GDPAI Lect3 Can we make AI systems fairer? | Bias Mitigation techniques in AI

追加: 2026-05-09

[00:00:02]Hello student, welcome back. Welcome to this uh third lecture. So far what we learned in the previous two lecture that how bias enter into AI system and we also learn how they can be measured or fairness can be measured or the biasness. Now we have a very practical question that once we know or we have detected the unfairness what can we do about it means simply saying that this system is not good or that system is not good it's very easy to point out errors but uh pointing errors we should have a solution so now question come can we make AI system fairer fairer means that detection ction of bias is only the beginning but the real challenge is how to mitigate them means whether it was during the AI model making or during the data collection or when it was already deployed already AI system running how to mitigate them so let's uh move forward so that become the question the bias mitigation technique in EI and in this lecture uh why mitigation matters because uh we need to uh have this into our mind that AI bias in AI can scale into discrimination. Small bias can lead into discrimination and this discrimination may be intentional or may be unintentional.

[00:01:43]Means the people who have designed the system they may not have designed intentionally but sometime uh things are happening so fast in AI. The proper care of data collection and the bias mitigation strategy has not been put into place.

[00:02:02]They are simply getting the data and soon run it and train it and release the model very fast. So the bias is intentionally getting into there and it can scale up fast. Now AI because of biasness it can and amplify the unfairness because unfairness for a group of people if that group is large then it can also go to a very large scale. Now that leads to mitigation is essential especially for high impact domain like if you are creating an AI system for societal justice or for large scale uh deployment for uh welfare measures for distributing any welfare kind of measures health care system these are all large where the biasness can lead it to bigger uh impact.

[00:02:59]small places like you know image not been displayed correctly or there are errors in the text image text generation or error in the image generation those are small thing or a website created using AI has some biasness that is a small uh thing which doesn't impact uh in that large scale but in a health care or welfare distribution by the state government or governmental If bias exists there then it can impact in large scale. So if we have a biasness in hiring or loan system or policing surveillance then harm is not only technical it becomes social and sometimes it become legal uh problem that is why we need to mitigate uh bias and it matters.

[00:03:53]So in this lecture what we are going to learn today is that where bias uh mitigation can happen mean in the whole AI pipeline where are the places we can do the bias mitigation and what are the techniques available to mitigate bias at those places common methods for reducing the bias and why uh we should think about inclusive design from the beginning when we are making an AI system. So we should think that from the beginning itself we design an AI system which by design inclusive and uh has a lesser bias. Uh we should mean when we are an AI developer we all the time think about training the model but we see that just uh clone the GitHub data set is there model is there base code is there just train it just use it change something and uh get that thing out. But if we have designed an AI system by thinking inclusivity and fairness in mind then that system will have lesser bias. Now there are different kind of approaches technical approaches human- centered approaches.

[00:05:04]So how to combine those to reduce the bias mitigation or mitigation of the bias or reduce the bias is going to be the focus of this lecture.

[00:05:16]Now bias mitigation when we think just like in the previous lecture we saw that fairness there is no universal fairness metric to judge uh fairness of a system.

[00:05:28]Similarly there is no universal bias mitigation technique. Now if someone asks you that use one bias mitigation make my system completely bias-free. You simply say to them it is not possible because there is no one technique which can mitigate all types of biases. We can reduce the bias like before training if we have uh already we are building an AI system and we are uh going to train the system. So we can use some strategy and mitigation method where we can reduce the biasness before the training. Now we already have collected the data and we are about to train the system. Now during the training also we can reduce the bias by incorporating some mitigation strategy. Now training is already done and our system is going to be deployed. So after training also we can use the mitigation strategy to reduce the biases. Now before if you are designing a new system then we can also think of building an inclusive system design for AI model or AI system which has inclusivity and diversity and fairness taken into account. So bias mitigation is basically not a single trick. It is a set of interventions and method across this EI pipeline like data, model, prediction and the whole design process. So let's move forward how them one by one.

[00:07:05]So there are three technical ways to mitigate the bias as per the AI pipeline like first is pre-processing means can we improve the data before the training reduce the biasness in the data can we improve the sampling method can we improve uh the labeling method can we relook into the data and it distribution does it have the biasness or not so pre-processing step we can Then next comes the in processing means can we modify during the training process can we modify the learning algorithm can we introduce the strategy which helps in reducing the bias during the training like if aggregation of the bias is happening can we deploy or use a method which can overcome or reduce the aggregation related biasness during the training process. Then comes the post-processing method like after the training is done can we adjust the output like often we are using pre-train model um and simply uh like many people are building the AI system using different kind of foundational model in the back end from open AI from claud from rock or llama or uh gemini many places but those models are been trained and often For a small company, a small team of two people, five people, 10, 20 people, it is not feasible to really go uh do this foundation for this foundational model making, data collection, pre-processing, training, all this is not feasible because they often require billions of dollars and huge amount of manpower and infrastructure power. Most uh people which are small company or small team they do not have uh let's say 40,000 GPU system or a billions of dollar for training to pay for cloud to do train the model and also billions of dollar to for data collection. What we you uh what normally people are getting is pre-trained model and once we have a pre-trained model can we adjust the output after the training. So there are these are the three technical ways.

[00:09:28]Now if you take this uh nutshell that what are the different mitigating uh what are the different methodologies are there by which we can mitigate the bias in AI model. So there are AI life cycle if you go through there are pre-processing uh in processing and post-processing.

[00:09:48]For pre-processing there are steps are there for in processing there are steps are there. post-processing where there are steps are there. We will look into them uh moving forward.

[00:10:00]Then methodologies. So if we classify like like this like for pre-processing we can have the uh fair representation we can have optimized pre-processing we can remove the disparity then we can also improve the sampling method. We can uh reway the data. We can rebalance the data like in in processing. We can use adversarial learning uh prejuditize removal thing.

[00:10:32]Then we can also do hyperparameter optimization so that biasness can be reduced like once the model we are training we can use different kind of hyperparameter to overcome the some certain kind of biases. Calibration we can use then uh different other techniques are there like constant optimization then grid search so that we can do a quicker search but a better search then post-processing we can do a equalized code for post-processing we can do a threshold optimization of the post-process model then we can also do deterministic retraining sometimes the model is there like you take the pre-trained model and after that you have to use it for your domain. So you can use some amount of data from your domain to retrain it so that for your domain uh where the data is not represented in the pre-trained model then it can be retrained so that you can reduce the biasness.

[00:11:38]Now fairness can be introduced through reinforcement learning the adversarial reinforcement. Then there are regularization we can use. There are hyperparameter optimization. We can use fairness constraint. We can also we can also do the causal approach finding the causality of uh biasness and using those approaches. So let's take with the one simple example that we are there is an AI system many case study which is uh doing the hiring decision uh automatically. Now this hiring model is trained on historical recruitment data. So in a company historical datas are there. Now the problem exists is that in past there were fewer women working in the workforce. Okay. So since there are less women has worked in the past uh with time more and more women are into the workforce but this data this model is trained using the past data. So there were fewer women in the workforce. So the fewer women are represented in the past hiring. Then this due to this model has learned some bias pattern and qualified candidate they have only got from one particular group and from other group even the qualified candidate is there they are getting a lower score by the AI model.

[00:13:06]Now how to fix this thing for this model? Now we can do pre-processing based that means fixing the bias before we are training the model. So one thing we can do data cleanup that we need to evaluate in the data that whether there is sporious error ambiguity and uh other problems exist in data or not. If there are already un uh unbalanced data is there do we need to reway before taking this data for a training purpose.

[00:13:43]Then we can also balance the group.

[00:13:46]Means if you do not have enough data from one sample or group, can we get more data or can we use the data balancing strategy to balance the data?

[00:13:58]Even after that can we remove uh proxies biased proxies from it or like improving the representation like past the female datas were less. Can we get more data and that can we get more data means can we have more representation? So, so that is or if you are building a system where we have collected the data where one class or few class has underrepresented can we ask our u uh company or uh that can we collect more data of other groups as well because currently data is uh biased and we need to mitigate the biasness in this stage itself.

[00:14:45]So what are the different way also we can do the the pre-processing the data uh mitigation of the biases. One is a oversampling like sometime increase examples from under represented groups we can oversample it. For example if we have a group which has uh 10% data and other group has 90%. So we can do oversampling from lesser represented group and by that we can uh match the data sometime under sampling like we can reduce the sample from over represented group like if the data has 10% from one group 90% from other group from the group which has 90% we can under sample it that means we can take all the 10% sample from this group and randomly chosen 10% sample from is 90%. So then it can become balanced. For example, if you see in this that before the balancing group A had a lot of data and group B had less data. So now after balancing we could have we are able to make it equal equal. So that way over sampling under sampling we can uh do this uh data level u uh samp representation problem uh reduction.

[00:16:06]Now we have balanced the data. Uh is that enough?

[00:16:11]Now we might have balanced the data but that is not enough because of various reason.

[00:16:18]If let's say in a in a healthcare domain most country the nurses are female and not male. Most of the nurses are female.

[00:16:30]Now if we are synthetically balancing the data it may not remove the historical bias because currently also if you are going to use the AI system the sample most data in the nursing domain will come from female not from male. Now we might use proxy but proxies are not perfect. It may still encode unfairness. So that risk is there. Now we have balanced the data but still it can produce unfair decision.

[00:17:08]Why can it happen? Yes, it can happen.

[00:17:12]Now the representation matters.

[00:17:16]Representing from various group matters but fairness is not guaranteed just because data set look balanced.

[00:17:25]data set might look balanced but it might be not representing everything from the different groups in the data.

[00:17:33]So that also uh can be a challenge.

[00:17:38]Next comes what to do during the training. So mitigating the bias during the training.

[00:17:45]So in the deep learning or AI model training we have something called loss function or objective function. how based on which the model uh estimate the error and try to learn to reduce the error. So that is called loss function.

[00:18:02]So can we design a fairness aware loss function?

[00:18:07]If not can we do adversarial debiasing?

[00:18:12]Sometime we do a regularizer for debiasing in the model training. In our adversarial model uh training we have many advers loss function. So sometime we do a regularizer to uh rebalance the uh debiasing.

[00:18:30]Then we can add fairness constraint in the uh model training. Then the if we are not doing the regularization in the loss function then we can also do regularization in the whole training pipeline for getting the equitable outcomes.

[00:18:49]So now after this in processing or during the model training time mitigation strategy we have already trained the model. Now do what we if we do not have any other thing to do let's say we cannot go for data set collection like you join a company the data set has been collected model has been trained now you have to uh make decision after the training. So in that case you can use the threshold adjustment process to do the bias correction or reduction. You can do the calibration of the output and you can make uh special specialized process for group specific decision correction.

[00:19:33]There are uh fairness aware uh releabeling we can do at the after the training process. So if you want to see this pipeline thing like learning pair representation there are uh we can apply AI model on refined data we have to first we can do data regularization and based on the fairness criteria we can compute fear representation given the data regularization criteria then during the training uh you can see that original data is there then we have transformed the data and then we are passing the transform data to create the AI model. Now to transform the data we are using the uh uh transforming using the some kind of disc variables. So that can be uh help in modifying the data in the original base.

[00:20:33]So that's called optimized pre-processing. If that is not visible or already done then we can rebuy the data like we have a training data then we train the model then we can use this validation data to ree the model and we can have a task loss we can add a fairness loss so that we can uh reway the data then the already discussed resampling oversampling it's given in the more detail and references.

[00:21:07]Then comes the adversial debiasing. So I was telling you that uh we can add the adversial debiasing method to improve the uh biasing in the model training time and sometime data is giving a prejudice so we can use methodology to remove the prejudizes. So that is a regularization term is added the L prejudice in the after the output in the training phase.

[00:21:37]So now question comes we have in processing pre-processing in processing post-processing which mitigation approach should we choose again the right strategy depend on uh like pre-processing when the data is the main issue then we use a pre-processing step in processing when you are in a control of model training then you can do the in processing or postprocessing when you only control the decision at the end you can can do it or you can do all you can do everything if everything is in your control.

[00:22:14]Now going forward so the problem is same but different mitigation choices can be there right the same hiring model like you know you have a AI based hiring that is a problem statement now same problem that ging the bias so pre-processing step in this model is that we rebalance the data remove the bias proxy features in the training time we train with fairness constraint In the after the training time we adjust the decision threshold. So that is once the training is done that time. So takeaway is that different types of intervention can target different sources of unfairness like in the previous lecture we saw that there are different types of unfairnesses there.

[00:23:03]So to handle different types of unfairness we need to we may need to use different intervention or mitigation strategy.

[00:23:14]Now what are the different uh technical fixes are there? So technical fixes are necessary but not always sufficient. So a model can look fair in technically but yet still fail people if users are excluded from design. That means only the AI developer are building the system. You have not talked to the users. So since users were out of the design or during the data collection different diversity population has not been involved during the data collection. That means the team which is getting the data. A different diverse team need to be there. The team which is training the model that also should have diversity. The pre-processing of the data should have a diversity post-processing. So in the company also diverse groups should be there. If the users were excluded, diverse users were excluded or users itself were excluded from the design like we are building a healthare system and we haven't talked to the medical professional. So that can also introduce the biases. Now the accessibility was ignored like we haven't taken the different types of diverse group and so we make an AI system and in our team we did not had a people with representing the disability group. So what typically will happen that we will not know the disability group uh different problem they are facing. So the system will be designed will be having the accessibility issue like like red green color blindness is very common and if you don't know in none of none of us in our test team is red green color blind so we will not know the challenges of them and we will make the system which look red and green and that can make them uh system uh as accessibility was ignored or a blind people how they access the device or a people with hearing impairment. If your system has a voice model and the people have a hearing impairment then accessibility issue was ignored there. Then comes the cultural context.

[00:25:29]Now for especially in the developed countries these developed countries are mostly in the western countries and let's say the Europe or Norway Norway is predominantly homogeneous means here the majority of the people are uh eating the same thing speaking the same language culturally the same thing and following the same religion.

[00:25:55]If the cultural contexts are been ignored there, so the different cultural people who are going to be using the system in Norway and they have not represented. So those aspect will not be taken into account and then the AI system built like that will be unfair towards may be unfair towards them.

[00:26:17]There can be the problem that harms were not anticipated means we designed a system and thinking that perfect and everything everything we taken into account but some aspect of uh AI system output where it could do a harm uh we have not anticipated and later on we come to know that uh it is working but then this particular thing is creating the harm like recently uh Anthropic Anthropic CEO came up uh in the public and he mentioned that uh the model he has developed or his company has developed is so powerful that it can break the security aspect of almost every system exist in the whole world whether it is Windows system whether it is Chrome system or Mozilla, Linux, Apple anything. So they became uh if they would have released the system. So recently just today I learn I read the news that the Mosilla used this anthorpic model to find out 249 vulnerabilities in their Mozilla Firefox browser and when they found out and they now can recover it remove it and then they can their system will be more secure. But what if they have released the thing? So the people who have the malicious intention, they could have used this system and uh used the vulnerability in the Mozilla Firefox.

[00:27:50]This vulnerability can be in the Chrome, can be in the Safari, can be any browser or can be any system, Android system, Apple system, any system or any banking system or any governmental system, any encryption system. So this model could have been used to uh bypass those vulnerability but uh sometime we do things we have not anticipated and that can lead to harm. So now once we know that uh mitigation strategy is there next comes this can we design a system which is from the beginning inclusive and diversity adher so let's see how to do thing now fairer AI will require more than a model level correction means we have we as a AI developer you always think like model training model model model right but if we really want to make a fairer AI uh system we have to think more than model level correction like pre-processing in processing post-processing we may have to think beyond that and that is that leads to why inclusion matters in AI so why I'm saying that that we have to think beyond the model level pre-processing in processing post-processing Because if important groups are not represented in design, testing and deployment, the system which we have developed may serve some users and may not serve other users or may serve the other user poorly.

[00:29:32]So we why we are it is important because we are building an AI system that shapes real decision in healthcare, finance, education, legal and many aspect. So exclusion of some people or some group can produce harm and it can introduce inequity or inequality.

[00:29:54]Now diversity is not an option while building an AI system. It is a design necessity. So nowadays uh most uh company which are cashri like Google, Facebook, Apple, OpenAI uh then um Grock, XI, all these anthropic they all uh have diversity taken into account because diversity is a design necessity and everyone knows it.

[00:30:28]So even you belong to a smaller company so you need to take this into account.

[00:30:34]If you cannot hire full-time uh people at least you can take the feedback from diverse population or diverse team for a partial timing or partial time so that at least something you can capture from them.

[00:30:52]Now what are the common pitfalls when we are designing the AI system? So the common pitfall is one is everyone knows that data set data set can be biased.

[00:31:05]Right? Now the other thing which just now know that our team only coming from one particular uh strata or one particular uh ethnicity or one particular uh representing one particular type of uh society then nondiverse development team in AI can lead to uh problem in AI design. Now next is the uh when we designing a system we have ignored uh in inequality in structurally. So when we are designing we have structurally ignored the inequality we are introducing and then same then other thing is the accessibility feature like our AI system design is not inclusive.

[00:31:58]It's not accessible to certain type of people with a disability or other kind of trouble like uh it's not blind friendly, it is not uh um hearing aid friendly, it is not for different purpose, different like people with disabilities is not friendly. So accessibility failures will be introduced if not taken into account. So the thing is that sometime unfairiness is not caused by bad intent. No one most of us are not having bad intent that by purpose we are designing this kind of pitfalls but it happens because certain people voices and realities were never meaningfully captured or included. So we have to be careful about it.

[00:32:48]Now what are the principle by which we can make the design inclusive AI system.

[00:32:56]So we have to make by rule equity by default. We need to have equity by default and we have to design with affected community. For example, if you're designing an AI system which is to uh which is catered for let's say in the Norway for Sami people. Now in uh in your team who are developing this AI system there was no one Sami no one from Sami language I uh means if there's there's no one from Sami language then that uh design you team will not be able to capture the troubles and problems and needs of the affected community. For example, if you are designing a E AI system for some very far away, let's say uh some people in uh ethnic group from Australia and the person representing that ethnic group is not into the system. So it is likely that we are going to miss some of the component while designing because we have not included the affected communities.

[00:34:13]Then comes the transparency and explanability. So when we are designing the system, let's make the data record and model parameters and post-processing whatever we are doing which will affect the final outcome we should make it transparent.

[00:34:32]And second thing is we should design an system where explanability is incorporated. So explanability methods will be there so that if anything wrong or why why inference is coming why the system is giving this inference whether it is prediction or classification or other thing which my system is doing we should trace back that in system where is the error is coming so adding transparency and explanability will help making the inclusive design of AI system finally the cultural and contextual sensitivity as I said contextual if you are making system for Sami people we should have people from Sami community. If you're making system from some ethnic group in Australia, we should have people from that community. Now if you are designing a system, let's say Norway and then Norway represent a cosmopolitan society.

[00:35:26]Here every religion people are there and we have over diverse population uh diverse country population exist. Now if in our AI system team we do not have we do not have cultural sensitivity we do not have uh that aspect then we will design a system which will more or less represent a homogeneous society because 98% is homogeneous here and the 2% which has not been represented will be affected.

[00:35:56]So inclusive designs means that we should be building awareness building with awareness of who might be affected by the system on who otherwise will be excluded.

[00:36:12]Now what are the different strategies?

[00:36:13]Simple strategy to build an inclusive AI system is we should have an audit of the data set for representation gaps. So audit of data set is there like nowadays many people are fearing that AI will be coming so fast into the everywhere that job losses will be there. So I am telling this AI auditor will be a new job in the AI based world. So audit of the data set for representation gap applying uh bias mitigation during the training using fairness metrics during the evaluation process involving multiple interdisciplinary teams so that we can have inclusive AI thing and we should be testing it across different user groups. So whoever is our user of a system, we should be testing with those user groups so that we will be able to capture if our system is doing uh not as it's meant to be for different groups.

[00:37:16]So so point is that fairness of AI needs both technical and systemic solution. So technical solution is data balancing fairness aware training uh output adjustment fairness metric while systemic is uh solution is that we should have a diverse team who is building AI system. We should have participatory design means we should have a participation from the various groups. We should take into accessibility into account and transparency of the whole AI system development and since AI training is continuous process and we should have continuous editing at various scale at various uh time stamps with this we talk about lot mitigation strategy do we have a practical nutshell place to think about it yes so the practical chess This is not exhaustive but just in one nutshell for reducing the bias. So before training we should think is the data representative of the problem. So target problem statement or target population for which we are building the system. During the training uh our fairness constraint or bias aware method have been used or needed. After the training uh our when the training is done when it is giving the outcome is the outcome given by the model is fair across group if not fair what can we do about it. So once we deployed it uh we have to think who may be excluded or getting harm with our system design.

[00:39:00]So we learned various mitigation strategy of uh and also how to build inclusive AI system or take into account different uh things while building the AI system which is inclusive and diversity aware. So the takeaway lecture of today's lecture is that bias mitigation can happen before building the AI system, during the building the AI system and after training the AI system. Balancing data helps but it does not solve the problem all always right. So we can balance the data but that doesn't mean it is going to solve everything. There are technical fairness are there and design inclusivity means when when we making the AI system we should be have inclusive design must work together technical fairness as well as design uh inclusivity then fairness is uh requires ongoing evaluation is a continuous process not that one time recorrected uh then it is done if our model is getting retrained redata and also during In the deployment phase we should have the fairness evaluation that is it still working the way it should work or some kind of biasness have start coming up. Now with this in the next lecture what you are going to learn about it that inclusive and diverse AI system design and where we go beyond model correction and ask bigger question how do we design system which is inclusive from the beginning and with this I would like to say thank you and follow for the next lecture if you are interested and hope