IBM THINK 2019 Day 3 Recap: Reducing Bias with IBM

Warning: this content is older than 365 days. It may be out of date and no longer relevant.

IBM THINK 2019 Day 3 Recap: Reducing Bias with IBM

In today’s IBM THINK 2019 Day 3 Recap, we look at the problem of bias in AI and machine learning, the three locations in the development process where bias can creep in, and how IBM helps us mitigate bias with two key products, Fairness 360 and IBM Watson OpenScale.

Learn more:
– Fairness 360 (open source)
– IBM Watson OpenScale

FTC Disclosure: Trust Insights is an IBM Registered Business Partner. Any transactions you make through us with IBM benefits the company financially.

IBM THINK 2019 Day 3 Recap: Reducing Bias with IBM

Watch this video on YouTube.

Can’t see anything? Watch it on YouTube here.

Listen to the audio here:

Download the MP3 audio here.

Machine-Generated Transcript

What follows is an AI-generated transcript. The transcript may contain errors and is not a substitute for watching the video.

In today’s episode recapping day three at IBM think two major things that really stood out from the day there was a ton of good stuff. One of the things I really enjoy about this show is being able to walk up to people who are senior or master level data scientists and machine learning experts and ask them questions like, Hey, can you take a look at this code that I’m working on and have them go, Oh, that’s how I do that. And then get productive feedback about ways you can improve your own code. So that was that’s been a lot of fun. But there were two initiatives that I thought were super important. One is called fairness 360, and one is called Watson open scale

when it comes to bias in our data. We’ve talked a lot about this. You’ve heard a lot about this

in the news in major stories about companies big companies like Amazon.

getting in trouble for having bias status of bias AI

causing problems and hiring and discriminating in certain groups. And a story I tell it from one of my keynote talks is Police Department trying to predict recidivism and failing miserably at it because of bias.

There are three places that bias can exist in the path towards AI. You can have bias in the data where as you’re collecting the data itself is biased

and that it wasn’t collected properly. There was data missing or in many cases there are problems in the population data itself. Example I said often is African American healthcare data is inherently biased and flawed because

of systemic discrimination within the society itself. So that’s one place bias can occur. Second place bias can occur is in the model creation when you’re creating an

I’m model which means that the software is taking the data and learning from it.

The software can create biases,

simple example against what we’re hiring example. Even if you say I don’t want to use gender or age, if you had things like someone’s Facebook movies that they like, and books and music they listen to, you can infer a with a high degree of accuracy a person’s age, and gender, and, and even ethnicity.

And so you can create a model that has bias in it.

And then the third place that a bias can occur is even if you’ve done a good job with the data, you’ve done a good job with

creating the model. The third place it can occur is in production when you deploy the model and then new data starts coming in. Because you’ve trained you’ve you’ve built it, trained it and now it’s now it’s in production bias can happen after the model has been deployed.

Because you can get unforeseen data or unusual data that you didn’t expect to get the classic example. This is when Microsoft released its tail, Twitter bot, and people fed it new, highly corrupted data from the perspective of the things you’d want to talk to a robot about, and turn into a racist, pornographic, but in 24 hours, thank you, Mandy.

And so these are the sort of the three places bias can occur. Now,

mitigating bias means taking all three places and trying to put safeguards in place for them. So truly announcements that IBM made at the show and have deployed product for that that was really interesting. One is called fairness 360. This is an open source, totally free set of 10 Python libraries or Python algorithms that you apply to the data side and it will look at your data and say there are problems in the data there are populations that are underrepresented.

You specify like, I want to make sure that

say we’re doing we’re doing a test on gender, it will look at and say, Hey, I think that this there’s an issue here

gender is under represented by this or or races under represented by this, you would specify protected classes and then it would notify you that there are significant problems,

you can do a lot of that manually with various machine learning software, but having some pre packaged stuff that can you can just put into your workflow and immediately get identify it is really, really powerful. The second place that you can

mitigate bias is when you’re doing things like feature selection in Watson studio Watson studio as you’re building your models, you can identify like there’s something wrong here. But the third one and thing I think is probably the most powerful of the in house products is one called Watson open scale. This is for managing the models when they are in progress.

And it is doing what it’s called perturbation testing,

where it’s constantly, you know, sort of holding back certain data and swapping it in and out as the models running as the models running

and identifying, hey, things are starting to drift over this, this model starting to exhibit bias. And like fairness, 360, you specify protective glasses, you say, you know, gender must remain at a 5050 split plus or minus, you know, 1% each way,

you know, African American representation must remain at, you know, 17% plus or minus

a certain guardrails and it will either read to help you return and fix the models or the worst case just shut it down. If a model goes completely off the guardrails.

If you think about it,

there is

with open scale you’re talking about.

essentially doing repairs on the car while you’re driving it right or or repairing the plane wallets in flight is extremely difficult thing to do and the fact that it has biased detection mitigation built into it so while you’re managing your models in production is really an incredible engineering feat to be able to say like yeah we can we can fix the model while it’s in production now I have to pull it down and not have to wait for the next round of retuning because it’s your open skill will help you tune it and fix it even if you didn’t intend in the original deployment model to have it be tuned in production. So super, super interesting and important

product announcement and deployment.

There is not enough thought being given it in the machine learning world to

the mitigation advice. It is not a part of the average

checklist for when someone is is putting together

what they want the software to do. And so having these tools be available within

a production environment and having in some cases like with open scale, having it be baked in so that you can’t not have bias detection is really important. It’s really important. It’s really powerful to help make sure that AI serves the outcomes that we want to achieve.

The The third thing about open skills bias detection, I thought was really interesting was that

it prohibits or, or reduces inferred variable bias.

So you may say, going back to the example from earlier

I want to

not use gender but if you’re out your model put together an inferred variable lovely between music and book and move.

likes you have a problem, you’re still you’re still effectively creating a gender valuable when you have protected classes in open scale and you say I gender is protected, this is the outcome, even if the model creates that inferred variable is still being held on course, like holding the steering wheel so that the the car doesn’t drift

by those protected classes. So you say race and gender and age and protected classes. And no matter what the model comes up with the mix and match in those inferred variables, it’s still cannot say okay, I’m going to now start discriminating against using this inverter variable because it’s over here at saying no

gender still have to say 5050 and what about this? Nope, gender. So let’s say 5050.

And that is also really really powerful something again, if you’re if you build a model and you’re not thinking about bias open scales over the backstop to prevent you from going off the rails and doing something you didn’t intend to do. And then waking up one day going oh, my God, are you

Our system discriminates against women.

So really, really useful stuff Take a look. You can see Watson open scale on on the IBM website and you can actually try it out the lite version if you want to try it out with your with your models and they do have to be in production through one of the many machine learning systems that actually deployed so really cool stuff.

If you have any questions leave them below. As always, please leave please subscribe to the newsletter and and the YouTube channel and I’ll talk to you soon.

Machine-Generated Transcript

Comments

Leave a Reply Cancel reply

Pin It on Pinterest