 8 00:00:17,720 --> 00:00:18,560 like our brains do. 9 00:00:18,560 --> 00:00:21,040 Yeah, potentially revolutionizing computing. 10 00:00:21,040 --> 00:00:23,680 But as this paper points out, 11 00:00:23,680 --> 00:00:26,040 figuring out how to actually program them 12 00:00:26,040 --> 00:00:27,560 is a huge challenge. 13 00:00:27,560 --> 00:00:29,440 Right, it's a huge E challenge. 14 00:00:29,440 --> 00:00:32,520 Because we can't just rely on the traditional methods we use. 15 00:00:32,520 --> 00:00:35,520 The traditional methods for programming regular computers 16 00:00:35,520 --> 00:00:37,160 just won't cut it. 17 00:00:37,160 --> 00:00:38,720 Okay, so why is that? 18 00:00:38,720 --> 00:00:41,840 Well, the paper highlights five key differences. 19 00:00:41,840 --> 00:00:43,920 Five key differences between traditional 20 00:00:43,920 --> 00:00:45,240 and brain-inspired computers. 21 00:00:45,240 --> 00:00:46,080 Yeah. 22 00:00:46,080 --> 00:00:47,600 But I think for our deep dive today, 23 00:00:47,600 --> 00:00:50,160 we should focus on the two that are most mind-boggling 24 00:00:50,160 --> 00:00:52,840 and really impact how we have to think about programming. 25 00:00:52,840 --> 00:00:54,040 Sounds good to me. 26 00:00:54,040 --> 00:00:56,040 So, let's start with plasticity. 27 00:00:56,040 --> 00:00:56,880 Okay. 28 00:00:56,880 --> 00:00:58,920 In traditional computers, the hardware stays the same 29 00:00:58,920 --> 00:01:00,440 no matter what program is running. 30 00:01:00,440 --> 00:01:04,160 But brain-inspired computers are different. 31 00:01:04,160 --> 00:01:05,040 Different how? 32 00:01:05,040 --> 00:01:08,960 They actually physically change during operation. 33 00:01:08,960 --> 00:01:09,800 Whoa! 34 00:01:09,800 --> 00:01:13,520 Imagine a computer that rewires itself as it learns. 35 00:01:13,520 --> 00:01:14,520 That's wild. 36 00:01:14,520 --> 00:01:16,520 How do we even begin to control that? 37 00:01:16,520 --> 00:01:18,160 That's one of the big questions researchers 38 00:01:18,160 --> 00:01:19,000 are grappling with. 39 00:01:19,000 --> 00:01:19,840 Yeah. 40 00:01:19,840 --> 00:01:22,680 It leads us to the second key difference, stochasticity. 41 00:01:22,680 --> 00:01:23,840 Stochasticity. 42 00:01:23,840 --> 00:01:24,960 In traditional computers, 43 00:01:24,960 --> 00:01:26,720 we want everything to be predictable. 44 00:01:26,720 --> 00:01:27,720 Yeah, you give an input, 45 00:01:27,720 --> 00:01:29,080 you get a specific output. 46 00:01:29,080 --> 00:01:29,920 Right. 47 00:01:29,920 --> 00:01:33,640 But brain-inspired systems are inherently noisy. 48 00:01:33,640 --> 00:01:34,720 Noisy. 49 00:01:34,720 --> 00:01:35,960 There's randomness built in. 50 00:01:35,960 --> 00:01:37,680 So, it's like trying to program a computer 51 00:01:37,680 --> 00:01:38,960 that has a mind of its own. 52 00:01:38,960 --> 00:01:39,960 Well, kind of. 53 00:01:39,960 --> 00:01:40,800 Yeah. 54 00:01:40,800 --> 00:01:43,280 Making decisions based on probability rather than strict rules. 55 00:01:43,280 --> 00:01:44,120 Exactly. 56 00:01:44,120 --> 00:01:45,560 Debugging that must be a nightmare. 57 00:01:45,560 --> 00:01:46,600 You can say that. 58 00:01:46,600 --> 00:01:50,560 So, these two differences, plasticity and stochasticity, 59 00:01:50,560 --> 00:01:52,720 really force us to rethink 60 00:01:52,720 --> 00:01:54,880 what it even means to program a computer. 61 00:01:54,880 --> 00:01:55,720 Exactly. 62 00:01:55,720 --> 00:01:57,440 Our usual approach is to give the computer 63 00:01:57,440 --> 00:01:59,360 a step-by-step list of instructions. 64 00:01:59,360 --> 00:02:01,200 Yeah, that's the traditional way. 65 00:02:01,200 --> 00:02:02,560 But if the hardware's changing 66 00:02:02,560 --> 00:02:04,200 and there's randomness involved, 67 00:02:04,200 --> 00:02:06,880 those instructions might not have the intended effect. 68 00:02:06,880 --> 00:02:07,720 Right. 69 00:02:07,720 --> 00:02:09,920 This paper argues that we need to move away 70 00:02:09,920 --> 00:02:14,000 from that traditional command and control style of programming 71 00:02:14,000 --> 00:02:16,600 and embrace new approaches that are more aligned 72 00:02:16,600 --> 00:02:18,360 with how brains work. 73 00:02:18,360 --> 00:02:20,360 So, what are some of those approaches? 74 00:02:20,360 --> 00:02:22,160 Well, the paper mentions a whole bunch of them. 75 00:02:22,160 --> 00:02:23,000 It does. 76 00:02:23,000 --> 00:02:25,680 Before we dive into specific methods, 77 00:02:25,680 --> 00:02:28,800 I think it's helpful to understand this overarching idea 78 00:02:28,800 --> 00:02:30,200 the author's present. 79 00:02:31,200 --> 00:02:33,120 Computing with physical systems. 80 00:02:33,120 --> 00:02:34,720 Computing with physical systems. 81 00:02:34,720 --> 00:02:36,560 It suggests that we need to consider 82 00:02:36,560 --> 00:02:38,920 the physical properties of the hardware 83 00:02:38,920 --> 00:02:41,040 as an integral part of the computation. 84 00:02:41,040 --> 00:02:43,200 So, we can't just treat the brain-inspired computer 85 00:02:43,200 --> 00:02:44,160 as a black box. 86 00:02:44,160 --> 00:02:45,640 We can't just feed it instructions 87 00:02:45,640 --> 00:02:47,200 and expect predictable outcomes. 88 00:02:47,200 --> 00:02:48,040 Right. 89 00:02:48,040 --> 00:02:50,160 We need to understand how its physical structure 90 00:02:50,160 --> 00:02:53,320 and dynamics are shaping the computation itself. 91 00:02:53,320 --> 00:02:55,280 It's like the difference between playing a piano 92 00:02:55,280 --> 00:02:56,600 and listening to a recording. 93 00:02:56,600 --> 00:02:57,440 Yes. 94 00:02:57,440 --> 00:02:59,480 With a recording, the music is fixed. 95 00:02:59,480 --> 00:03:02,000 But with a piano, the sound you produce 96 00:03:02,000 --> 00:03:04,360 depends on how you interact with the physical instrument. 97 00:03:04,360 --> 00:03:05,200 Exactly. 98 00:03:05,200 --> 00:03:06,440 It's a great analogy. 99 00:03:06,440 --> 00:03:08,760 So, in neuromorphic computing, 100 00:03:08,760 --> 00:03:11,760 the program isn't just a set of instructions. 101 00:03:11,760 --> 00:03:14,480 It's more like a description of how the computation 102 00:03:14,480 --> 00:03:16,560 will emerge from the physical interactions 103 00:03:16,560 --> 00:03:18,360 within the system. 104 00:03:18,360 --> 00:03:19,800 And that leads to a whole new way 105 00:03:19,800 --> 00:03:21,960 of thinking about programming languages. 106 00:03:21,960 --> 00:03:22,800 Right. 107 00:03:22,800 --> 00:03:24,120 Because in traditional programming, 108 00:03:24,120 --> 00:03:27,160 we use languages to write those step-by-step instructions. 109 00:03:27,160 --> 00:03:28,000 Exactly. 110 00:03:28,000 --> 00:03:30,640 But if we're not giving explicit instructions anymore, 111 00:03:30,640 --> 00:03:33,160 then what role do programming languages play? 112 00:03:33,160 --> 00:03:34,000 Yeah. 113 00:03:34,000 --> 00:03:35,720 What role do they play in this new framework? 114 00:03:35,720 --> 00:03:38,560 Well, the authors use the term programming language 115 00:03:38,560 --> 00:03:40,640 in a much broader sense. 116 00:03:40,640 --> 00:03:41,480 Broader how? 117 00:03:41,480 --> 00:03:44,000 It could be any formal language that allows us 118 00:03:44,000 --> 00:03:45,960 to communicate with the physical system 119 00:03:45,960 --> 00:03:48,720 and shape how the computation unfolds. 120 00:03:48,720 --> 00:03:50,840 So, instead of thinking of programming languages, 121 00:03:50,840 --> 00:03:53,040 it's just ways to write instructions. 122 00:03:53,040 --> 00:03:55,280 We need to consider them as ways to describe 123 00:03:55,280 --> 00:03:57,800 the computation that will emerge from the physical system. 124 00:03:57,800 --> 00:03:58,640 Precisely. 125 00:03:58,640 --> 00:04:02,080 And this leads us to a discussion of programming paradigms. 126 00:04:02,080 --> 00:04:03,440 Programming paradigm. 127 00:04:03,440 --> 00:04:06,080 The different approaches we can take to programming 128 00:04:06,080 --> 00:04:08,200 based on certain principles or theories. 129 00:04:08,200 --> 00:04:09,040 Okay. 130 00:04:09,040 --> 00:04:10,400 So, let's unpack some of those paradigms. 131 00:04:10,400 --> 00:04:11,240 Right. 132 00:04:11,240 --> 00:04:12,440 The paper talks about a whole bunch of them, 133 00:04:12,440 --> 00:04:14,840 from the familiar to the downright futuristic. 134 00:04:14,840 --> 00:04:15,680 It does. 135 00:04:15,680 --> 00:04:18,520 It covers traditional paradigms like imperative programming. 136 00:04:18,520 --> 00:04:19,640 Imperative programming. 137 00:04:19,640 --> 00:04:22,360 Where you give the computer a sequence of commands. 138 00:04:22,360 --> 00:04:23,000 Right. 139 00:04:23,000 --> 00:04:24,200 Declarative programming. 140 00:04:24,200 --> 00:04:25,360 Declarative programming. 141 00:04:25,360 --> 00:04:27,840 Where you describe the desired outcome. 142 00:04:27,840 --> 00:04:28,400 Okay. 143 00:04:28,400 --> 00:04:30,560 Rather than the specific steps. 144 00:04:30,560 --> 00:04:31,000 Yeah. 145 00:04:31,000 --> 00:04:33,000 But then it gets really interesting when it dives 146 00:04:33,000 --> 00:04:34,640 into automated programming. 147 00:04:34,640 --> 00:04:35,880 Automated programming. 148 00:04:35,880 --> 00:04:36,200 Yeah. 149 00:04:36,200 --> 00:04:38,480 That sounds like something straight out of a sci-fi movie. 150 00:04:38,480 --> 00:04:39,440 It kind of is. 151 00:04:39,440 --> 00:04:41,880 Are we talking about computers that program themselves? 152 00:04:41,880 --> 00:04:43,920 In a way, yes. 153 00:04:43,920 --> 00:04:47,080 Automated programming involves using algorithms 154 00:04:47,080 --> 00:04:48,680 to generate programs. 155 00:04:48,680 --> 00:04:49,560 Okay. 156 00:04:49,560 --> 00:04:53,080 Often based on a set of specifications or examples. 157 00:04:53,080 --> 00:04:55,080 So, the computer is doing the programming. 158 00:04:55,080 --> 00:04:56,480 In a sense, yes. 159 00:04:56,480 --> 00:04:57,480 That's wild. 160 00:04:57,480 --> 00:05:01,200 The paper highlights techniques like metaprogramming, 161 00:05:01,200 --> 00:05:03,760 program synthesis, constraint programming, 162 00:05:03,760 --> 00:05:05,280 and inductive programming. 163 00:05:05,280 --> 00:05:08,200 So, it's a way to let the computer take on more 164 00:05:08,200 --> 00:05:10,120 of the heavy lifting in the programming process. 165 00:05:10,120 --> 00:05:11,080 Exactly. 166 00:05:11,080 --> 00:05:11,360 Okay. 167 00:05:11,360 --> 00:05:12,520 So, that's automated programming. 168 00:05:12,520 --> 00:05:13,640 Right. 169 00:05:13,640 --> 00:05:16,800 What other paradigm shifting approaches 170 00:05:16,800 --> 00:05:18,320 does this paper explore? 171 00:05:18,320 --> 00:05:21,800 Well, one that's particularly relevant to neuromorphic 172 00:05:21,800 --> 00:05:23,800 computing is machine learning. 173 00:05:23,800 --> 00:05:25,120 Machine learning. 174 00:05:25,120 --> 00:05:26,240 Everyone's talking about it these days. 175 00:05:26,240 --> 00:05:26,840 It seems like it. 176 00:05:26,840 --> 00:05:29,760 How does it fit into this brain-inspired world? 177 00:05:29,760 --> 00:05:31,640 Well, machine learning can be seen 178 00:05:31,640 --> 00:05:33,760 as a type of automated programming. 179 00:05:33,760 --> 00:05:34,280 Okay. 180 00:05:34,280 --> 00:05:37,560 Instead of explicitly programming every step, 181 00:05:37,560 --> 00:05:40,960 we provide the system with data. 182 00:05:40,960 --> 00:05:43,640 And let it learn the desired behavior. 183 00:05:43,640 --> 00:05:46,560 So, in the context of neuromorphic computing, 184 00:05:46,560 --> 00:05:49,880 we're essentially training the brain-inspired computer 185 00:05:49,880 --> 00:05:51,720 to perform the desired computation. 186 00:05:51,720 --> 00:05:52,240 Needs. 187 00:05:52,240 --> 00:05:55,120 Much like we train our own brains to learn new skills. 188 00:05:55,120 --> 00:05:56,080 That's a good way to put it. 189 00:05:56,080 --> 00:05:58,440 So, we're teaching the computer instead of programming it. 190 00:05:58,440 --> 00:05:58,920 Yes. 191 00:05:58,920 --> 00:06:02,560 And the paper talks about various machine learning 192 00:06:02,560 --> 00:06:05,280 paradigms that are relevant here, including supervised 193 00:06:05,280 --> 00:06:08,520 learning, reinforcement learning, and, of course, deep learning. 194 00:06:08,520 --> 00:06:09,440 Deep learning. 195 00:06:09,440 --> 00:06:11,400 The rock star of the AI world. 196 00:06:11,400 --> 00:06:12,000 Right. 197 00:06:12,000 --> 00:06:15,480 But adapting deep learning for neuromorphic systems 198 00:06:15,480 --> 00:06:17,200 presents unique challenges. 199 00:06:17,200 --> 00:06:19,360 Because of those key differences we talked about earlier. 200 00:06:19,360 --> 00:06:20,080 Exactly. 201 00:06:20,080 --> 00:06:22,600 Plasticity and stochasticity. 202 00:06:22,600 --> 00:06:25,280 So, we can't just take the deep learning algorithms that 203 00:06:25,280 --> 00:06:27,280 work so well in traditional computing. 204 00:06:27,280 --> 00:06:30,880 And apply them directly to neuromorphic systems. 205 00:06:30,880 --> 00:06:31,520 Right. 206 00:06:31,520 --> 00:06:32,560 Not quite. 207 00:06:32,560 --> 00:06:34,040 There's got to be some adaptation. 208 00:06:34,040 --> 00:06:35,480 There's a lot of research going on 209 00:06:35,480 --> 00:06:39,000 to figure out how to adapt deep learning for these brain-inspired 210 00:06:39,000 --> 00:06:39,760 computers. 211 00:06:39,760 --> 00:06:41,920 So, we're just scratching the surface of understanding 212 00:06:41,920 --> 00:06:44,640 how to harness the power of deep learning 213 00:06:44,640 --> 00:06:46,000 in the neuromorphic world. 214 00:06:46,000 --> 00:06:47,280 Yeah, I'd say so. 215 00:06:47,280 --> 00:06:49,520 What other approaches are being explored? 216 00:06:49,520 --> 00:06:52,280 The paper also discusses some non-digital programming 217 00:06:52,280 --> 00:06:54,880 methods that are really interesting, especially 218 00:06:54,880 --> 00:06:58,400 for the types of hardware often used in neuromorphic systems. 219 00:06:58,400 --> 00:06:58,960 OK. 220 00:06:58,960 --> 00:07:01,120 These methods borrow ideas from fields 221 00:07:01,120 --> 00:07:03,920 like control engineering and signal processing. 222 00:07:03,920 --> 00:07:06,280 So, we're looking beyond the traditional world of computer 223 00:07:06,280 --> 00:07:09,640 science to tackle the challenges of neuromorphic programming. 224 00:07:09,640 --> 00:07:12,000 It's a truly interdisciplinary field. 225 00:07:12,000 --> 00:07:14,880 It requires drawing on a wide range of expertise. 226 00:07:14,880 --> 00:07:16,360 This is all so fascinating. 227 00:07:16,360 --> 00:07:18,120 It feels like we're standing on the precipice 228 00:07:18,120 --> 00:07:19,800 of a whole new era of computing. 229 00:07:19,800 --> 00:07:20,920 It does, doesn't it? 230 00:07:20,920 --> 00:07:23,120 But we're also just beginning to understand 231 00:07:23,120 --> 00:07:25,880 how to program these incredible machines. 232 00:07:25,880 --> 00:07:27,000 That's the beauty of it. 233 00:07:27,000 --> 00:07:28,880 There's so much left to explore. 234 00:07:28,880 --> 00:07:29,480 There is. 235 00:07:29,480 --> 00:07:31,400 In the next part of our deep dive, 236 00:07:31,400 --> 00:07:33,480 we'll delve into some specific approaches 237 00:07:33,480 --> 00:07:36,720 to neuromorphic programming that the paper highlights 238 00:07:36,720 --> 00:07:38,320 as particularly promising. 239 00:07:38,320 --> 00:07:39,280 I can't wait. 240 00:07:39,280 --> 00:07:42,520 Welcome back to our deep dive into neuromorphic programming. 241 00:07:42,520 --> 00:07:44,320 I'm ready to get into the nitty gritty. 242 00:07:44,320 --> 00:07:46,920 Last time we explored why programming, 243 00:07:46,920 --> 00:07:49,880 these brain-inspired computers, is so different. 244 00:07:49,880 --> 00:07:50,160 Right. 245 00:07:50,160 --> 00:07:51,280 It's a whole new world. 246 00:07:51,280 --> 00:07:53,600 Now, let's look at some specific approaches. 247 00:07:53,600 --> 00:07:53,840 OK. 248 00:07:53,840 --> 00:07:55,120 Let's get concrete. 249 00:07:55,120 --> 00:07:57,880 Researchers are exploring these to tackle this challenge. 250 00:07:57,880 --> 00:07:58,920 Oh, where do we start? 251 00:07:58,920 --> 00:08:01,440 Let's begin with neuromorphic co-design. 252 00:08:01,440 --> 00:08:02,280 Co-design. 253 00:08:02,280 --> 00:08:05,480 It's not strictly a programming paradigm, 254 00:08:05,480 --> 00:08:09,680 but rather a fundamental approach to designing these systems. 255 00:08:09,680 --> 00:08:11,600 So what exactly does that involve? 256 00:08:11,600 --> 00:08:13,960 It's all about closely integrating 257 00:08:13,960 --> 00:08:16,520 the development of both the hardware and the software. 258 00:08:16,520 --> 00:08:17,160 Together. 259 00:08:17,160 --> 00:08:19,200 Ensuring they work together seamlessly 260 00:08:19,200 --> 00:08:20,800 for a specific application. 261 00:08:20,800 --> 00:08:21,240 OK. 262 00:08:21,240 --> 00:08:24,120 Remember these brain-inspired computers? 263 00:08:24,120 --> 00:08:27,120 They use their physical properties for computation? 264 00:08:27,120 --> 00:08:30,120 So the hardware design is crucial to how we program them. 265 00:08:30,120 --> 00:08:30,880 That's right. 266 00:08:30,880 --> 00:08:32,560 It's like an architect designing a building. 267 00:08:32,560 --> 00:08:33,360 Yeah, like an architect. 268 00:08:33,360 --> 00:08:35,320 I can't just draw up a fancy blueprint. 269 00:08:35,320 --> 00:08:35,720 Right. 270 00:08:35,720 --> 00:08:39,200 They need to consider the materials, the location. 271 00:08:39,200 --> 00:08:40,760 And how everything will fit together. 272 00:08:40,760 --> 00:08:41,680 That's a great analogy. 273 00:08:41,680 --> 00:08:44,920 So co-design seems perfect for specialized applications. 274 00:08:44,920 --> 00:08:47,080 It has been very successful in developing 275 00:08:47,080 --> 00:08:49,240 specialized neuromorphic chips. 276 00:08:49,240 --> 00:08:52,760 Chips that excel at tasks like recognizing patterns? 277 00:08:52,760 --> 00:08:53,400 Yes. 278 00:08:53,400 --> 00:08:55,040 Or processing sensory data? 279 00:08:55,040 --> 00:08:56,520 Almost like artificial senses. 280 00:08:56,520 --> 00:08:57,080 Exactly. 281 00:08:57,080 --> 00:09:01,000 OK, so co-design seems perfect for specialized applications. 282 00:09:01,000 --> 00:09:03,920 But what about creating software that can adapt? 283 00:09:03,920 --> 00:09:05,480 It's a different neuromorphic system. 284 00:09:05,480 --> 00:09:08,840 To tackle a wider range of problems. 285 00:09:08,840 --> 00:09:10,760 That's where machine learning methods come in. 286 00:09:10,760 --> 00:09:11,680 Machine learning. 287 00:09:11,680 --> 00:09:13,400 Remember how we talk about machine learning? 288 00:09:13,400 --> 00:09:15,840 Yeah, letting the computer learn from data. 289 00:09:15,840 --> 00:09:18,240 Instead of giving it explicit instruction. 290 00:09:18,240 --> 00:09:20,040 It's like teaching a computer to ride a bike 291 00:09:20,040 --> 00:09:21,320 by letting it practice. 292 00:09:21,320 --> 00:09:21,920 Exactly. 293 00:09:21,920 --> 00:09:24,080 Instead of giving it a physics lecture. 294 00:09:24,080 --> 00:09:26,600 This approach is becoming increasingly popular 295 00:09:26,600 --> 00:09:28,040 in neuromorphic computing. 296 00:09:28,040 --> 00:09:28,960 Especially with what? 297 00:09:28,960 --> 00:09:33,400 Especially with the rise of spiking neural networks or SNNs. 298 00:09:33,400 --> 00:09:34,080 SNNs. 299 00:09:34,080 --> 00:09:34,960 Right. 300 00:09:34,960 --> 00:09:36,200 We touched on those before. 301 00:09:36,200 --> 00:09:38,520 Yes, they are those brain-like networks. 302 00:09:38,520 --> 00:09:40,000 They're the ones that communicate using 303 00:09:40,000 --> 00:09:41,640 spikes of electrical activity. 304 00:09:41,640 --> 00:09:44,120 Mimicking how real neurons work. 305 00:09:44,120 --> 00:09:44,640 OK. 306 00:09:44,640 --> 00:09:48,400 SNNs have the potential to be much more energy efficient. 307 00:09:48,400 --> 00:09:49,520 More efficient than what? 308 00:09:49,520 --> 00:09:51,400 Then the artificial neural networks 309 00:09:51,400 --> 00:09:53,720 used in traditional deep learning. 310 00:09:53,720 --> 00:09:54,360 Interesting. 311 00:09:54,360 --> 00:09:56,400 And they're much better at processing information 312 00:09:56,400 --> 00:09:57,960 that changes over time. 313 00:09:57,960 --> 00:10:01,280 So SNNs are like the next generation of neural networks. 314 00:10:01,280 --> 00:10:03,960 Taking inspiration directly from the brain's own communication 315 00:10:03,960 --> 00:10:04,440 system. 316 00:10:04,440 --> 00:10:05,440 OK, I see. 317 00:10:05,440 --> 00:10:07,720 Researchers are exploring different machine learning 318 00:10:07,720 --> 00:10:10,400 methods that work well with these spiking networks. 319 00:10:10,400 --> 00:10:13,400 Going beyond just adapting traditional deep learning 320 00:10:13,400 --> 00:10:14,240 techniques. 321 00:10:14,240 --> 00:10:15,240 Precisely. 322 00:10:15,240 --> 00:10:18,760 So we can't just copy and paste from the traditional world 323 00:10:18,760 --> 00:10:19,560 of AI. 324 00:10:19,560 --> 00:10:21,520 We need to develop methods tailored 325 00:10:21,520 --> 00:10:23,280 to these unique SNNs. 326 00:10:23,280 --> 00:10:24,000 I don't. 327 00:10:24,000 --> 00:10:28,720 And one particularly fascinating area is online learning. 328 00:10:28,720 --> 00:10:29,400 Online learning. 329 00:10:29,400 --> 00:10:30,200 What does that mean? 330 00:10:30,200 --> 00:10:33,280 It's all about allowing brain inspired computers 331 00:10:33,280 --> 00:10:35,680 to learn and adapt continuously. 332 00:10:35,680 --> 00:10:36,440 Continuous. 333 00:10:36,440 --> 00:10:38,240 As they interact with the world. 334 00:10:38,240 --> 00:10:41,720 So no more training offline with massive data sets. 335 00:10:41,720 --> 00:10:42,360 Exactly. 336 00:10:42,360 --> 00:10:45,360 So instead of teaching a network, everything up front. 337 00:10:45,360 --> 00:10:46,880 We let it learn on the fly. 338 00:10:46,880 --> 00:10:49,800 Kind of like how our own brains constantly adapt 339 00:10:49,800 --> 00:10:50,920 to new experiences. 340 00:10:50,920 --> 00:10:51,640 Exactly. 341 00:10:51,640 --> 00:10:53,840 And one promising way to implement this. 342 00:10:53,840 --> 00:10:55,000 Just through plasticity. 343 00:10:55,000 --> 00:10:55,800 Plasticity, right? 344 00:10:55,800 --> 00:10:58,080 Remember how we talked about brain inspired computers 345 00:10:58,080 --> 00:11:00,000 physically changing during operation? 346 00:11:00,000 --> 00:11:01,320 Yeah, they're rewiring themselves. 347 00:11:01,320 --> 00:11:03,480 Well, plasticity refers to their ability 348 00:11:03,480 --> 00:11:06,160 to modify their own structure and connection. 349 00:11:06,160 --> 00:11:06,880 Their connections. 350 00:11:06,880 --> 00:11:08,840 In response to new information. 351 00:11:08,840 --> 00:11:11,880 Almost like rewiring their circuits as they learn. 352 00:11:11,880 --> 00:11:12,440 Exactly. 353 00:11:12,440 --> 00:11:13,960 So we're not just training the network. 354 00:11:13,960 --> 00:11:16,760 We're letting it re-architect itself as it learns. 355 00:11:16,760 --> 00:11:17,720 That's mind blowing. 356 00:11:17,720 --> 00:11:21,160 The challenge, of course, is guiding that rewiring process. 357 00:11:21,160 --> 00:11:21,640 Oh, way back. 358 00:11:21,640 --> 00:11:24,360 To make sure the system learns what we want it to learn. 359 00:11:24,360 --> 00:11:25,000 Right. 360 00:11:25,000 --> 00:11:28,240 Researchers are using techniques like evolutionary algorithm. 361 00:11:28,240 --> 00:11:29,400 Evolutionary algorithm. 362 00:11:29,400 --> 00:11:30,360 And meta learning. 363 00:11:30,360 --> 00:11:30,840 OK. 364 00:11:30,840 --> 00:11:33,720 To figure out the best plasticity rules. 365 00:11:33,720 --> 00:11:34,480 Plasticity rules. 366 00:11:34,480 --> 00:11:36,880 That govern how the system modifies itself. 367 00:11:36,880 --> 00:11:41,360 OK, so online learning through plasticity is one path. 368 00:11:41,360 --> 00:11:43,360 What other approaches are out there? 369 00:11:43,360 --> 00:11:46,960 Another fascinating avenue is the use of evolutionary methods. 370 00:11:46,960 --> 00:11:48,040 Evolutionary methods. 371 00:11:48,040 --> 00:11:49,720 You know how natural selection has 372 00:11:49,720 --> 00:11:52,280 shaped the incredible diversity and complexity of life 373 00:11:52,280 --> 00:11:52,800 on Earth? 374 00:11:52,800 --> 00:11:53,040 Right. 375 00:11:53,040 --> 00:11:54,360 Survival of the fittest and all that. 376 00:11:54,360 --> 00:11:55,080 Exactly. 377 00:11:55,080 --> 00:11:57,280 So instead of designing a solution ourselves. 378 00:11:57,280 --> 00:11:59,920 We're setting up a sort of digital ecosystem 379 00:11:59,920 --> 00:12:03,520 where solutions can evolve organically. 380 00:12:03,520 --> 00:12:06,240 So a survival of the fittest but for neural networks. 381 00:12:06,240 --> 00:12:07,560 That's a great way to put it. 382 00:12:07,560 --> 00:12:09,120 And over many generations. 383 00:12:09,120 --> 00:12:12,400 The algorithm discovers incredibly efficient and effective 384 00:12:12,400 --> 00:12:13,760 network architectures. 385 00:12:13,760 --> 00:12:14,400 Exactly. 386 00:12:14,400 --> 00:12:16,440 And this approach doesn't require the network 387 00:12:16,440 --> 00:12:17,640 to be differentiable. 388 00:12:17,640 --> 00:12:21,600 Meaning it can be applied to a broader range of SNN models. 389 00:12:21,600 --> 00:12:22,080 OK. 390 00:12:22,080 --> 00:12:24,000 Including some that are difficult to train 391 00:12:24,000 --> 00:12:26,560 with traditional deep learning techniques. 392 00:12:26,560 --> 00:12:28,280 This is all incredibly fascinating. 393 00:12:28,280 --> 00:12:30,360 But I have to admit it's a lot to keep up with. 394 00:12:30,360 --> 00:12:30,920 I know. 395 00:12:30,920 --> 00:12:32,320 There are so many approaches. 396 00:12:32,320 --> 00:12:35,240 Podesign, machine learning, online learning. 397 00:12:35,240 --> 00:12:37,360 Plasticity, evolutionary methods. 398 00:12:37,360 --> 00:12:39,560 It feels like we need a whole new dictionary just to talk 399 00:12:39,560 --> 00:12:41,680 about neuromorphic programming. 400 00:12:41,680 --> 00:12:45,400 It's a rapidly evolving field with its own unique concepts 401 00:12:45,400 --> 00:12:46,360 and terminology. 402 00:12:46,360 --> 00:12:46,800 Right. 403 00:12:46,800 --> 00:12:49,200 But the key takeaway is that we're moving away 404 00:12:49,200 --> 00:12:51,560 from the old way of programming. 405 00:12:51,560 --> 00:12:53,600 The old command and control way. 406 00:12:53,600 --> 00:12:55,960 Where we dictate every step to the computer. 407 00:12:55,960 --> 00:12:57,440 Instead, we're embracing approaches 408 00:12:57,440 --> 00:12:59,400 that allow the computation to emerge. 409 00:12:59,400 --> 00:13:00,920 From the dynamic interactions. 410 00:13:00,920 --> 00:13:04,160 From the interactions within these brain inspired systems. 411 00:13:04,160 --> 00:13:07,080 It's a paradigm shift in how we think about computation 412 00:13:07,080 --> 00:13:07,600 itself. 413 00:13:07,600 --> 00:13:08,560 It really is. 414 00:13:08,560 --> 00:13:12,320 And this shift is leading to some truly ingenious approaches 415 00:13:12,320 --> 00:13:13,120 to programming. 416 00:13:13,120 --> 00:13:14,000 Let's let's next. 417 00:13:14,000 --> 00:13:15,840 In the next part of our deep dive, 418 00:13:15,840 --> 00:13:18,120 we'll explore even more of these approaches. 419 00:13:18,120 --> 00:13:18,560 Really? 420 00:13:18,560 --> 00:13:22,040 Something called spiking neuromorphic algorithms. 421 00:13:22,040 --> 00:13:22,400 OK. 422 00:13:22,400 --> 00:13:24,680 And the concept of higher abstractions. 423 00:13:24,680 --> 00:13:26,160 Stay tuned, listeners. 424 00:13:26,160 --> 00:13:28,680 Because we're just scratching the surface 425 00:13:28,680 --> 00:13:32,080 of this exciting journey into the future of computing. 426 00:13:32,080 --> 00:13:35,040 Welcome back, listeners, for the final part of our deep dive 427 00:13:35,040 --> 00:13:38,440 into the fascinating world of neuromorphic programming. 428 00:13:38,440 --> 00:13:41,400 We've explored why programming these brain inspired 429 00:13:41,400 --> 00:13:43,640 computers is such a unique challenge. 430 00:13:43,640 --> 00:13:45,120 And we've looked at some high level 431 00:13:45,120 --> 00:13:46,360 approaches for tackling it. 432 00:13:46,360 --> 00:13:46,600 Right. 433 00:13:46,600 --> 00:13:49,520 Now let's get even more concrete and see how some of these ideas 434 00:13:49,520 --> 00:13:50,680 are being put into practice. 435 00:13:50,680 --> 00:13:51,080 OK. 436 00:13:51,080 --> 00:13:51,560 Let's do it. 437 00:13:51,560 --> 00:13:54,640 We'll explore spiking neuromorphic algorithms. 438 00:13:54,640 --> 00:13:56,520 Spiking neuromorphic algorithms. 439 00:13:56,520 --> 00:13:58,760 And the concept of higher abstractions. 440 00:13:58,760 --> 00:13:59,920 Higher abstractions. 441 00:13:59,920 --> 00:14:02,120 Both of which are key to unlocking 442 00:14:02,120 --> 00:14:04,040 the true potential of these systems. 443 00:14:04,040 --> 00:14:04,280 OK. 444 00:14:04,280 --> 00:14:07,520 Let's start with these spiking neuromorphic algorithms, 445 00:14:07,520 --> 00:14:08,320 or SNAs. 446 00:14:08,320 --> 00:14:09,160 OK. 447 00:14:09,160 --> 00:14:10,480 What makes them so special? 448 00:14:10,480 --> 00:14:13,480 SNAs are algorithms specifically designed 449 00:14:13,480 --> 00:14:16,480 to work with the unique way spiking neural networks 450 00:14:16,480 --> 00:14:17,240 compute. 451 00:14:17,240 --> 00:14:19,360 So they're tailored for SNNs. 452 00:14:19,360 --> 00:14:19,800 Yes. 453 00:14:19,800 --> 00:14:23,360 Remember SNNs use spikes of electrical activity 454 00:14:23,360 --> 00:14:24,120 to communicate. 455 00:14:24,120 --> 00:14:25,880 Just like real neurons in the brain. 456 00:14:25,880 --> 00:14:28,560 Well, SNAs capitalize on these spikes 457 00:14:28,560 --> 00:14:31,560 to perform computations in a way that traditional algorithms 458 00:14:31,560 --> 00:14:32,760 simply can't. 459 00:14:32,760 --> 00:14:35,600 So SNAs are like a custom design toolkit 460 00:14:35,600 --> 00:14:37,800 for speaking the language of the brain. 461 00:14:37,800 --> 00:14:38,960 Yeah, you could say that. 462 00:14:38,960 --> 00:14:41,960 Taking advantage of those precise spikes of activity. 463 00:14:41,960 --> 00:14:42,480 Exactly. 464 00:14:42,480 --> 00:14:45,360 They leverage the timing and dynamics of these spikes. 465 00:14:45,360 --> 00:14:45,840 OK. 466 00:14:45,840 --> 00:14:49,080 To do things that are either very difficult or impossible 467 00:14:49,080 --> 00:14:50,560 for traditional algorithms. 468 00:14:50,560 --> 00:14:52,720 Can you give us some examples of what these SNAs can do? 469 00:14:52,720 --> 00:14:54,480 What kind of problems can they solve? 470 00:14:54,480 --> 00:14:57,240 The paper mentioned some really fascinating applications. 471 00:14:57,240 --> 00:14:59,400 For instance, researchers have developed SNAs 472 00:14:59,400 --> 00:15:02,600 that can perform calculations on sets of numbers. 473 00:15:02,600 --> 00:15:05,600 Nibbulate graphs, solve constraint satisfaction 474 00:15:05,600 --> 00:15:09,040 problems, and even model physical phenomena 475 00:15:09,040 --> 00:15:10,320 like the spread of heat. 476 00:15:10,320 --> 00:15:12,200 That's a pretty diverse set of applications. 477 00:15:12,200 --> 00:15:12,720 It is. 478 00:15:12,720 --> 00:15:15,600 And as we develop more sophisticated SNAs, 479 00:15:15,600 --> 00:15:18,160 we'll unlock even more of the computational power 480 00:15:18,160 --> 00:15:20,120 of these brain inspired systems. 481 00:15:20,120 --> 00:15:22,320 It sounds like SNAs have the potential 482 00:15:22,320 --> 00:15:24,960 to revolutionize many different fields. 483 00:15:24,960 --> 00:15:25,880 They absolutely do. 484 00:15:25,880 --> 00:15:26,380 OK. 485 00:15:26,380 --> 00:15:29,760 So SNAs sound incredibly powerful and versatile. 486 00:15:29,760 --> 00:15:30,160 Yeah. 487 00:15:30,160 --> 00:15:32,440 But I imagine they're also pretty complex 488 00:15:32,440 --> 00:15:33,920 to design and implement. 489 00:15:33,920 --> 00:15:35,200 Well, they can be. 490 00:15:35,200 --> 00:15:39,520 Is there any way to make programming with SNNs more 491 00:15:39,520 --> 00:15:40,280 accessible? 492 00:15:40,280 --> 00:15:41,120 More accessible. 493 00:15:41,120 --> 00:15:43,880 Especially for those of us who aren't neuroscientists 494 00:15:43,880 --> 00:15:46,080 or computer engineering wizards. 495 00:15:46,080 --> 00:15:47,120 That's a great question. 496 00:15:47,120 --> 00:15:49,840 And it brings us to the idea of higher abstractions. 497 00:15:49,840 --> 00:15:50,960 Higher abstractions. 498 00:15:50,960 --> 00:15:52,200 Think about it like this. 499 00:15:52,200 --> 00:15:55,160 When you write a program in a high level language, 500 00:15:55,160 --> 00:15:56,000 like Python. 501 00:15:56,000 --> 00:15:56,400 Python. 502 00:15:56,400 --> 00:15:56,760 OK. 503 00:15:56,760 --> 00:15:58,960 You don't have to worry about the low level details 504 00:15:58,960 --> 00:16:01,080 of how the computer's processor actually 505 00:16:01,080 --> 00:16:02,560 executes each instruction. 506 00:16:02,560 --> 00:16:03,200 Right. 507 00:16:03,200 --> 00:16:06,400 The language provides you with a higher level of abstraction. 508 00:16:06,400 --> 00:16:07,760 A higher level of abstraction. 509 00:16:07,760 --> 00:16:10,120 Allowing you to focus on the logic of your program. 510 00:16:10,120 --> 00:16:10,640 All right. 511 00:16:10,640 --> 00:16:12,760 Rather than the nitty gritty of machine code. 512 00:16:12,760 --> 00:16:14,840 It's like using a map to navigate a city. 513 00:16:14,840 --> 00:16:15,680 Yes. 514 00:16:15,680 --> 00:16:18,920 Instead of memorizing every single street in Alleyway. 515 00:16:18,920 --> 00:16:21,160 And in the world of neuromorphic programming. 516 00:16:21,160 --> 00:16:21,440 Oh, OK. 517 00:16:21,440 --> 00:16:23,960 Researchers are developing tools and frameworks 518 00:16:23,960 --> 00:16:26,440 that provide similar high level abstractions 519 00:16:26,440 --> 00:16:28,160 for working with SNNs. 520 00:16:28,160 --> 00:16:30,280 So these tools would allow us to program 521 00:16:30,280 --> 00:16:33,160 brain inspired computers without needing 522 00:16:33,160 --> 00:16:36,360 to understand all the intricate details of how 523 00:16:36,360 --> 00:16:38,360 neurons communicate and interact. 524 00:16:38,360 --> 00:16:39,360 That's the goal. 525 00:16:39,360 --> 00:16:40,680 That would be a game changer. 526 00:16:40,680 --> 00:16:44,080 These higher abstractions aim to make neuromorphic programming 527 00:16:44,080 --> 00:16:45,800 more user friendly. 528 00:16:45,800 --> 00:16:46,680 User friendly. 529 00:16:46,680 --> 00:16:47,840 And accessible. 530 00:16:47,840 --> 00:16:50,800 Opening it up to a wider range of developers. 531 00:16:50,800 --> 00:16:53,320 The paper mentions a couple of specific examples 532 00:16:53,320 --> 00:16:54,440 of these high level tools. 533 00:16:54,440 --> 00:16:55,560 Can you tell us about them? 534 00:16:55,560 --> 00:16:58,680 Of course, one example is the neural engineering framework 535 00:16:58,680 --> 00:16:59,800 or Nengo. 536 00:16:59,800 --> 00:17:00,600 Nengo. 537 00:17:00,600 --> 00:17:03,480 Nengo allows programmers to describe the computations 538 00:17:03,480 --> 00:17:05,120 they want to perform at a higher level. 539 00:17:05,120 --> 00:17:05,880 A higher level. 540 00:17:05,880 --> 00:17:08,600 Using concepts from dynamical systems theory. 541 00:17:08,600 --> 00:17:08,960 OK. 542 00:17:08,960 --> 00:17:11,600 Then it automatically translates those descriptions 543 00:17:11,600 --> 00:17:13,480 into networks of spiking neurons. 544 00:17:13,480 --> 00:17:16,720 So you don't need to be a neuroscience expert to use Nengo? 545 00:17:16,720 --> 00:17:17,600 Not necessarily. 546 00:17:17,600 --> 00:17:19,600 You can describe what you want to achieve 547 00:17:19,600 --> 00:17:21,040 in a more abstract way. 548 00:17:21,040 --> 00:17:21,880 Right. 549 00:17:21,880 --> 00:17:25,120 And Nengo takes care of all the messy details of implementing 550 00:17:25,120 --> 00:17:26,520 it with those spiking neurons. 551 00:17:26,520 --> 00:17:27,560 Exactly. 552 00:17:27,560 --> 00:17:28,240 That's incredible. 553 00:17:28,240 --> 00:17:31,480 It makes designing complex brain inspired systems much more 554 00:17:31,480 --> 00:17:33,360 accessible and intuitive. 555 00:17:33,360 --> 00:17:35,200 What's the other example of the paper highlights? 556 00:17:35,200 --> 00:17:38,440 The other example is Lava, an open source framework designed 557 00:17:38,440 --> 00:17:41,240 specifically for neuromorphic computing. 558 00:17:41,240 --> 00:17:45,360 So Lava is like a toolbox full of pre-built components 559 00:17:45,360 --> 00:17:48,200 and utilities for working with brain inspired computers. 560 00:17:48,200 --> 00:17:49,200 That's a great analogy. 561 00:17:49,200 --> 00:17:50,920 Does it have those SNAs we were talking about? 562 00:17:50,920 --> 00:17:51,480 It does. 563 00:17:51,480 --> 00:17:55,120 Lava includes a growing library of neuromorphic algorithms 564 00:17:55,120 --> 00:17:55,840 and tools. 565 00:17:55,840 --> 00:17:58,040 Including support for those SNAs we talked about. 566 00:17:58,040 --> 00:18:00,400 It also provides higher level abstractions 567 00:18:00,400 --> 00:18:02,080 for describing computations. 568 00:18:02,080 --> 00:18:02,560 OK. 569 00:18:02,560 --> 00:18:05,160 And managing the complexities of SNNs. 570 00:18:05,160 --> 00:18:07,960 It sounds like Lava is trying to build an entire ecosystem 571 00:18:07,960 --> 00:18:09,840 for neuromorphic programming. 572 00:18:09,840 --> 00:18:10,960 Kind of is. 573 00:18:10,960 --> 00:18:12,440 Giving developers everything they 574 00:18:12,440 --> 00:18:14,560 need to build sophisticated applications 575 00:18:14,560 --> 00:18:16,400 for these new types of computers. 576 00:18:16,400 --> 00:18:18,480 And the best part is that it's open source. 577 00:18:18,480 --> 00:18:19,960 Open source. 578 00:18:19,960 --> 00:18:21,440 So anyone can contribute. 579 00:18:21,440 --> 00:18:23,400 So anyone can contribute to its development 580 00:18:23,400 --> 00:18:25,720 and help push the field forward. 581 00:18:25,720 --> 00:18:28,200 This is all so incredibly exciting. 582 00:18:28,200 --> 00:18:28,760 It is. 583 00:18:28,760 --> 00:18:31,520 It feels like we're truly on the verge of a revolution 584 00:18:31,520 --> 00:18:32,280 in computing. 585 00:18:32,280 --> 00:18:32,840 I agree. 586 00:18:32,840 --> 00:18:35,680 But as we've discussed throughout this deep dive, 587 00:18:35,680 --> 00:18:37,040 it's still early days. 588 00:18:37,040 --> 00:18:38,080 Very early. 589 00:18:38,080 --> 00:18:40,920 What are some of the key challenges that lie ahead? 590 00:18:40,920 --> 00:18:43,440 One of the biggest challenges is developing 591 00:18:43,440 --> 00:18:46,680 a unified theory of neuromorphic computation. 592 00:18:46,680 --> 00:18:48,040 A unified theory. 593 00:18:48,040 --> 00:18:49,920 A framework that can bridge the gap 594 00:18:49,920 --> 00:18:52,640 between the physical dynamics of these systems 595 00:18:52,640 --> 00:18:55,880 and the abstract concepts we use to describe computation. 596 00:18:55,880 --> 00:18:58,120 We need a way to connect the language of the brain 597 00:18:58,120 --> 00:18:59,760 with the language of computer science. 598 00:18:59,760 --> 00:19:00,280 Exactly. 599 00:19:00,280 --> 00:19:01,640 Without a common framework. 600 00:19:01,640 --> 00:19:02,840 It's hard to make progress. 601 00:19:02,840 --> 00:19:05,200 It's difficult to compare different approaches, 602 00:19:05,200 --> 00:19:06,280 share knowledge. 603 00:19:06,280 --> 00:19:06,600 Right. 604 00:19:06,600 --> 00:19:08,280 And build upon each other's work. 605 00:19:08,280 --> 00:19:09,800 It sounds like a daunting challenge, 606 00:19:09,800 --> 00:19:11,600 but also an incredibly exciting one. 607 00:19:11,600 --> 00:19:12,920 It absolutely is. 608 00:19:12,920 --> 00:19:14,680 There's so much we still don't understand 609 00:19:14,680 --> 00:19:16,400 about how the brain computes. 610 00:19:16,400 --> 00:19:19,200 And those mysteries hold the key to unlocking 611 00:19:19,200 --> 00:19:21,560 the true potential of neuromorphic systems. 612 00:19:21,560 --> 00:19:24,240 We're like explorers charting a new frontier. 613 00:19:24,240 --> 00:19:26,000 And every discovery brings us closer 614 00:19:26,000 --> 00:19:29,840 to understanding the very nature of intelligence itself. 615 00:19:29,840 --> 00:19:31,320 That's beautifully put. 616 00:19:31,320 --> 00:19:33,600 And it sounds like the journey is just beginning. 617 00:19:33,600 --> 00:19:34,560 It really is. 618 00:19:34,560 --> 00:19:39,120 As we delve deeper into this world of neuromorphic programming, 619 00:19:39,120 --> 00:19:43,000 we'll not only create more powerful and efficient computers, 620 00:19:43,000 --> 00:19:46,040 but we'll also gain a deeper understanding of our own minds. 621 00:19:46,040 --> 00:19:46,920 That's a great point. 622 00:19:46,920 --> 00:19:49,640 It's been a pleasure exploring this fascinating topic 623 00:19:49,640 --> 00:19:50,320 with you. 624 00:19:50,320 --> 00:19:51,200 Likewise. 625 00:19:51,200 --> 00:19:54,320 And to our listeners, I encourage you to keep your minds 626 00:19:54,320 --> 00:19:56,480 open and your curiosity peaked. 627 00:19:56,480 --> 00:19:57,720 Curiosity peaked. 628 00:19:57,720 --> 00:19:58,440 Who knows? 629 00:19:58,440 --> 00:20:01,080 Maybe you will be the one to discover the next breakthrough 630 00:20:01,080 --> 00:20:02,400 in neuromorphic computing. 631 00:20:02,400 --> 00:20:03,960 That's a great note to end on. 632 00:20:03,960 --> 00:20:04,440 Thanks. 633 00:20:04,440 --> 00:20:06,720 Thank you for joining us on this incredible journey 634 00:20:06,720 --> 00:20:09,480 into the future of computing.