1 00:00:00,05 --> 00:00:05,04 - [Instructor] We are back on the Azure AI demo screen. 2 00:00:05,04 --> 00:00:10,03 The ethics in the world of H-AI are constantly evolving. 3 00:00:10,03 --> 00:00:12,03 We touched on human-centered AI 4 00:00:12,03 --> 00:00:16,05 and the data sets we use to build our autonomous vehicles. 5 00:00:16,05 --> 00:00:18,06 Now let's take a look at what building 6 00:00:18,06 --> 00:00:22,02 with data privacy in mind might look like in H-AI. 7 00:00:22,02 --> 00:00:25,01 Instead of physically walking through an aisle 8 00:00:25,01 --> 00:00:28,05 and looking for shipments, which can be time consuming 9 00:00:28,05 --> 00:00:30,08 as well as dangerous, 10 00:00:30,08 --> 00:00:32,07 depending on the size of the products, 11 00:00:32,07 --> 00:00:36,08 warehouse robots are there to safely take our place. 12 00:00:36,08 --> 00:00:39,06 They can map every aisle and shelf, 13 00:00:39,06 --> 00:00:44,04 but they also capture people who are working and moving. 14 00:00:44,04 --> 00:00:46,06 This is the reason it's highly critical 15 00:00:46,06 --> 00:00:50,01 that data privacy is at the forefront of our thoughts 16 00:00:50,01 --> 00:00:52,04 when we are using H-AI. 17 00:00:52,04 --> 00:00:54,06 Let's look at an example. 18 00:00:54,06 --> 00:00:56,07 You've seen this picture before. 19 00:00:56,07 --> 00:00:58,08 It's a picture in a warehouse, 20 00:00:58,08 --> 00:01:02,02 and we are doing object detection using Azure AI, 21 00:01:02,02 --> 00:01:04,01 and it has put bounding boxes 22 00:01:04,01 --> 00:01:08,04 and identified four people in here. 23 00:01:08,04 --> 00:01:10,02 We discussed about degree of confidence. 24 00:01:10,02 --> 00:01:12,04 That's not what this demo is about. 25 00:01:12,04 --> 00:01:14,03 But look at this. 26 00:01:14,03 --> 00:01:19,07 We see Azure saying it's picked up four people, 27 00:01:19,07 --> 00:01:25,05 but, it's capturing people's faces, features. 28 00:01:25,05 --> 00:01:27,01 It shows what they're wearing. 29 00:01:27,01 --> 00:01:30,03 It shows headgear such as helmet. 30 00:01:30,03 --> 00:01:32,07 It is showing a robotic arm at the back. 31 00:01:32,07 --> 00:01:34,05 I'm not worried about the object here. 32 00:01:34,05 --> 00:01:38,02 I'm thinking about the people and there's lot of wires 33 00:01:38,02 --> 00:01:42,00 and cables and one person is at the laptop. 34 00:01:42,00 --> 00:01:43,02 There's expressions on each 35 00:01:43,02 --> 00:01:45,06 of these faces, which looks different. 36 00:01:45,06 --> 00:01:48,00 They all seem to be looking at the same thing. 37 00:01:48,00 --> 00:01:51,04 So there is so much information of people. 38 00:01:51,04 --> 00:01:54,06 So when an autonomous agentic AI 39 00:01:54,06 --> 00:02:00,07 or even a simple object detection model looks at a scenery 40 00:02:00,07 --> 00:02:04,02 or an environment, it sees a person. 41 00:02:04,02 --> 00:02:06,06 That's what it is saying it is seeing a person, 42 00:02:06,06 --> 00:02:08,08 but it is seeing more than just the person 43 00:02:08,08 --> 00:02:11,03 because it is able to see the expressions 44 00:02:11,03 --> 00:02:14,01 and their outfit and a lot more information. 45 00:02:14,01 --> 00:02:16,00 In some cases, we zoom in 46 00:02:16,00 --> 00:02:18,06 and we can even see the name tag of that person. 47 00:02:18,06 --> 00:02:22,05 So it is important for us to understand what we need 48 00:02:22,05 --> 00:02:26,01 to white out if we are giving a picture. 49 00:02:26,01 --> 00:02:28,09 In some data sets, when you go to download, 50 00:02:28,09 --> 00:02:32,00 especially from Europe where they follow GDPR very, 51 00:02:32,00 --> 00:02:34,08 very diligently, you can see the faces 52 00:02:34,08 --> 00:02:37,03 and expression of people are white out 53 00:02:37,03 --> 00:02:40,07 because we can do facial recognition 54 00:02:40,07 --> 00:02:42,03 and identify these people, 55 00:02:42,03 --> 00:02:45,02 and that is against GDPR compliance. 56 00:02:45,02 --> 00:02:48,08 So I want you to think about ethics as the right thing 57 00:02:48,08 --> 00:02:52,07 to do, but it's also about following the data privacy laws 58 00:02:52,07 --> 00:02:57,09 of the state or country and for GDPR compliance. 59 00:02:57,09 --> 00:03:00,09 So if this is a real person and you're saving their picture 60 00:03:00,09 --> 00:03:02,06 and this is going to the cloud 61 00:03:02,06 --> 00:03:04,00 and you're saying, "Hey, I'm just going 62 00:03:04,00 --> 00:03:09,07 to let the AI identify pictures of people in here", 63 00:03:09,07 --> 00:03:13,00 then you'll have to think about the GDPR compliance 64 00:03:13,00 --> 00:03:18,03 of permission of using their faces and storing them. 65 00:03:18,03 --> 00:03:20,02 And again, you have to decide are you using 66 00:03:20,02 --> 00:03:23,07 that in the cloud or in the Edge? 67 00:03:23,07 --> 00:03:24,08 If this is a face 68 00:03:24,08 --> 00:03:26,08 and a person's features 69 00:03:26,08 --> 00:03:30,02 that are stored right on the Edge 70 00:03:30,02 --> 00:03:33,06 and Edge AI is running inference and making decisions 71 00:03:33,06 --> 00:03:36,06 and moving on, or is this information stored? 72 00:03:36,06 --> 00:03:37,05 Think about this. 73 00:03:37,05 --> 00:03:40,03 If this image is just being stored as data 74 00:03:40,03 --> 00:03:44,04 and it is on the Edge and running inference 75 00:03:44,04 --> 00:03:47,04 and being able to make some predictions, 76 00:03:47,04 --> 00:03:51,02 you might think this is an H-AI problem in some workflow, 77 00:03:51,02 --> 00:03:55,01 but, if this data is stored with the timestamp, 78 00:03:55,01 --> 00:03:59,09 it is informing who was here, what were they looking at, 79 00:03:59,09 --> 00:04:01,04 whether it is... 80 00:04:01,04 --> 00:04:02,06 It could be a company offsite. 81 00:04:02,06 --> 00:04:03,09 It could be something else. 82 00:04:03,09 --> 00:04:06,04 So there's a lot more information that goes 83 00:04:06,04 --> 00:04:09,06 with the location and timestamp associated with the person 84 00:04:09,06 --> 00:04:13,04 that you are not even thinking as H-AI use case 85 00:04:13,04 --> 00:04:15,01 because that's not the use case. 86 00:04:15,01 --> 00:04:17,05 So those are the things you should be thinking about. 87 00:04:17,05 --> 00:04:19,02 What are you storing? 88 00:04:19,02 --> 00:04:21,00 What is the data privacy of it? 89 00:04:21,00 --> 00:04:25,01 And what other information is that data 90 00:04:25,01 --> 00:04:28,07 or that AI inference results carrying in it 91 00:04:28,07 --> 00:04:33,03 that you might want to be thinking about for data privacy 92 00:04:33,03 --> 00:04:35,06 and also for ethics compliance? 93 00:04:35,06 --> 00:04:38,01 We'll do a demo challenge next, 94 00:04:38,01 --> 00:04:41,01 and I want you to think about this 95 00:04:41,01 --> 00:04:45,09 and decide, okay, what kind of things are in this image 96 00:04:45,09 --> 00:04:46,08 and what am I omitting? 97 00:04:46,08 --> 00:04:47,08 What am I keeping? 98 00:04:47,08 --> 00:04:50,00 Let's do a fun challenge next.