DeBias — Ending Personal Bias with AI
Why not?
Xenophobia.
One of the ugliest most disgusting issues humanity is facing.
It has lead up to countless genocides and massacre of millions of innocent people because of something they often have no control over — their race, skin color, body size…
“Genocides and massacres are not something we see a lot of today so shut up”
I am not going to talk about all the genocides you’re probably unaware of around the world AS YOU’RE READING THIS ARTICLE RIGHT HERE…
But even if we don’t see physical portrayal of xenophobia in Western countries, there are some other forms of it…
But there’s one particular form I believe is most destructive of all — systematic racism…
Why?
Systematic racism is not the usual opponent you openly fight, it’s rather the friends that betrays your trust and stabs you in the back.
Systematic racism is running in our blood, embedded into our society and making millions of lives miserable.
How do we know it’s a thing?
That’s a great question!!! I am glad you asked 🤗
Since this form of bias is never explicitly expressed — and is not intentional in first place. It’s very difficult to recognize and acknowledge.
However, we can prove its effects through a variety of statistics that show ongoing inequalities discriminating certain groups.
Where is that even coming from?
There quite a few ways we can grow our systematic personal biases — that includes the content we consume, the people we surround ourselves with, our personal experiences with people from certain groups…
With social media taking over our lives, there is more education around such issues… However, it’s also an additional medium for the spread of hate and Xenophobia.
PROBLEM??
A lot of us — people from minority groups can probably relate — have been around someone who claims to be open-minded and acceptant, but in reality, are automatically biased towards certain groups and against others.
They do not mean to hurt anyone intentionally, and there is no blame on them because that is a result of what’s happening around them…
But we can’t deny the fact that that is an ISSUE.
Now, you might say “if they knew about it, they would likely work on being more intentional about their interactions”
Ladies and gentleman…
THERE LIES THE KEY TO THE PROBLEM.
“if they knew about it”
Meaning the problem here is awareness, people portraying systematic racism are not AWARE of what’s happening, they are UNABLE to identify it.
But what if we could help them IDENTIFY any biases that they might have.
That’s what I am currently trying to do.
HOW?
You’re about to learn everything about that project, from start to finish, but without further ado, let’s get into it!
P.S: If you’re not a big fan of long articles, I promise this one will be a quick read — No longer than 7 minutes.
Introducing…
DeBias
The WHAT
Tell me about the project in 7 sentences.
On the most basic level, DeBias is an AI/ML powered application that analyzes users’ online patterns to identify their personal biases to help them be more aware and intentional about their interactions with others.
Technical Aspect
There are four main components to the DeBias project. 3 out 4 of them are to actually collect the data that would be inputted for the major component — the unsupervised learning model.
Natural Language Processing (NLPs)
Since part of the application will be analyzing users’ comments and written content, I will be using Natural Language Processing to detect bias in text.
Object Detection
The other part of the application will be analyzing visuals the user is looking at. Obviously, the software needs to be able to determine whether pictures contain human faces so that it can classify what groups the people in the picture would fall into.
For this part, I am considering 2 options — either building out my own YOLOv3 model or using the existing facial-recognition library on Python.
Image Classification
Once the software detects human faces on pictures, it needs to classify it based on factors like age, gender and ethnicity. This one will play a major role in the process, so it specifically needs a high level of accuracy.
Unsupervised Learning Model
This is the final and arguably the most important component. The model takes in different metrics (e.g. how much time users spend per post/per group, likelihood of user commenting/liking posts for different groups…) as inputs and clusters them to predict any biases that the user might have after a certain amount of time.
Challenges?
There are a couple of problems that might come up as the project progresses…
Training Data
I did some research around, there are existing datasets I can use for my Natural Language Processing model, my Object Detection model, and my image classification model.
However…
For the unsupervised learning model, there are no existing datasets. What does that mean?
I LITERALLY HAVE TO BUILD MY OWN DATASET 😭
It can’t be as painful as thatttttt
It might not be as difficult as building out datasets that involve patients’ data… but there are some serious issues that come out from that either way…
For a higher accuracy for the model, I would need a fairly large dataset. But the dataset involves privacy concerns.
Getting people to share their online patterns’ metrics is no easy-task. Additionally, there is a whole strategy to be defined in terms of collecting data. Is it going to be volunteer based? Monetary compensation for participants?
Privacy Concerns from Users
Let’s suppose we were actually able to build a solid dataset and build out a decent model…
It’s time to deploy the project… I need to get users on board.
But from pitching the idea to some industry experts and potential users, something that kept coming up was usage of data.
Especially with the recurring data privacy controversies that broke off trust between social media companies and users, getting people to trust DeBias with such sensitive data like their personal biases might not seem like a good idea.
Obviously, no data will be collected aside from training, therefore nothing will be shared with third parties. But good luck to me talking people into that!!
User Activity Frequency
Essentially, the more a user is active on social media and online in general, the more accurate the models’ predictions will be. So as it is of now, this solution would ABSOLUTELY not work for people who spend very little time around social media.
What now?
At this point, I built out my image classification and object detection models. So the next steps would be the Natural Language Processing and the Unsupervised Learning Models.
How long is that going to take?
I honestly can’t really tell… Depending on what I have going on over the few weeks, I might do it on and off. Maybe I will double down on it throughout next summer, but that is yet to be determined.
Conclusion
There you go…
Everything you possibly need to know about the DeBias project (for now).
Interested in learning more as the project is progressing?
Feel free to follow me on Twitter, connect with me on LinkedIn, and subscribe to my newsletter to receive monthly emails from me.
