M(AI)ND READER

Inspiration

While the availability of massive data sets has made it easy for artificial intelligence (AI) to derive new insights, AI systems are powered by pattern recognition and classification, which broadly speaking often drive exclusionary and discriminatory social processes. In the case of google photos, for example, a user was offended when an image of two black friends was tagged as “GORILLAS.” Lack of diversity in AI industry has reached ‘a moment of reckoning’ over these biased systems where we must consider the implications this technology may cause as a consequence of its use.

One way to help mitigate negative consequences is to allow people to provide feedback so that machines can keep adjusting their view of the world. In fact, human-in-the-loop is a branch of artificial intelligence that leverages both human and machine intelligence to create machine learning models. 

But how can we make it scalable, in a way like how Google Image Labeler game allowed users to label images to improve Google's image search quality?

What it does

This project’s goal is to allow you capability to report any offensive images that might arise, whether you think the image was classified wrongly by ml5 via its dataset of 15 million images or whether the associated words to AI's guess are off color. This game was a proof of concept for engaging users to help train datasets by alerting creators and API maintainers of these pitfalls.

How we built it

Vanilla JavaScript, APIs, and ML5 for image classifier!

Challenges we ran into

Steep learning curve - interfacing with different APIs and creating a game and really learning about promises and asynchronous functions!

Accomplishments that we're proud of

I made a game with an awesome and noble concept behind why it was created.

What we learned

A lot about promises and asynchronous functions! And it's my first time working with ml5's image classifier!

What's next for M(AI)ND READER

Give it ability to inform API creators/maintainers of concerns with offensive image and word associations.

Built With

• css
• datamuse
• html
• javascript
• lol
• ml5