foreign AI stage my name is Gary Abella and I'm honored to have been asked to actually host the rest of the afternoon on this stage I have founded myself a billing based agency called codepan and I've spent the last 10 years really working through the trials and tribulations of taking AI research and actually pulling them into real world use cases that are actually fit for purpose and that's the entire theme of the applied AI stage and we have some of the most inspiring and thought-provoking speakers and leaders from across the AI ecosystem ranging from my fellow startup entrepreneurs to some of the movers and shakers of the corporate world presenting on this very stage here today so before I introduce our first Speaker I wanted to kick off the proceedings with a quote that I hope will really set the right mindset and who better to quote than Socrates himself who said that the only true wisdom is knowing that you know nothing and so the only thing that I ask of all of you
in here today is to have an open mind be excited to learn from each other today and to kick off proceedings I'm delighted to introduce our first inspirational speaker Mr Sean Moore having sold his uh first facial recognition startup true face a few years ago he now serves as the chief AI officer for his acquirer of panjam Sean has flown in directly from the U.S to really try to convince us skeptical Europeans on why he believes this highly controversial area of AI has a place in our society today so please welcome Sean to the sage good morning thank you all for being here um so quick agenda for the day an introduction to me and my background the landscape around facial recognition how bias can insert itself into AI specifically facial recognition the ethical implications of that and then we'll tie it up with a conclusion so as mentioned I'm Sean Moore I spent about 10 years building true face which is a leading Global facial recognition company one of the most important things to
us at the company was these two words Humanity first we're starting to see pushback around AI right now Elon Musk came out and said we need to Halt all AI the trouble with that is you have an international group of individuals who are all working separately to build these products so unless you have an international regulation you're not going to be able to do that so you know I think we have to ask the question how do we enforce or how do we ensure that at every single process of development and deployment we think about Humanity first this has to be a non-negotiable guiding decision that is made at every single point of that Journey from the investors who are investing to the researchers researching to the designers the product delivery the customer support Humanity has to come first so as mentioned I sold my company in 2021 to a group called Pangea pangium shares this principle with us this was the most important thing during the acquisition was they they believe that this technology
can be a global technology as long as Humanity comes first so a brief history of facial recognition many people do not know this started in 1960 with a man by the name of Woodrow Wilson Bledsoe it was through a joint effort with the intelligence agency in the U.S he was tasked with taking hand-written geometry of a face and trying to compare it and overlay it took about 30 years before a computer could do this and in 1991 we saw the first computer run a correct facial recognition analysis so about 2 000 pictures were outputted into the system they found the one individual the one input image they were looking for took another 10 years for us to see this in a commercial use case in the Super Bowl in the United States it's the first time in history that facial recognition have been used on people at this scale so 60 70 000 people that were going through effectively CCTV cameras scanning their face to figure out if they had warrants or if there was
a reason that they shouldn't be there or be investigated further this was ultimately shut down due to privacy concerns a few months later in 2010 uh this is when I I believe facial recognition was brought to the market uh in terms of mass adoption Facebook announced that they would be using facial recognition to auto tag your friends your family people in your pictures so not a consumer product but you're interfacing it sort of knowing sort of not knowing how it was happening and then in 2017 the iPhone x everyone thought this would crush facial recognition companies because now Apple can do it so how are you going to compete this was the turning point for adoption Apple put facial recognition in people's hands so it showed people the power that facial recognition has to unlock your phone to transact to prevent fraud and so this was truly a pivotal moment for all facial recognition providers and then a few years ago I showed the same history at the time there were not 1 billion cameras
in the world there are currently 1 billion cameras in the world so this gives you an idea of the scale in which computer vision or artificial intelligence can be deployed this morning it was announced that Princeton has come up with a camera that's the size of a grain of salt so incredibly impressive somewhat scary whoever gets their hands on it how are they using it it's being pitched as being used for medical care now that you can ingest this grain of salt and effectively scope your body but in the hands of an adversary how do they use this you know we're unsure and so the current facial recognition landscape you know we went from pen and paper to these massive computers down to a credit card size chip that sits in your pocket that can run facial recognition this is a Raspberry Pi the Raspberry Pi's are capable of storing between a hundred thousand and one million faces that run completely offline at the edge it's extremely impressive that we can now run these complex
computational algorithms on something of that size on 2015 it was when we introduced deep learning to facial recognition so prior to that it was clunky it was slow required a ton of compute power with the introduction of machine learning and deep learning really we could take these massive 50 60 70 million person data sets scanned through them pattern match and understand you know who was who in a very very short period of time one of the the things that we like to claim as a company which has been tested is we can scan through a billion faces a second so the population of the Earth in under 10 seconds to find a person this is what the power of deep learning has done to facial recognition so you've got improved accuracy you've got precision and recall and speed and now we have we have ways to test this so we have ways for third-party uh evaluators to to evaluate our claims to evaluate the accuracy nist in the US the National Institute of Standards and
technology is probably the biggest one they test over three I think 350 companies algorithms so you can see how you stack up the test ranges from a one to a 12 million identification of one to a 1.2 million identification and so you're seeing the different uh kind of different verticals in which this is being tested Visa border passports and so just different applications but it does it has proven all right we have been able to prove that even at the 1 to 12 million we score into 99 point x percent X percentage of accuracy it's increasingly accurate due to deep learning with all this said it's not without its challenges um you know people are very aware of the challenges and with bias and we'll get to that in a few minutes privacy and Trust if the public doesn't trust this technology apology you know there's going to be an uphill battle forever really and then data protection as you all are probably more acutely familiar than we are in the US you know data
protection is extremely important how the data transfers from the user uh the ability not to sell or the the requirement of the contract not to sell that data and then be able to remove yourself from those platforms and then the regulatory environment um you know it's incred increasingly difficult difficult for companies like ours to operate across the globe because of the regulatory environment is split by jurisdiction you know the EU has their requirements we have a per state requirement we don't have Federal Regulations so you've got 50 states to deal with and so it's just you know it's it's increasingly important that we can understand this environment in the regulatory situation and how we handle these these regulations oh I think we won't want to go any um so this is just an example or an overview of how facial recognition actually works you take the input image you look at the face detection and the landmark detection so you're detecting where the nose is where the eyes are where the cheeks are basically the
structure of the face you normalize it you extract those features and it creates a facial recognition template so a numerical representation of your face so never actually transacting images or video video frames we are transacting this numerical representation of your face you've got one to n on top meaning a one to many search where you're looking at the similarity score to previously generated templates in that database and then on bottom the one-to-one match where you're taking is this picture a match with this person or this picture and then what is the confidence score of that so as mentioned you know we're never transacting images we don't store the images we're only storing effectively two kilobytes of data on people which is the representation of their face through about 500 to 600 numerical values this is four different ways you can introduce bias in this process data collection the algorithmic design itself feature selection and the deployment in the use case so I know it's a lot of words and I apologize but I think it's
important for people to be able to read this sampling bias is when you're taking a data set used to train AI that is not represented of the real world application or population so if I'm training a database for CCTV recognition and I use it in Access Control it's not going to work as well so if you're trying to specify the use case in which you're training for you're going to introduce bias labeling bias all these images for facial recognition have to be labeled by someone right now it's not been automated and it's an extremely redundant and boring task and so you're adding the human element into it do humans have Prejudice or stereotypes that are influencing how they're labeling these faces now you've added labeling bias the algorithmic design bias so without knowing it you can have inherent biases in the way in which you're processing sensitive features of the face which could be the gender it could be this certain you know area of the eyes and that overfitting which is the unability for
us to generalize well if you have over fitted training data meaning if you have taken training data that is Consolidated around a specific gender ethnicity or group you're going to have overfitting and it will perform well on that group but no one else feature selection so irrelevant or redundant features it's features that are not relevant to the problem such as gender so that should not be relevant to whether or not I know is Sean Sean but we will find out later that it actually is and we cannot explain that yet the omitted variable bias so kind of as it sounds you're excluding important features that are relevant and can be correlated to other features so you're providing less of a map of the face to actually understand and then deployment and use case this is how you work with your customers so if the context is different than the one trained for or designed for as sampling bias you're going to have issues with deployment the customer will not be satisfied recognition rates will be
low false positives will be high and then feedback loops if you don't investigate your own algorithms and continue to to look at why they are performing the way they are and you keep adding more data you keep adding more faces to it you're just going to create this Perpetual cycle of bias and so why do the why does this bias exist um it's a great question that we're trying to solve right now and we think we have you know somewhat of the solution nailed down commercially available data uh it's it's cheaper so it's easier to come by but it is typically always concentrated around one group one ethnic group one gender and so you're you're by by way of using this data creating bias if you're not augmenting the commercially available data the country-based data so each country has specific data sets that they can license and if you are not getting a global population a representation of a global population in your training data you're going to introduce country-based bias data collection costs so
Google famously paid people five dollars per identity in Times Square to start training their facial recognition models you have to train on tens of thousands of millions of Faith tens of thou are sorry tens of millions of faces so you add that up very very quickly not many startups not many companies can afford that type of cost the data has the data cost has since gone down but it is still very expensive you have flawed data so you've got data that comes in blurry side of face and you're you're essentially infecting your algorithms with bad data to misperform or to have bias in it talked a little bit about unrepresentative or incomplete data if you're not representing the global population and you've got a deployment in the airport you're going to have a really bad time having to deal with the questions around why it's not working for people of specific countries coming through over representation the opportunities for false match and then on the deployment your quality angle and lighting this is something that
the industry is always speculating that the quality of the camera capture solution is not great the algorithm will not perform that well it's just been proven at the last Department of Homeland Security technology rally that the camera actually plays a significant role in whether the algorithm will produce the right result so everyone's been blaming the algorithms for not being up to speed or being as accurate as it could be but now we've been able to associate that the camera also plays a role in that decision and then I'm not going to bore you all with all of these but but it is important to run through a few and the reason that these are on the screen is because these are the things that we have to go through and look at When developing this technology and ensure that they do they are not present in our Solutions historical bias so a lack of diversity used in facial recognition algorithms prior to 2015 these algorithms that were being developed were based on very old data
data that was very specific to the country or to the region or the use case and so if you're perpetuating that you've got this historical bias now on your algorithm and if you don't completely start over or at least try and augment that data set this is going to be a very large issue for you representation bias if you've got one race that your data is made up of it will perform very poorly across everyone else self-explanatory measurement bias here's where some of the camera camera issues come in so the conditions in which a photo are taken are not standardized this will lead to inaccuracies so if you're not standardizing the way you're collecting data and then the way you're capturing data in the real world you're going to have a mismatch on performance evaluation bias talked about it earlier evaluation data is not representative of the target population aggregation bias you're aggregating data from multiple sources with no normalization so if you're taking cell phone footage if you're taking CCTV footage if you're taking
Access Control footage throwing it on there and not understanding how it's impacting your algorithm you're going to have different performance across those deployments population bias it's not intended for the the global population you're going to be very good with a specific population the time period and then the algorithmic bias which I think is the most important one here but it is the way that the algorithm is being designed to prioritize certain features again this is not always explainable so we don't always know why the algorithm is prioritizing certain features but we can run tests to understand is it and if so what is the consequence of that so here's an example of algorithmic bias so it's a deep learning model that's taking the top images as the input image and then matching them to the bottom and you'll see here the focus that we were able to cut uncover was that it was focusing in on a feature a feature of color here and so you're looking at the albino wallaby and the second the
second picture and blow at the cats and so this correctly predicted that the cat was the same as the wallaby and so when you have this bias in there you're going to get outcomes like this and you can imagine the implications of this so that leads us to you know how how do we as an organization how do we as an industry start to quantify bias in algorithms and there's really only one resource out there right now and it's called fairface it's an independent party has about 70 000 faces in their collection each categorized and I think the seven major ethnicities and so this is a data set in which we can take we can run our algorithm through and it Compares every face template against one another to General generally that that similarity score as mentioned earlier and so it will tell us you know how accurate we are across the different the different ranges of ethnicity and this is just an example of year-over-year progress that we've made as a company internally and
so you'll see on the horizontal axis 0.3 which is in the middle of that graph is an operating threshold that we would work with a client at you'll see from version 4 to version five it's almost at parity and so we were able to to use this exam if UL evaluation as a way to understand where were we underperforming where do we need more data and then go out there collect that data and then retrain and resubmit the algorithm and so again just a way for us as an industry to understand how you actually quantify this issue of bias and ethnicity so this is the one of the most interesting papers that has come out in the recent years uh was the Department of Homeland Security their technical paper series to quantify to the extent in which race and gender features determine identity so as mentioned earlier are the algorithms prioritizing features are they prioritizing race or gender and if so what is the out what is the impact what is the outcome the big
takeaway here is that the algorithms are learning features associated with gender and race and not just the uniqueness of the face so not just is this Sean it is looking at and understanding there are features that are clustered together that make up gender and ethnicity which is an extremely important realization for the industry so the question we have to then ask ourselves and our customers is this paper found that if you remove 10 of the features which describe ethnicity you have a drop in performance but you have a more fair algorithm so would you prefer as an organization to be selling a more more fair algorithm or a more accurate algorithm would you as a buyer prefer to have more accuracy or a more fair algorithm and so it's a really important philosophical question as well as a question that we now have to answer as an industry you know what what what should we be doing um and again I I show you this to state that these are very very difficult decisions that
require committees at companies to go through their pros and the cons uh and to go through how the market adopts this technology so it's really really important for us to understand where we sit and what can be done and you know what communication we have to have with the client what communication we have to have with the with the public and how this plays a role into the development of facial recognition technology so one of the big initiatives that I'm really proud of that I spearheaded this past year was the opening of a lab in Dallas Texas at Southern Methodist University it's called the intelligence systems and bias examination lab and so the the entire mission of this kind of academic and and private company partnership is to investigate facial recognitions facial recognition Technologies examine bias understand how to mitigate it and we can do this using research standards peer review scientific studies and so SMU is the vehicle now which we have as a resource to really dive into this question the most important
outcome is explainability if we can explain why this is happening why the algorithm is prioritizing features why it's not performing as well in certain individuals then we can mitigate the problem right now we're mitigating the problem based on assumptions and so you know we're showing that we are we are gaining traction with that problem and with those assumptions but it would be much more clear to the public and to us if we know exactly what is causing this and so um you know what are the implications of all this and and how do we adopt and how do we adjust uh as an industry you know as a company building this technology the most important thing on it's again being discussed being discussed today is with Innovative technology You're Building very quickly um and you're not thinking about you know all the regulation around you you're thinking about how can I deploy a product to Market that we're happy with that we're comfortable with in our case that is putting Humanity first but how you
know how do you make those decisions in a time to Market when regulation is lacking two three four five years in 2016 2017 China and Russia both came out and said whoever wins the AI race wins world domination that is something that you know has kind of provoked I would say other nations to really focus on artificial intelligence and to understand what can be done and then how do we put guard rails to ensure humanity is put first it's very very difficult question to ask uh very very question to answer but you know as an organization operating globally how do you ensure that you are meeting regulation as they come out you know the AI paper out of the EU is very different than what we're seeing in the U.S for regulation state by state so it's ensuring that you're building these solutions to spec for the the most strict of these you know of these regulations are these AI papers uh so that you can just you know you can be able to deploy
your technology in those countries without fear of repercussions without fear of breaking the law so one of the things that we've explored with the U.S government and are continuing to explore where we're hearing more about is dedicated Federal data sets so if the federal governments are allowed to put together these data sets that then are given access to specific vendors you know can we reach that parity across all ethnicities that we're looking for now the data has to be you know it has to be provided in the right way potentially in a sandbox environment where there's no ownership and and that's all stuff that we have to work through but you know it's an interesting thought to have if that is a possibility because you then bring the algorithm uh designers to the table when you democratize the data piece data has always been what has led the path here if you remove that as the hurdle then then you have to look to the design of the algorithm itself data augmentation and synthetic data
we've been hearing about synthetic data for training now for quite a few years in my personal experience it's not there yet we're unsure why we think it has to do with the way in which the faces are generated to be symmetrical many human faces are not symmetrical and so when you do have that symmetry you're reducing the accuracy of the synthetic data so I have I'm hopeful that we can eventually use that but as of today we cannot and then defining bias and then measuring it you know no one has really put forth a generally a genuinely accepted definition of bias and artificial intelligence everyone kind of operates on their own internal definition with their ethics committees uh principles such as when Microsoft puts them out and so a you know a global definition or even a federal definition of bias and then how you can measure it I think is a really important step forward bias lab certifications that's what I hope SMU can do with with Isabel is be able to certify that
these algorithms that are being being used for boarding airplanes that are being used to access sporting events have no bias or they are are near or at parity across the global population and then this gets into deployment is how do you deploy this technology so if you're running it in your Cloud you know you inevitably have access to that information and that creates a security concern or privacy and data transparency concern if you were self having your customer self hold host these Solutions then you're running it in their infrastructure they own the data they operate the data and ultimately they have you know they have this say in what happens to that data so it's really important to understand the dynamic between the two and provide that you know that level of transparency back to the customer what what's happening to this data why do you have my face to begin with why are you using it why do you need it and then what happens once I transact once I access the building once
I fly whatever the use cases and then customer compliance uh extremely extremely important for our industry that we work directly with the customers to understand their use case to review their policies to ensure that they you know pass the ethical guidelines of our company but they also have ethical procedures in place and then the evaluations pre-deployment and post post deployment and so once it's actually being used if it's in the same or if it's in a self-hosted environment it's very difficult to know how that technology is is actually being used it's one thing for them to tell you how it's being used but it's another thing if you don't have any oversight or or visibility into it so it's really important that facial recognition vendors are developing these relationships with customers to understand their use case to help them design for it to help them to optimize for it with camera placements with how they enroll individuals and so you know this kind of brings the the presentation together here but as I mentioned Humanity
first has got to be a non-negotiable for the design development deployment of this technology and now now I hope that you all have Awareness on the decisions that companies like ours have to make to deploy this technology so the biases that exist the different points in which they can be introduced to the process and how I think we should go about mitigating them so really appreciate you all being here this morning morning thank you for having me to Berlin I hope to see you next year thank you thank you very much Sean uh right on time so I really appreciate you as a vendor actually taking the effort to talk about the biases and actually talk about the hard things that we need to consider so kudos to you and thank you for your time Sean will be around for the rest of the day so if you have any questions please feel free to grab him great thank you Sean right so we're gonna jump