Use your human brain to avoid artificial intelligence disasters

Earlier this year, I helped mentor a local edition of fast.ai’s Practical Deep Learning for Coders. Each mentor gave a brief talk on a given week’s subject, adding to the material covered in the recorded lectures. My talk (embedded below) supplemented the data ethics lesson. While the mere mention of the word ethics can elicit instant yawns from some people, the main message for me is that it’s critical for humans to think about the context and consequences of deploying machine learning models.

Unfortunately, this message sometimes gets muddied amidst the outrage about specific applications that conflict with the values of the outraged parties. But I believe it’s possible to transcend narrow moralities and agree that better outcomes arise when humans think deeply about their deep learning systems. Or to put it more bluntly, any fool can build machine learning models, but it takes thoughtful humans to build good artificial intelligence applications.

Of course, what constitutes good is an open question, which I touched on in the talk. Other key points include:

• The modelling context is much broader than any machine learning model. Considering context is where human brains shine.
• Thoughtlessness can have a negative impact on society and on your career.
• Moral values vary across time, space, cultures, and individuals, e.g., along five moral foundations.
• Any data scientist, machine learning engineer, or modern human should develop their critical thinking skills. The Calling Bullshit course from the University of Washington is a great starting point – essentially Data Literacy 101.
• Bullshit is easier to detect than call. Deciding on a level of bullshit calling is like tuning a model’s learning rate.

A good chunk of the talk was spent on the case study on criminal machine learning from the Calling Bullshit website. I was pleased with the level of engagement on this segment, especially since a lockdown forced us to deliver the class online at short notice. You can watch the full talk below (my part ends after 24 minutes), view the slides here, and check out supplementary materials from all mentors on GitHub.