[MUSIC] Now, let us talk about facial recognition, the application probably most familiar and known to all of you, and how deep learning the AI is affecting it. So what is modern facial recognition? Actually, it started quite early just like AI in the 60s. Woodrow Wilson Bledsoe actually had the first system that gave ideas to people and agencies who wanted to do facial recognition, right? That makes sense. So from then on in the 70s and the 90s, especially law enforcement agencies started to craft their own systems and started deploying them and using them. So it got easier and more widespread and relevant in the 2000s when we see that it can be used for big events such as Super Bowl to identify people in the crowd and identify them. And even local police department starts to have their own databases. So we can see it's quite prevalent at that time and especially coming to our era now. But what we forget is that there is an older and older facial recognition system mainly us. We always use this as a means of figuring out other facial features to understand who they are, where they're thinking or might be feeling. It also gives us this model then you can start to anticipate what they might do, identification and location and tracking is key and the basic things so that you might start to mirror or at least predict what others may do. Along with that you can start to see more details such as their emotions, their facial expressions, right? How people express just like the way I'm trying to emote right now there is this whole person kind of communication and that is not possible without visual system i.e. For us it's a lot of facial recognition in there. And through that we know that this is crucial for the social construct and how people interact and build connections or treat each other. So this whole natural human thing has been now automated. And as we have seen in this course from GANs topic that eventually everything we can emulate and automate the march of technology will be such that it will do that. And deep learning, basically, open the access wide open especially to everybody and using mobile devices and we can see that actually face ID is the standard features of iPhone and many other mobile devices. But that doesn't mean that the privacy has to be compromised, right? For example, Apple did make a conscious effort to use synthetic faces to train such a system. So there's always this dynamic between what can we use as a synthetic data or simulated data versus reality. But regardless, all the systems have to work on real people. The process as this chart is illustrating starts from an image or video any scene that you have. So the first part just like object detection would be to detect a face within that bigger scene, and then have to track it if there's any movement if it was a video, something like that. And then have to align itself in terms of the location the size and the pose. And then use these neural networks to do feature extraction so that it can match based on the database that it has stored of that particular person's face to the new input that's coming in, and then do a match whether it's the same person or not. So this diagram in the process should be very familiar to you at this point and it's nothing magical about it. We're just using all these components that's been built since the 60s to basically automate and lower the threshold that almost everybody is familiar with face ID. So there are actually four types of models that tends to be popular and it kind of maps the performance increase in leveling of such systems for you to be familiar with it. The first is DeepFace. Again, it's based on deep learning so it tends to have the word deep in it. So DeepFace is of course the first one that proved that neural networks and deep learning can achieve near human-level performance. And human performance is defined as 97.5% but turns out Deep ID and subsequent systems and innovations are doing better than even humans at 99.1%. So FaceNet by Google is further innovation that allows rapid calculation of similar things and matching distances so that inference of the output of models can be even faster and more accurate. Finally, VGGFace and you may remember VGG is a deep neural network that we're familiar with tuning even better and show how you can do that by collecting a very large training dataset. But any time you collect very large training datasets to train these deep CNN models there's always privacy concerns and that's exactly what we'll see next. And in fact the more accurate it gets, the most dangerous they get. So now there's actually a counter trend of not using them intentionally. Just very recently November 2021, Facebook shutdown use of facial recognition in their wider system but that's in contrast to let's say China where it's actually able to scan and identify 1.4 billion citizens in just one second, right? And that was quite a long time ago. So the systems are only getting better and deployed more at even schools and all these different points where they want to identify people. Even in US since 2018, FBI has had access to 412 million facial images database for their searches because governments, even local governments, progressive cities like San Francisco, Oakland, Boston, they're trying to actively ban the government and agencies from using them. And there's pressure on the tech giants like Amazon, Microsoft, IBM to not sell them to the law enforcement agencies. So there are all these trends going on whereas the technology is increasing people are starting to feel the threat because of the pervasiveness and easy access and high availability. So now we have to grapple with all these issues that we've been talking about in this course. What are the ethical and societal implications of just deploying these systems? Not just the tech accuracy itself. As we have covered. These are highly available and sophisticated systems that are being deployed that affects everybody in society. And in this report by Australian Government Human Rights Commission, they've identified three characteristics that all AI or technology systems should have. One fair, two accurate, and three accountable. Nobody should have any problems with such definitions. But actually, what is the meaning of fair, accurate, and accountable as we have covered in the ethnic section, and then who will be responsible for ensuring that such things are being met? No abuses being perpetrated? These are bigger concerns, right, and that involves everybody in society, especially you. So we want you to take a few moments, read the report, and reflect on the implications from such a report and a government body. [MUSIC]
