We are working together at Colby to analyze AI applications and develop a “tool users’” understanding of AI. Follow along with our January 2022 course via our syllabus and calendar here: AI: What Lies Beneath.
Day 1: What is AI and why do we care?
First year, US-based college students in 2021:
- were in elementary school when Siri came out, and when Facebook first released face recognition
- were toddlers when Google machine translation came out
Day 2: Our values
Members of “generation AI” assume that ML solutions are continually learning; to them, the idea of lifelong reinforcement learning is intuitive, and they are disappointed to learn that many ML models are never trained again once deployed. They are also, not coincidentally, the first generation to be surveilled by tech almost from birth.
Day 3: Bias, fairness and humans in AI
Members of “generation AI” may be less interested in using obvious AI — chatbots, machine translation — but are still subject to hidden AI.
Days 4–5: Just what is in that AI application?
We are only able to do this course this way because of open source and open science. Without these things, we would not be able to get “inside” an AI application to understand it deeply. On the other hand, open source and open science have led to increased risks, including — the same week we discussed open source — stories like this.
Day 6: Running and testing AI applications
Always look at your data.
Day 7: What is in the data?
“The absence of critical engagement with canonical datasets disproportionately negatively impacts women, racial and ethnic minorities, and vulnerable individuals and communities at the margins of society” — Prabhu and Birhane, Large datasets: A pyrrhic win for computer vision?
Day 8: How can we attack it?
Creativity is one of the hardest skills.
Day 9: Guest speakers!
From one of the class participants: “This [most crash test dummies being modeled after men] is an interesting issue that would not have occurred to me; it presumably did not arise through intentional discrimination, but shows how easily we can unknowing create systems that favor certain groups over others, even when intentions (such as safety) are pure.”
Day 10: Working together
In the workplace, collaboration skills are critical to success.
Day 11: And the winners are…
What astonishes me is how quickly the transformer architecture has become the state of the art architecture across many AI applications. It’s not just that this architecture gives state of the art results for one type of AI; it gives state of the art results for many types of AI, leading to an implosion of techniques over the past few years. And it’s not just that this architecture is being used, but the same models (trained on the same data) are being used across widely different AI applications.
Acknowledgements:
- The participants in this course reviewed and approved all blog posts except for day 11. They approved the publication of their presentations, application audits, model cards and data sheets through a blog post that I wrote after the final course session.
- The word clouds were created using wordart.com.