Announcing the 8-Bit Bias Bounty Winners

Originally Posted: December 19, 2022

We are pleased to announce the winners of our 8-bit Bias Bounty. The competition drew challengers from around the world, ranging from seasoned ML practitioners to students. Our winners will receive cash prizes, and all contestants may opt-in to receive fun swag, including Amazon Kindles and Bias Bounty stickers. We also wish to thank our sponsors at Reality Defender, Robust Intelligence, Microsoft, and AWS. Bias Bounty Winners All winners are graded based on a publicly available rubric. Our winners are:

Supervised Model:

• First Prize: Prakhar Ganesh aka Breeze ($6,000).

• Second Prize: Yonatan Perry aka To Arr is Pirate ($4,000)

• Third Prize: Justin Chung Clark aka Squashbuckler ($2,000)

Best Unsupervised Model:

• Yonatan Perry aka To Arr is Pirate ($4,000)

Judge’s Commendation for Most Innovative Approach:

• Andrew Bean

• Jonathan Rystrøm

• Lujain Ibrahim aka OII crew

About Our Winners

Prakhar Ganesh is a Masters of Computing (AI specialization) student at NUS. He did his bachelors in Computer Science and Engineering from IIT Delhi in 2019 and then worked at Advanced Digital Sciences Center, Singapore as a Research Engineer for two years. His research interests involve studying the challenges that arise in machine learning when moving from the sandbox to the real world — including fairness, privacy, adversarial robustness, and model compression — by uncovering the learning dynamics of neural models. Bias Buccaneers: What activity or approach did you take that you think made your approach stand out? Prakhar Ganesh: Additional of “background class” (i.e., all images that were not faces) to denoise training data and also (I hope) pass the random prediction tests during validation. Also, inal training on a combined train and test dataset (i.e. all publicly available images). What was the most challenging aspect of the competition that took extra consideration or effort? Accuracy of skin tone classification. While reading the competition details, I thought I would end up spending more time on making sure there is less disparity. However, by the end, I didn't even consider disparity penalty and was more focused on getting some decent accuracy for skin tone classification. That was definitely the most challenging aspect of the competition, and even then the final accuracy achieved was quite low. There is a huge margin of improvement for the same. What advice do you have for aspiring Bias Buccaneers? Gear your solution to the competition specifications, but only after spending an adequate amount of time making more general improvements. The latter is your baseline, and it needs to be strong. But the former can help you squeeze a few extra points in the final ranking. ____________________________________________

Yonatan Perry is a seasoned software engineering leader with over 20 years of experience in programming, machine learning, cyber security, and innovation. Having joined Cybereason as their first VP of R&D, he has been directing Innovation and Data Science research there for the last several years. He is passionate about data, science, the people behind them, and the startups that drive them. He is also the father of two ninja girls and a 2nd Dan black belt Karateka.

Bias Buccaneers: What activity or approach did you take that you think made your approach stand out?

Yonatan Perry: This challenge was unique in that it aimed to mimic real life by having a complex score plane where different avenues for value often compete and sometimes directly contradict. My focus was constantly on finding the best balance between the components of the score, and I think this paid the most dividends — something I think every applied machine learning professional out there should keep at the top of their priority list. I also spent significant time on harnessing previous work and on making sure the data set was as clean as could be, two activities that can save a lot of time down the road and boost the benefit of every other implementation choice. I think one of the biggest wins there was identifying early a PNG stream that gave away a photo being from Imagenet, which in turn was leveraged for a training set for separating all of them from faces. Finally, it’s important to not leak train data into test or validation .This can sneak up on you in surprising ways, especially with all the chaos of hacking a competition solution late into the night.

What was the most challenging aspect of the competition that took extra consideration or effort?

The hardest challenge by far was the monk scale prediction. The fact that being off by one was considered as bad as being completely off had thrown my models off track all too often. And it didn’t help that to my eyes the difference between two neighboring classes was imperceptible. I think I was able to offset this difficulty a bit by training regressors and then applying thresholds on their output, but for sure this is the area where I felt the challenge was greatest.

What advice do you have for aspiring Bias Buccaneers? Get out there, sign up, and join the competition! Nothing is more educational than tackling a real-life challenge on your own, learning from experience, and painfully optimizing repeatedly. I also think it’s often useful to stand on the shoulders of giants wherever possible, so go ahead, pick out existing tools for each of the components of your future solution, and only then measure the impact of each one and iterate on that part where the potential benefit is the greatest. And remember: To Err is Human, To Arr is Pirate! 🏴‍☠️ ____________________________________________

Justin Chung Clark does data science work at the Berkman Klein Center for Internet & Society at Harvard University. Justin is interested in ways machine learning can improve well-being (and not just concentrate wealth/power). Bias Buccaneers: What activity or approach did you take that you think made your approach stand out? Justin Chung Clark: Out-of-distribution detection was probably important.

Bias Buccaneers: What was the most challenging aspect of the competition that took extra consideration or effort?

Justin Chung Clark: The most difficult part was coming to terms with the nature of the challenge itself. Generally, I don't think algorithms that can sort people by skin tone, age, or gender are a positive thing to have in society. Finding a little clarity around when I feel these sorts of algorithms might be necessary, and how they might be used while minimizing dual-use concerns took extra effort. What advice do you have for aspiring Bias Buccaneers? It's important to get a good feel for all the technical tools and pieces, but don't get hyper-focused on the technical aspects. They are less important than the social aspects of these projects.

____________________________________________

Outstanding Contribution

The judges were impressed with the incredibly low inference time and data-centric nature of our winner for most innovative model, and while it did not match the performance of the winning models, we wanted to provide a special commendation for this achievement. OII Crew (Andrew, Jonathan, Lujain) are Master's and PhD students in the Social Data Science program at the Oxford Internet Institute. Our backgrounds are in computer science, mathematics, and cognitive science. Our research interests vary and include human-computer interaction, machine learning, and the societal implications of AI. What activity or approach did you take that you think made your approach stand out? OII Crew: Our unsupervised approach does a nearest neighbor search on images generated by DALL-E. This gives us more control over the intersectional makeup of the data, without relying on labelled examples. It also gives us the opportunity to leverage the increasingly impressive abilities of generative models, though they come with their own set of bias difficulties.

Our supervised approach uses two main ideas: 1) bad data gives bad models, and 2) make sure your training objective matches the behavior you want in the final model. Based on these ideas, we focused on cleaning the dataset to only include real faces, and balancing through augmentation to help reduce class disparities without overfitting. We also modified the training algorithm to incorporate the final objective function rather than just cross-entropy loss in pursuit of the second goal. What was the most challenging aspect of the competition that took extra consideration or effort? Learning about and working with the data! What advice do you have for aspiring Bias Buccaneers? Understanding the data you're dealing with is as important as understanding the model you are building. ____________________________________________

Thank You to All Our Participants!