How to teach Machine Learning to empower learners to speak up for themselves
Yim, what have you been working on for the past 2 years?
Thanks for asking!
It’s this paper that just got accepted at the ACM International Computing Education Research (ICER) conference! Here’s a pre-print of the paper: Learning Machine Learning with Personal Data Helps Stakeholders Ground Advocacy Arguments in Model Mechanics (Yim Register & Amy J. Ko, 2020) Gotta love academic titles am I right? But don’t worry, I’ve saved you the trouble of reading the paper by summarizing it in this blog post. Because A) honestly, it is long and fancy and a bit jargony and B) because how else would you get to read all my jokes?
Additionally, you could watch my presentation here. I “doodled my conference talk”, because I wanted to do something creative and fun to keep me sane right now.
Machine Learning is all around us
Our lives are shaped by machine learning systems. Google Maps learns which routes you tend to take, Facebook suggests ads and friends to you based on your history, and ALEXA uses speech recognition to understand what you scream at her from the other room. Computer vision (a subcategory of machine learning) is used to detect your face for a silly filter or to detect the presence of tumors on a medical scan. Have you ever played with text prediction and seen what your phone thinks you’re gonna say next? That’s machine learning too. ML is all around us.
But many of us move through the world without a great understanding of how these systems work, where the data comes from or where it’s going, or how these systems can fail. The ways that we traditionally teach ML are not set up to help us empower learners to answer these questions.
We often teach on generic or irrelevant data sets, present complicated equations, or rely on opaque programming packages that hide a lot of what is actually happening in the system. It is not the norm to critically evaluate the design and applications of ML systems, even in the ML classroom.
Imagine being shown this, and then being asked to point out problems with ML in the real world:
That’s gonna be a hard Nope! from me, and I’m the Machine Learning person! Pardon me while my eyes glaze over!
So what should we do instead?
Involve the learner’s culture, identity, and personal experience
Learning sciences, critical pedagogy, computing education research, psychology, social science, and cognitive science all point us towards the idea that involving the learner’s relevant experiences can benefit their learning. What if we took that to the machine learning playing field? That means allowing the learner to represent themselves, explore their own data, and point out how models might fail for their specific circumstances. Teach the models in context, on real data, while giving the gift of discovery to the learner, and maybe you’re on to something…
Empowered voices help shape choices
If we succeed in teaching ML in a personalized and contextual way, could we shape how we interact with ML in the world? Keep in mind that people are coming up with new ML technology all the time; a new app for this a new app for that. Companies adjust their algorithms constantly, and new “Big Data Solutions” are solving the world’s toughest problems, right? But one problem with that is the disconnect between the ML-savvy engineers/designers/researchers and the stakeholders on the other end. The entire point of involving the learner in ML education is to involve them in ML real-world scenarios. We can empower one another to speak up, ask the “right” questions, demand better from technology companies, and advocate for ourselves and others when we think a model has harmed us in some way.
Some questions an empowered user might ask about ML models in the world:
- What data and online behavior led to me only seeing photos of white and thin people in my recommended ads?
- How do models used in law enforcement contribute to racist judicial decisions?
- When I “Like” something on Facebook, how does that affect what news articles I will see in my newsfeed?
- Do my friends see the same things that I do when they search the Internet for information? How about my parents?
- What are engineers doing to ensure that ALEXA can understand my accent or way of speaking? (alternatively, what can I do to make sure I keep my privacy?)
- Do your machine learning models encode gender as binary?
The Experiment
- Build something that automatically teaches a Machine Learning concept using personal data
- Ask novice (pre-screened) learners to write self-advocacy arguments against a model that has made a mistake that personally affects them
- Test if learners using their personal data behave differently than those who didn’t (e.g. learned more, wrote more, mentioned more concepts from the tutorial, talked about themselves more, asked more critical questions, pointed out problems that others didn’t)
I created a tool called LearnMyData that allowed students to predict what grade they might get in a college course based on how interested they were in the material (kind of meta, isn’t it?) Learners input the grades they got for their past 5 courses and how interested they were in those courses (1–7 scale), and then looked at the resulting linear regression model (prediction line) on their own data. (Fun fact! For me personally, the more interested in a course I am, the worse I tend to do! I think it’s because I’m most interested in classes that are really hard for me).
There were 3 conditions:
Facts: Learners never saw the LearnMyData tool and instead received a brochure-style fact sheet about how linear regression works. This mimicked a kind of textbook or lecture instruction without a walkthrough with actual data.
Impersonal: Learners saw the LearnMyData tool but did not input their own data. Instead, they learned about a hypothetical student who had increasing grades with increasing interest.
Personal: Learners input their own grades and interest ratings for their past 5 courses, and walked through the tutorial looking at their own linear regression line.
Next, participants saw two scenarios. The first one was using the Interest-to-Grade model.
“The instructor of a college course uses a model to identify which students might need extra support and help throughout the quarter. The instructor uses the linear model that you saw in the tutorial. At the beginning of the course, the instructor collects everyone’s Interest Level, and makes a prediction for their Grade, based on last year’s Interest-to-Grade data. If the model predicts a grade lower than an 75, the instructor will intervene and offer extra help. So, if a student who rates their interest at a 2 tends to score below a 75, the instructor will intervene with a new student who rates their interest at a 2. List as many critiques of this model as you can. Try to use what you learned in the tutorial to make your case. Next, write a convincing argument of how you might advocate for yourself as a student in this scenario (someone being affected by the model). Imagine you are making your case to the instructor, or someone else enforcing the results of the model.”
The second was a Financial Aid modeling example.
“The financial aid office has to make tons of decisions in order to give out aid. Usually, they offer an amount and won’t change it unless a family “appeals” the process because it is not enough. They try to predict how much aid they should give as accurately as they can, so that they offer an amount that a family won’t appeal. They use a model that uses the number of siblings that a student has to predict how much money their family will need. They use last year’s data, looking at families who were “happy” (did not appeal) with the offer the office gave. So, if families with 3 children tend to need $20,000 in aid, that’s what the office will budget for a new family with 3 children. Next, write a convincing argument of how you might advocate for yourself as a student or family in this scenario (someone being affected by the model). Imagine you are making your case to the financial aid office, or someone else enforcing the results of the model.”
We had 51 participants, randomly assigned to one of the three conditions. 37 of them had actually taken an introductory programming course, but weren’t familiar with data science. The authors inductively coded all of the self-advocacy arguments (without knowing the condition of the participant), and discovered emerging “codes” or topics that appeared in the writing. Next, they went through and identified which of those codes were present in the self-advocacy arguments (again, not knowing the condition of the participant they were looking at). In the end, you can count up the codes after uncovering which condition they were in, and discover any differences between the conditions. Because the codes and conditions are “categorical” we can perform Kruskal Wallis chi-squared tests to detect differences, and then evaluate effect size with Cramer’s V. This is a very simplified version of the methods and analysis! You can read more in the paper 😉.
Learning on Personal Data helps self-advocacy arguments to talk about the actual mechanisms of ML models
We discovered a number of important codes that appeared in participant’s writing. They talked about all kinds of things! What we wanted to see was evidence that they were learning how the machine learning was actually working. Some things they mentioned in their arguments that showed evidence for this was mentioning things like:
Causality: “correlation is not causation”
Confounds: “what about students who already know the material?”
Construct Validity: “interest level doesn’t necessarily mean a student’s ability”
Additional Features: “what if you included motivation also?”
Model Performance: “the fit will be really spread out and make wrong predictions”
Outliers: “generalization from past students doesn’t take me into account as a potential outlier”
Additionally, other things showed up like “you could game the system by pretending you weren’t interested in order to get extra help”.
So you count up what everyone mentioned and then take a look at it split up by condition. Learners in the Personal condition used more of the mechanisms of ML in their self-advocacy arguments, specifically mentioning other Features the models could use and commenting on Model Performance. They also mentioned how learners could “Game the System” more than the other conditions (not shown), and were more likely to say the model scenario was actually “good” but with some flaws. Goodley’s Self-Advocacy, “Learning Difficulties,” and the Social Model of Disability and Empowerment, self-advocacy, and resilience outline some effective self-advocacy strategies such as negotiation and articulation of the problems the advocates are facing. The learners in the Personal condition were able to effectively communicate critiques of the model scenarios and provide alternative solutions in their advocacy arguments.
What this study is missing and next steps
This is just the start. Given this one study in this one context on this one algorithm, we don’t know a ton. But this work outlines the rest of my PhD career: to discover strategies for teaching machine learning that empower people to advocate for themselves in the world when these models make mistakes. Lots of prior hard work guides these ideas: AI literacy, Feminist Data Science, Self-Advocacy for Disability, Critical Race Theory, Computing Education, Machine Learning Education, Personalized Learning, Constructivist and Experiential Learning, Gender Studies, Participatory Modeling, and more.
The project I’m now working on takes this work into real-world “dirty” data: Facebook recommender systems. I’ve developed a system to teach a recommendation algorithm called “collaborative filtering” on people’s own Facebook ad recommendations. Hopefully through working with your own real-world data, we can uncover pathways for self-advocacy on a larger scale.
Next steps are a magical world of possibility! The world is scary and hard and there’s lots to advocate about, especially within a domain like machine learning where it’s hard to know what is going on. My mission is to make it easier to speak up for yourself within complex systems that affect you and your experience in the world.
Takeaways for YOU, whoever you are
Whoever you are, you have a place in shaping an ML-literate society. You are an important and vital part of a movement. Researchers, social media influencers, parents, businessowners, data scientists, teachers: we all interact with machine learning and we all have a voice to keep questioning and critiquing the technology in the world. The big takeaway from this work is that by involving your own experiences, culture, and backgrounds in the conversation about machine learning, you can bring new insights and understanding. Don’t be afraid to ask questions, critique data sources, and advocate for yourself whenever you can! Your voice matters more than you might realize.
My current creative (not academic!) project on these topics is my book Life Lessons from Algorithms, where I discuss machine learning algorithms through a PTSD recovery lens. You can find updates and excerpts on Instagram here.
