6. Brain-Behavior Predictive Modeling: My Journey
Value-Based Machine Learning | Intro | Accuracy | Utility | What’s Next
• Data Scientist @ UMich
• Predicting BrainAge or Cognition
2017 2023
2022
2021
2020
2019
2018
Whistler: Whistler:
“Value-Based
Machine Learning”
• Ph.D. candidate @ Donders Institute
• Big Data Normative Modeling +
Transfer Learning to Clinical Datasets
Whistler:
“Developmental Mega Sample”
7. Brain-Behavior Predictive Modeling: Current Status
Combine a bunch
of data from
open datasets Fit a bunch of different
algorithms, ranging
from simple to super
complicated
Realize there is little
overlap in available
phenotypes across these
datasets. You are left with
age, sex, maybe cognition
(if you’re lucky).
Realize that there isn’t a
lot of signal in the data,
and that you can’t even
predict age that well
(maybe within ~3-5years)
Publish your results anyways….
a) being super optimistic and
slightly overselling the
interpretation and potential.
b) sharing your honest viewpoint
(using MRI doesn’t help much).
Have trouble finding a journal that
will publish this perspective.
a) repeat
b) leave for a data science industry
job or another field “where ML can
have more impact”
Value-Based Machine Learning | Intro | Accuracy | Utility | What’s Next
8. Brain-Behavior Predictive Modeling: Bingo Card
Fluid
Intelligence
Brain Age
Poor
Reliability
Reference to
Marek et al.
Nature Paper
No confound
correction...
“could be
motion”
HCP /
ABCD /
UKBiobank
r = 0.28
“Has clinical
potential
(one day)”
“We need a
bigger
sample size”
Value-Based Machine Learning | Intro | Accuracy | Utility | What’s Next
14. • The quest for high performance.
• A narrow objective of becoming more
accurate, and an immediate (short term)
action plan for how to achieve this goal
(minimize the loss function on a
particular set of data).
Definition: Accuracy
Value-Based Machine Learning | Intro | Accuracy | Utility | What’s Next
15. • Consider utility to be more closely aligned with
the model’s purpose (i.e., answering the research
question and adding real world value).
• Utility looks at the bigger picture and makes
creative adjustments to align with the ultimate
research goal and real-world application.
Definition: Utility
Value-Based Machine Learning | Intro | Accuracy | Utility | What’s Next
16. • In the process of setting up the optimization problem we convinced ourselves
that it makes sense to optimize for accuracy because it is more easily
mathematically formulated than utility…
• … But if you zoom out to look at the bigger picture you realize the goal of the
A.I. field is to do useful stuff that makes life easier for humans, not to create
intelligence (become more accurate).
Bringing Together Optimization, Accuracy, & Utility
Value-Based Machine Learning | Intro | Accuracy | Utility | What’s Next
17. Measuring Accuracy
Value-Based Machine Learning | Intro | Accuracy | Utility | What’s Next
• A single standard metric that represents the models’ ability to predict
observations in the test set.
18. Optimization for Accuracy
• A specific loss function is used to improve model during training/validation. Often
same metric is used to evaluate “out of sample” performance in test set.
Value-Based Machine Learning | Intro | Accuracy | Utility | What’s Next
19. Optimization for Accuracy
• Benchmarking A comparison of model performance to another model.
• “Best” model is determined by being more accurate than the others.
Value-Based Machine Learning | Intro | Accuracy | Utility | What’s Next
20. • There is no consensus regarding the definition of success in creating artificial
intelligence meaning there is no finish line or upper limit on attaining A.I.
• Without a clear definition of goals and a vision of what success looks like,
how will we know when we have reached the goal?
• What does it mean to become infinitely more intelligent?
• What purpose does it serve to have a world full of agents (machines or humans)
that are super intelligent?
• Goodhart’s Law: “When a measure becomes a target, it ceases to be a good metric.”
Limitations of Accuracy
Value-Based Machine Learning | Intro | Accuracy | Utility | What’s Next
21. (Abstract) Limitations of Accuracy
• Soccer team example…. the star player who only thinks about themselves
(accuracy) vs. the team captain who puts the team first (utility).
Value-Based Machine Learning | Intro | Accuracy | Utility | What’s Next
Jamie Tart
vs.
Roy Kent
22. (Concrete) Limitations of Accuracy
• High accuracy does not imply:
• reproducibility
• meaningfulness (that the features used are better than random)
• does not come with explainability
• equal accuracy doesn’t imply that two models have learned in the same way
Value-Based Machine Learning | Intro | Accuracy | Utility | What’s Next
27. Value-Based Machine Learning | Intro | Accuracy | Utility | What’s Next
Predicting patient pain score instead of radiologist’s dx
Optimization for Utility (fairness is priority value)
• Use knee X-rays to predict patients’ self-reported experienced pain,
instead of using standard measures of pain severity (radiologist dx).
• Relative to radiologist dx, which accounted for only 9% of racial
disparities in pain, using self reported pain labels accounted for 43%
of racial disparities in pain (4.7× more than radiologist dx).
28. Value-Based Machine Learning | Intro | Accuracy | Utility | What’s Next
• Equal opportunity & Multi-objective optimization
Fairness
Accuracy
Optimization for Utility (fairness is priority value)
29. Value-Based Machine Learning | Intro | Accuracy | Utility | What’s Next
Optimization for Utility (efficiency/useability is priority value)
• Optimizing for teamwork, AI learning to complement humans.
https://pcnportal.dccn.nl/
• Sharing pre-trained models & creating accessible tools
30. Value-Based Machine Learning | Intro | Accuracy | Utility | What’s Next
Benefits of Utility
• Collaborative, efficient, well-defined purpose.
• Functional (real depth and meaning) rather than attractive (shallow,
surface-level appeal).
• An opportunity to think deeply and align your models with your purpose.
• Creative thinking and problem solving is required.
• More of a challenge… thus more satisfying solutions will be created.
31. Value-Based Machine Learning | Intro | Accuracy | Utility | What’s Next
Roadblocks
• Cognitive biases make us focus on simpler problems.
• As problem complexity increases, we shift responsibility, and think along the
lines of “this is out of my expertise, it is someone else’s problem to solve”.
Ambiguity Effect Bandwagon Effect Status Quo Bias
Loss Aversion
32. Value-Based Machine Learning | Intro | Accuracy | Utility | What’s Next
Roadblocks
During
Development
In the Wild
(Real world)
Stationary Data
Single Decision Maker
Complex, Non-stationary Data
Many Stakeholders
34. Value-Based Machine Learning | Intro | Accuracy | Utility | What’s Next
• Many other fields have defined utility and figured out how to optimize for it.
• Let’s learn from them.
Future Directions
Human-Computer Interaction (HCI)
Ethical A.I.
Value-based healthcare
Behavioral Economics
35. Value-Based Machine Learning | Intro | Accuracy | Utility | What’s Next
• Utility (and value priorities) will always depend on the context.
• We need open communication and guidelines about making these decisions.
Future Directions
36. Value-Based Machine Learning | Intro | Accuracy | Utility | What’s Next
• Too much (tunnel vision) focus on accuracy of predictive models.
• We have lost track of our “why” and this has created a lack of model utility.
• It should be a priority to define our values which will help us build a better plan
for moving towards these goals and values.
• Optimizing for utility is an abstract and creative practice that requires diverse
perspectives and input. It should be an on-going process.
Take Home Messages
37. Charlie-Mop
Acknowledgments
Value-Based Machine Learning
Chandra Sripada, Mike Angstadt, Daniel Kessler, Liza
Levina, Ivy Tso, Alex Weigard, Jenna Wiens
Ph.D. supervisors: Andre Marquand, Eric Ruhé, &
Christian Beckmann.
Lab members: Seyed Mostafa Kia, Thomas Wolfers,
Mariam Zabihi, Charlotte Fraza, Pieter Barkema, Stijn
de Boers, Barbora Rehák Bučková
Donders Institute, Nijmegen University of Michigan, Ann Arbor
Historically, A.I. as a field has overpromised solutions and underperformed on bringing scientific advancements into the real world.
Shifting priorities to focus on utility over intelligence will help make the goals of A.I./ML more explicit + actionable and thus will improve scientific communication through creating more realistic public expectations and building trust.
During an optimization step, the model parameters are iteratively updated such that the loss function (i.e., mean squared error) is minimized within the training data set.
To summarize, when setting up the optimization step of a machine learning model, we are deciding what is right and what is wrong.
For example, maintaining high accuracy while simultaneously using less computational resources which saves money and reduces carbon emissions.
An opportunity to reframe our research questions to better align with our true purpose and vision.
An extreme simplification of a model’s performance and traits.
Does not capture reliability, validity, complexity, fairness, useability, etc.
The goal is to achieve the highest accuracy, lowest mean squared/absolute error, highest correlation between predicted and observed.
Continuous version of the binary winner and loser, followed by a ranking-based comparison.
Contributes to the “replication is all we need” attitude, and a lack of thinking about true innovation.
We propose a simple, interpretable, and actionable framework for measuring and removing discrimination based on protected attributes. We argue that, unlike demographic parity, our framework provides a meaningful measure of discrimination, while demonstrating in theory and experiment that we also achieve much higher utility.
Usefulness means saving time/energy/costs/resources
Making utility explicit (so that we can mathematically model it) is more challenging than mathematically modeling accuracy/performance.
We favor simple-looking options and complete information over complex, ambiguous options
We’d rather do the quick, simple thing than the important complicated thing, even if the important complicated thing is ultimately a better use of time and energy.
In practice, there is often a single decision maker (ML developer), and the underlying population is assumed to be stationary.
This is not true in the wild (real world setting) where there are a lot of people involved, each with different value priority queue, and the data is of course incomplete and very messy.
Fairness and accuracy are often assumed to be in opposition, meaning there is a trade-off when optimizing for one over the other (i.e., optimizing for more predictive fairness leads to less accurate predictions or optimizing for accuracy results in less fairness).