1 00:00:00,000 --> 00:00:03,120 All right, ready to jump into some really fascinating AI research. 2 00:00:03,120 --> 00:00:03,800 Yeah, let's do it. 3 00:00:03,800 --> 00:00:09,640 Today we're looking at how AI can be used to understand driving scenes. 4 00:00:09,640 --> 00:00:10,920 Really understand, yeah. 5 00:00:10,960 --> 00:00:16,480 Not just, you know, recognizing objects, like moving beyond just seeing a stop 6 00:00:16,480 --> 00:00:18,720 sign and actually getting what it means. 7 00:00:18,760 --> 00:00:20,840 Yeah, yeah, it's purpose, the context. 8 00:00:20,840 --> 00:00:21,400 Exactly. 9 00:00:21,480 --> 00:00:25,640 And the paper that we're going to be talking about today is called Knowledge 10 00:00:25,640 --> 00:00:31,920 Graphs of Driving Scenes to empower the emerging capabilities of neurosymbolic AI. 11 00:00:32,080 --> 00:00:34,880 Quite a mouthful, but a really interesting paper. 12 00:00:35,000 --> 00:00:38,720 And it introduces this tool called DeCenica G. 13 00:00:38,840 --> 00:00:39,520 It's a cool tool. 14 00:00:39,600 --> 00:00:39,920 Yeah. 15 00:00:39,920 --> 00:00:41,120 Yeah, it's basically like a map. 16 00:00:41,160 --> 00:00:42,160 You can think of it that way. 17 00:00:42,160 --> 00:00:46,880 But instead of streets and buildings, it maps out objects, events and the 18 00:00:46,880 --> 00:00:49,360 relationships between them in driving scenarios. 19 00:00:49,480 --> 00:00:52,200 So like building a deeper understanding of what's happening on the road. 20 00:00:52,240 --> 00:00:53,520 Exactly, more connected. 21 00:00:53,520 --> 00:00:58,080 Okay, so how does it go from just data to this like insightful map? 22 00:00:58,120 --> 00:01:00,400 Well, it starts with massive data sets, right? 23 00:01:00,680 --> 00:01:03,320 The kind that are used to train self-driving cars. 24 00:01:03,360 --> 00:01:03,760 Okay. 25 00:01:03,920 --> 00:01:09,560 So images from cameras, LiDAR scans, GPS data, all that stuff. 26 00:01:09,680 --> 00:01:10,040 Okay. 27 00:01:10,200 --> 00:01:14,040 And it takes all that raw data and turns it into a knowledge graph. 28 00:01:14,200 --> 00:01:15,920 And what is a knowledge graph exactly? 29 00:01:15,920 --> 00:01:18,040 I'm picturing like lots of lines and bubbles. 30 00:01:18,040 --> 00:01:18,520 Is that right? 31 00:01:18,520 --> 00:01:20,640 You know, that's not a bad visualization actually. 32 00:01:20,640 --> 00:01:24,960 Basically, a knowledge graph connects different pieces of information and shows 33 00:01:24,960 --> 00:01:26,200 how they relate to each other. 34 00:01:26,560 --> 00:01:29,920 So yeah, you have your lines and your bubbles and they're all connected in a 35 00:01:29,920 --> 00:01:31,080 way that makes sense. 36 00:01:31,200 --> 00:01:35,680 So with TCNG, we're not just seeing a car and a pedestrian and a stop sign as 37 00:01:35,680 --> 00:01:37,120 like separate things. 38 00:01:37,120 --> 00:01:37,400 Right. 39 00:01:37,600 --> 00:01:40,320 But we're understanding how they all interact within the scene. 40 00:01:40,440 --> 00:01:41,040 Exactly. 41 00:01:41,040 --> 00:01:41,840 That's the point. 42 00:01:41,960 --> 00:01:42,240 Okay. 43 00:01:42,320 --> 00:01:45,720 It's like taking a puzzle and actually putting it together to see the full picture. 44 00:01:45,800 --> 00:01:48,600 And this is where the power of neural symbolic AI comes in, right? 45 00:01:48,600 --> 00:01:49,440 Exactly. 46 00:01:49,440 --> 00:01:54,160 That's what this paper is all about using TCNG to leverage this really interesting 47 00:01:54,160 --> 00:01:54,720 approach. 48 00:01:54,840 --> 00:01:55,120 Okay. 49 00:01:55,120 --> 00:01:58,080 Now, Neurosymbolic AI, I've heard that term. 50 00:01:58,400 --> 00:01:59,880 I'm not an AI expert though. 51 00:01:59,880 --> 00:02:01,120 So can you break that down for me? 52 00:02:01,400 --> 00:02:01,760 Sure. 53 00:02:01,760 --> 00:02:02,000 Yeah. 54 00:02:02,000 --> 00:02:06,240 It combines two different approaches to AI, you know, on the one hand, you have 55 00:02:06,240 --> 00:02:11,080 neural networks, which are great at recognizing patterns, but not as good at 56 00:02:11,080 --> 00:02:14,640 explaining how they do it, their logic. 57 00:02:15,160 --> 00:02:19,280 And on the other hand, you've got symbolic systems, which are very logical and 58 00:02:19,280 --> 00:02:24,160 good at reasoning, but they struggle with messy real world data. 59 00:02:24,320 --> 00:02:29,280 So it's like having like a really intuitive detective who can spot all the clues, 60 00:02:29,360 --> 00:02:29,720 Yeah. 61 00:02:29,720 --> 00:02:31,520 but can't really explain how they got there. 62 00:02:31,520 --> 00:02:32,400 Exactly. 63 00:02:32,400 --> 00:02:36,000 And then you have like a super logical analyst who needs everything to be 64 00:02:36,000 --> 00:02:38,080 perfect to solve the case. 65 00:02:38,080 --> 00:02:39,280 That's a great analogy. 66 00:02:39,440 --> 00:02:42,080 So Neurosymbolic AI brings these two together. 67 00:02:42,080 --> 00:02:42,400 Okay. 68 00:02:42,400 --> 00:02:47,520 So they can both learn from data and reason about it in a more, I guess, human 69 00:02:47,520 --> 00:02:48,160 like way. 70 00:02:48,160 --> 00:02:50,000 So how does that apply to DC and KG then? 71 00:02:50,000 --> 00:02:50,960 How are they used together? 72 00:02:51,280 --> 00:02:55,240 Well, DC and G with its knowledge graph, that provides the structure for 73 00:02:55,240 --> 00:02:57,520 understanding relationships and rules, right? 74 00:02:57,520 --> 00:02:58,000 Okay. 75 00:02:58,000 --> 00:03:02,080 And then you can use machine learning, things like neural networks, to analyze 76 00:03:02,080 --> 00:03:06,320 the patterns, make predictions within that structure, that framework. 77 00:03:06,800 --> 00:03:09,120 So DC and G provides like the blueprint. 78 00:03:09,200 --> 00:03:09,760 I like that. 79 00:03:09,760 --> 00:03:10,480 Yeah, the blueprint. 80 00:03:10,480 --> 00:03:12,480 And then the neural networks kind of fill in the detail. 81 00:03:12,480 --> 00:03:12,800 I was thinking. 82 00:03:12,800 --> 00:03:14,160 Make predictions based on that. 83 00:03:14,240 --> 00:03:15,360 Yeah, that's a good way to put it. 84 00:03:15,360 --> 00:03:19,840 And this leads to some really interesting, you know, developments in self-driving 85 00:03:19,840 --> 00:03:20,400 car technology. 86 00:03:20,400 --> 00:03:21,120 Yeah, absolutely. 87 00:03:21,120 --> 00:03:21,440 Yeah. 88 00:03:21,440 --> 00:03:27,040 So the paper highlights seven specific ways DC and G is being used to, you know, 89 00:03:27,040 --> 00:03:30,560 kind of improve AI's understanding of driving scenes. 90 00:03:30,560 --> 00:03:31,680 Let's start with the first three. 91 00:03:31,680 --> 00:03:32,720 What can you tell us about those? 92 00:03:33,200 --> 00:03:39,680 So the first three are really about enhancing machine perception, like how the AI sees 93 00:03:39,680 --> 00:03:41,120 and interprets the world. 94 00:03:41,120 --> 00:03:45,440 So giving self-driving cars a more human-like awareness? 95 00:03:45,440 --> 00:03:45,920 Exactly. 96 00:03:45,920 --> 00:03:47,040 That's a good way to put it. 97 00:03:47,040 --> 00:03:47,520 Okay. 98 00:03:47,520 --> 00:03:51,680 The first one is what they call knowledge-based entity prediction, or cake. 99 00:03:51,680 --> 00:03:52,240 Cakey. 100 00:03:52,240 --> 00:03:52,800 Okay. 101 00:03:52,800 --> 00:03:55,040 So imagine a car driving down the street. 102 00:03:55,040 --> 00:03:55,520 Okay. 103 00:03:55,520 --> 00:03:57,600 It sees a ball roll out into the road. 104 00:03:57,600 --> 00:04:00,880 Now, basic object detection would just be like, oh, there's a ball. 105 00:04:00,880 --> 00:04:04,640 But with Keepee, it uses the context from the knowledge graph. 106 00:04:04,640 --> 00:04:05,200 Okay. 107 00:04:05,200 --> 00:04:09,680 So knowing that kids often play with balls, it might predict that a kid is about to run 108 00:04:09,680 --> 00:04:11,120 into the street. 109 00:04:11,120 --> 00:04:14,720 So it's not just reacting to what it sees, it's like thinking ahead. 110 00:04:14,720 --> 00:04:15,360 Exactly. 111 00:04:15,360 --> 00:04:16,560 Anticipating. 112 00:04:16,560 --> 00:04:17,040 Wow. 113 00:04:17,040 --> 00:04:17,920 That's impressive. 114 00:04:17,920 --> 00:04:18,240 Yeah. 115 00:04:18,240 --> 00:04:19,600 Based on a deeper understanding. 116 00:04:19,600 --> 00:04:20,080 Okay. 117 00:04:20,080 --> 00:04:21,200 What about the second one? 118 00:04:21,200 --> 00:04:24,400 Second one is called explainable scene clustering typing. 119 00:04:24,400 --> 00:04:24,960 Okay. 120 00:04:24,960 --> 00:04:27,040 Think about all the different driving environments, right? 121 00:04:27,040 --> 00:04:29,760 You got your busy intersection, you got your quiet neighborhood. 122 00:04:29,760 --> 00:04:30,400 Exactly. 123 00:04:30,400 --> 00:04:31,920 School zone, highway at night. 124 00:04:31,920 --> 00:04:32,400 Yeah. 125 00:04:32,400 --> 00:04:34,960 They all require different driving behaviors, right? 126 00:04:34,960 --> 00:04:35,680 Mm-hmm. 127 00:04:35,680 --> 00:04:41,840 So DC and IG can actually group similar scenes together based on what's there. 128 00:04:41,840 --> 00:04:42,080 Yeah. 129 00:04:42,080 --> 00:04:44,160 The objects, the events, the relationships. 130 00:04:44,160 --> 00:04:47,920 So it's understanding not just what it's seeing, but what kind of environment it's in. 131 00:04:47,920 --> 00:04:48,720 Right, exactly. 132 00:04:48,720 --> 00:04:50,560 To inform the driving decisions. 133 00:04:50,560 --> 00:04:51,040 Okay. 134 00:04:51,040 --> 00:04:52,320 And it goes a step further. 135 00:04:52,320 --> 00:04:54,800 It gives labels to these groups. 136 00:04:54,800 --> 00:04:55,360 Oh, wow. 137 00:04:55,360 --> 00:05:00,960 So instead of just saying urban environment, it might be like school zone during pickup time. 138 00:05:00,960 --> 00:05:02,400 That's way more specific. 139 00:05:02,400 --> 00:05:04,160 Much more useful information. 140 00:05:04,160 --> 00:05:07,840 Yeah, it's like a driving instructor for the AI constantly analyzing. 141 00:05:07,840 --> 00:05:09,360 That's a great way to think about it. 142 00:05:09,360 --> 00:05:11,200 What about the third application? 143 00:05:11,200 --> 00:05:15,200 This one is called computing semantic similarity. 144 00:05:15,200 --> 00:05:15,520 Okay. 145 00:05:15,520 --> 00:05:20,000 And this is about understanding how similar two driving scenes are. 146 00:05:20,000 --> 00:05:21,360 Even if they look totally different. 147 00:05:21,360 --> 00:05:23,520 Yeah, even if they look different on the surface. 148 00:05:23,520 --> 00:05:25,600 Okay, so like what, give me an example. 149 00:05:25,600 --> 00:05:31,280 So imagine like a car turning left from around about in a busy city. 150 00:05:31,280 --> 00:05:31,840 Okay. 151 00:05:31,840 --> 00:05:37,840 And then another scene where a car is doing the same maneuver, but in a quiet suburb. 152 00:05:37,840 --> 00:05:38,320 Okay, yeah. 153 00:05:38,320 --> 00:05:38,960 Totally different. 154 00:05:38,960 --> 00:05:40,320 They look totally different. 155 00:05:40,320 --> 00:05:41,040 Fertilitially. 156 00:05:41,040 --> 00:05:42,720 But the core action is the same, right? 157 00:05:42,720 --> 00:05:43,120 Right. 158 00:05:43,120 --> 00:05:47,680 So DC and IG can actually see past those differences. 159 00:05:47,680 --> 00:05:50,160 So it's recognizing that underlying similarity. 160 00:05:50,160 --> 00:05:50,960 Exactly. 161 00:05:50,960 --> 00:05:53,520 And use the knowledge from one to apply to the other. 162 00:05:53,520 --> 00:05:57,520 So it can apply, yeah, can learn from different situations, even if they look different. 163 00:05:57,520 --> 00:05:58,080 Exactly. 164 00:05:58,080 --> 00:06:01,760 That's crucial for handling all sorts of situations, you know. 165 00:06:01,760 --> 00:06:04,480 These are the first three already show how powerful this is. 166 00:06:04,480 --> 00:06:06,480 What else can this knowledge graph do? 167 00:06:06,480 --> 00:06:07,760 Oh, there's so much more. 168 00:06:07,760 --> 00:06:08,480 Oh, there is? 169 00:06:08,480 --> 00:06:11,360 But I think we should probably let this sink in for a little bit. 170 00:06:11,360 --> 00:06:11,680 Okay. 171 00:06:11,680 --> 00:06:13,600 And we can explore the rest in part two. 172 00:06:13,600 --> 00:06:14,720 All right, looking forward to it. 173 00:06:14,720 --> 00:06:22,240 So we've seen how DC and IG helps self-driving cars like really see what's around them. 174 00:06:22,240 --> 00:06:22,720 Right. 175 00:06:22,720 --> 00:06:24,000 Become more aware. 176 00:06:24,000 --> 00:06:26,880 Yeah, but it goes beyond just perception. 177 00:06:26,880 --> 00:06:27,200 Okay. 178 00:06:27,200 --> 00:06:30,400 It can actually enhance the car's knowledge base. 179 00:06:30,400 --> 00:06:32,400 So you're talking about those next two applications you mentioned? 180 00:06:32,400 --> 00:06:34,640 Yeah, knowledge completion and augmentation. 181 00:06:34,640 --> 00:06:35,920 What does those actually mean? 182 00:06:35,920 --> 00:06:40,640 Okay, so think of it like this real world data is messy, right? 183 00:06:40,640 --> 00:06:42,800 Like sensors miss things. 184 00:06:42,800 --> 00:06:43,200 Yeah. 185 00:06:43,200 --> 00:06:47,120 Or there are relationships that aren't explicitly captured. 186 00:06:47,120 --> 00:06:47,520 Okay. 187 00:06:47,520 --> 00:06:50,560 So knowledge completion is like filling in those gaps. 188 00:06:50,560 --> 00:06:53,840 Using AI to make the knowledge graph more complete. 189 00:06:53,840 --> 00:06:58,080 So if a car sensor only sees part of a pedestrian, 190 00:06:58,080 --> 00:07:00,160 knowledge completion can help predict where they're going. 191 00:07:00,160 --> 00:07:03,680 Yeah, based on typical pedestrian behavior or traffic rules. 192 00:07:03,680 --> 00:07:05,840 So it's like having a detective on board. 193 00:07:05,840 --> 00:07:06,320 Exactly. 194 00:07:06,320 --> 00:07:07,360 She's seeing together the clues. 195 00:07:07,360 --> 00:07:08,800 Yeah, to get the full story. 196 00:07:08,800 --> 00:07:09,280 Okay. 197 00:07:09,280 --> 00:07:11,120 And what about knowledge augmentation? 198 00:07:11,120 --> 00:07:12,400 How is that different? 199 00:07:12,400 --> 00:07:15,600 Augmentation is about adding new information to the graph. 200 00:07:15,600 --> 00:07:15,920 Okay. 201 00:07:15,920 --> 00:07:20,960 So this could be real-time stuff like weather reports, traffic updates, 202 00:07:20,960 --> 00:07:25,280 or even insights that are derived from the existing data. 203 00:07:25,280 --> 00:07:28,560 So constantly learning and expanding its understanding. 204 00:07:28,560 --> 00:07:30,240 Exactly, making it more adaptable. 205 00:07:30,240 --> 00:07:31,280 And I want to number five. 206 00:07:32,080 --> 00:07:33,200 Semantic search. 207 00:07:33,200 --> 00:07:34,560 Semantic search. 208 00:07:34,560 --> 00:07:35,760 What is that all about? 209 00:07:35,760 --> 00:07:42,560 So imagine you have this huge library of driving scenarios all stored in the 210 00:07:42,560 --> 00:07:44,400 decen-e-g knowledge graph. 211 00:07:44,400 --> 00:07:44,720 Okay. 212 00:07:44,720 --> 00:07:48,240 Semantic search lets you search through it using natural language. 213 00:07:48,240 --> 00:07:52,640 So I could ask something like, find all the scenes where a car had to break hard to avoid 214 00:07:52,640 --> 00:07:53,440 a pedestrian. 215 00:07:53,440 --> 00:07:54,000 Exactly. 216 00:07:54,000 --> 00:07:55,840 And it would understand the meaning. 217 00:07:55,840 --> 00:07:56,720 Not just keywords. 218 00:07:56,720 --> 00:07:58,080 Right, the context of your query. 219 00:07:58,080 --> 00:07:58,320 Wow. 220 00:07:58,320 --> 00:08:01,360 So we would find all those scenes even if they don't have those exact words. 221 00:08:01,360 --> 00:08:01,680 Yeah. 222 00:08:01,680 --> 00:08:03,920 It's like having a librarian who knows what you mean. 223 00:08:03,920 --> 00:08:04,560 Okay. 224 00:08:04,560 --> 00:08:05,600 Very cool. 225 00:08:05,600 --> 00:08:07,280 And the last two applications. 226 00:08:07,280 --> 00:08:08,240 These get a little complex. 227 00:08:08,240 --> 00:08:09,520 Yeah, they sound pretty intense. 228 00:08:09,520 --> 00:08:12,800 So the sixth one is about causality. 229 00:08:12,800 --> 00:08:13,440 Causality. 230 00:08:14,320 --> 00:08:15,680 Understanding cause and effect. 231 00:08:15,680 --> 00:08:16,160 Yeah. 232 00:08:16,160 --> 00:08:17,600 And driving scenarios. 233 00:08:17,600 --> 00:08:19,920 Okay, but why is that important for self-driving cars? 234 00:08:20,800 --> 00:08:22,480 Seems kind of philosophical. 235 00:08:22,480 --> 00:08:26,160 It might seem that way, but it's crucial for safety. 236 00:08:26,160 --> 00:08:26,560 Okay. 237 00:08:26,560 --> 00:08:33,440 If we can teach AI to understand not just what happened, but why it happened, we can make them 238 00:08:33,440 --> 00:08:34,240 much safer. 239 00:08:34,240 --> 00:08:36,240 Okay, so if a car swerves. 240 00:08:36,240 --> 00:08:36,640 Right. 241 00:08:36,640 --> 00:08:38,880 Was it because another car cut it off? 242 00:08:38,880 --> 00:08:40,720 Or was there like a mechanical failure? 243 00:08:40,720 --> 00:08:41,360 Exactly. 244 00:08:41,360 --> 00:08:44,320 Understanding the cause helps prevent accidents. 245 00:08:44,320 --> 00:08:44,880 Makes sense. 246 00:08:44,880 --> 00:08:48,240 And that last application, cross-modal retrieval of complex data. 247 00:08:48,240 --> 00:08:49,280 Yeah, that was a mouthful. 248 00:08:49,280 --> 00:08:50,240 Something intimidating. 249 00:08:50,240 --> 00:08:51,840 It's actually simpler than it sounds. 250 00:08:52,320 --> 00:08:54,880 It's about finding information in one format. 251 00:08:54,880 --> 00:08:55,440 Okay. 252 00:08:55,440 --> 00:08:57,200 Based on a query in a different format. 253 00:08:57,200 --> 00:09:00,240 So like I could describe a scene using words. 254 00:09:00,240 --> 00:09:01,120 Yeah. 255 00:09:01,120 --> 00:09:04,400 And the system could find matching images or videos. 256 00:09:04,400 --> 00:09:06,480 Exactly, from the DeSana Geo-Knowledge Graph. 257 00:09:06,480 --> 00:09:09,840 So I could say, Shony all the times a car had to go through a roundabout in the rain. 258 00:09:09,840 --> 00:09:11,440 And it would find all the visuals. 259 00:09:11,440 --> 00:09:12,400 Wow, that's really cool. 260 00:09:12,400 --> 00:09:14,000 So it's like a translator for data. 261 00:09:14,000 --> 00:09:14,960 Exactly. 262 00:09:14,960 --> 00:09:17,360 Helps the AI access and analyze information. 263 00:09:17,360 --> 00:09:22,240 Now earlier you mentioned that DeSine KG uses data from a bunch of different driving data sets. 264 00:09:22,240 --> 00:09:22,800 Yeah. 265 00:09:22,800 --> 00:09:24,320 Can you talk about that a bit more? 266 00:09:24,320 --> 00:09:25,280 Why is that important? 267 00:09:25,280 --> 00:09:25,600 Sure. 268 00:09:25,600 --> 00:09:32,640 So they use data from data sets like Pandaset New Scenes, Waymo Open, Dataset, Kit, TTI. 269 00:09:32,640 --> 00:09:34,000 A lot of different ones. 270 00:09:34,000 --> 00:09:37,360 Yeah, these data sets capture lots of different scenarios. 271 00:09:37,360 --> 00:09:37,920 Okay. 272 00:09:37,920 --> 00:09:39,440 Like a different weather, different locations. 273 00:09:39,440 --> 00:09:41,440 It's about training on a variety of data. 274 00:09:41,440 --> 00:09:42,160 Exactly. 275 00:09:42,160 --> 00:09:44,880 Think of it like training an athlete. 276 00:09:44,880 --> 00:09:47,200 You know, if they only practice in one environment. 277 00:09:47,200 --> 00:09:48,720 They'll struggle when things change. 278 00:09:48,720 --> 00:09:49,200 Right. 279 00:09:49,200 --> 00:09:51,120 But if they train in lots of different places. 280 00:09:51,120 --> 00:09:52,320 They're more adaptable. 281 00:09:52,320 --> 00:09:53,840 Exactly, more resilient. 282 00:09:53,840 --> 00:09:54,720 That makes sense. 283 00:09:54,720 --> 00:09:58,240 So the AI isn't biased towards just certain types of scenarios. 284 00:09:58,240 --> 00:09:58,720 Right. 285 00:09:58,720 --> 00:10:00,160 Can handle the unexpected. 286 00:10:00,160 --> 00:10:03,440 All right, so we've seen all seven applications of DeSine KG. 287 00:10:03,440 --> 00:10:04,080 Yeah. 288 00:10:04,080 --> 00:10:06,640 And it's clear this resource is really promising. 289 00:10:06,640 --> 00:10:06,880 Yeah. 290 00:10:07,600 --> 00:10:11,600 But beyond all the technical stuff, I'm curious about the bigger picture. 291 00:10:11,600 --> 00:10:12,240 Okay. 292 00:10:12,240 --> 00:10:16,320 What does DeSine K tell us about the future of AI? 293 00:10:16,320 --> 00:10:20,720 So we spent this deep dive talking about DeSine K and how it can really transform 294 00:10:20,720 --> 00:10:23,360 the way AI understands these driving situations. 295 00:10:23,360 --> 00:10:24,400 Yeah, it's exciting stuff. 296 00:10:24,960 --> 00:10:27,120 But it's not just about the cool tech, right? 297 00:10:27,120 --> 00:10:28,160 No, definitely not. 298 00:10:28,160 --> 00:10:30,320 About the impact it can have in the real world. 299 00:10:30,320 --> 00:10:32,560 Exactly, real world impact. 300 00:10:32,560 --> 00:10:36,320 This research could actually make self-driving cars safer. 301 00:10:36,320 --> 00:10:37,760 Safer, more efficient. 302 00:10:37,760 --> 00:10:38,720 More reliable. 303 00:10:38,720 --> 00:10:39,840 And that's huge. 304 00:10:39,840 --> 00:10:40,640 It really is. 305 00:10:40,640 --> 00:10:43,120 It could change how we think about transportation. 306 00:10:43,680 --> 00:10:44,240 Yeah. 307 00:10:44,240 --> 00:10:50,080 I mean, when cars can navigate all those tricky situations with the same understanding as a human. 308 00:10:50,080 --> 00:10:52,240 It opens up so many possibilities. 309 00:10:52,240 --> 00:10:53,360 It really does. 310 00:10:53,360 --> 00:10:55,680 And it helps with that black box problem. 311 00:10:55,680 --> 00:10:56,800 It helps with the black box, yeah. 312 00:10:56,800 --> 00:11:00,240 That people have with AI, like not knowing how it makes decisions. 313 00:11:00,240 --> 00:11:01,040 Right, right. 314 00:11:01,040 --> 00:11:02,480 That lack of transparency. 315 00:11:02,480 --> 00:11:03,200 It can be scary. 316 00:11:03,200 --> 00:11:04,800 Especially with something like driving, yeah. 317 00:11:04,800 --> 00:11:06,240 You're handing over control. 318 00:11:06,240 --> 00:11:07,280 Exactly. 319 00:11:07,280 --> 00:11:10,400 But DeSine K, it lets us see the AI's reasoning. 320 00:11:10,400 --> 00:11:11,360 We can trace it. 321 00:11:11,360 --> 00:11:13,920 See what information led to a decision. 322 00:11:13,920 --> 00:11:16,160 So it's not just about getting the right answer. 323 00:11:16,160 --> 00:11:18,160 It's about the why. 324 00:11:18,160 --> 00:11:19,760 Yeah, understanding the process. 325 00:11:19,760 --> 00:11:21,040 The why behind the what. 326 00:11:21,040 --> 00:11:22,720 And that builds trust, right? 327 00:11:22,720 --> 00:11:23,040 Absolutely. 328 00:11:23,040 --> 00:11:25,040 Both for the people using the cars. 329 00:11:25,040 --> 00:11:26,240 And the people building them too. 330 00:11:26,240 --> 00:11:27,360 Yeah, the engineers. 331 00:11:27,360 --> 00:11:29,360 Yeah, that transparency is key. 332 00:11:29,360 --> 00:11:31,040 It also makes it easier to catch errors, right? 333 00:11:31,040 --> 00:11:33,200 Oh yeah, if you understand the reasoning. 334 00:11:33,200 --> 00:11:34,800 You can make sure it's sound. 335 00:11:34,800 --> 00:11:38,240 Fix any problems, make the systems more reliable. 336 00:11:38,240 --> 00:11:41,120 So it's not just about making AI smarter. 337 00:11:41,120 --> 00:11:44,240 It's about making it think more like us. 338 00:11:44,240 --> 00:11:45,680 More human-like, yeah. 339 00:11:45,680 --> 00:11:47,760 Understanding the world in a much deeper way. 340 00:11:47,760 --> 00:11:49,680 With nuance, with context. 341 00:11:49,680 --> 00:11:52,000 And that leads to a really interesting question. 342 00:11:52,000 --> 00:11:53,200 Okay, what's that? 343 00:11:53,200 --> 00:11:55,360 If AI can understand driving like this. 344 00:11:55,360 --> 00:11:57,280 Yeah, with this level of detail. 345 00:11:57,280 --> 00:11:58,800 What could it do? 346 00:11:58,800 --> 00:12:00,880 I mean, the possibilities are huge. 347 00:12:00,880 --> 00:12:02,800 What other fields could benefit? 348 00:12:02,800 --> 00:12:04,080 Think about healthcare. 349 00:12:04,080 --> 00:12:04,880 Healthcare, yeah. 350 00:12:04,880 --> 00:12:07,760 Analyzing patient data, finding patterns, 351 00:12:07,760 --> 00:12:09,040 personalized treatments. 352 00:12:09,040 --> 00:12:10,800 Finance, predicting markets. 353 00:12:10,800 --> 00:12:12,800 Exactly, or managing risk. 354 00:12:12,800 --> 00:12:15,360 Environment, understanding climate change. 355 00:12:15,360 --> 00:12:16,960 Oh, that's a big one. 356 00:12:16,960 --> 00:12:18,320 Modeling ecosystems. 357 00:12:18,320 --> 00:12:20,560 So many complex systems out there. 358 00:12:20,560 --> 00:12:22,320 It seems like we're just scratching the surface. 359 00:12:22,320 --> 00:12:23,120 I think so too. 360 00:12:23,120 --> 00:12:26,240 With DCNEG and this neuro-symbolic AI. 361 00:12:26,240 --> 00:12:29,120 It's a really exciting time to be in this field. 362 00:12:29,120 --> 00:12:29,840 I can imagine. 363 00:12:29,840 --> 00:12:31,600 Well, this has been an amazing deep dive. 364 00:12:31,600 --> 00:12:32,560 I've enjoyed it. 365 00:12:32,560 --> 00:12:34,880 Thank you so much for joining us and sharing your expertise. 366 00:12:34,880 --> 00:12:36,320 It was my pleasure, really. 367 00:12:36,320 --> 00:12:39,200 And to everyone listening, thanks for coming along on this journey. 368 00:12:39,200 --> 00:12:40,560 Yeah, thanks for listening. 369 00:12:40,560 --> 00:12:43,680 Into the world of knowledge graphs and self-driving cars. 370 00:12:43,680 --> 00:12:44,960 I hope you learned something. 371 00:12:44,960 --> 00:12:47,600 If you enjoyed it, be sure to subscribe and leave a review. 372 00:12:47,600 --> 00:12:49,280 Yeah, we'd love to hear from you. 373 00:12:49,280 --> 00:13:06,960 Until next time, keep exploring, keep learning, and keep diving deep.