- **phd classes should orient you to the near future** - ml and behav.econ will not exist without the other ### why be can't live wo ml - role of algorithm (systematically produce variation); algorithm:= a process or set of rules to be followed in calculations or other problem-solving operations, especially by a computer - in the next 10 yrs, much of hiring will be increasingly algorithmic - changing ppl is hard, but code is easy to rewrite (this is an opportunity) - building algorithm to eliminate bias - problem is human bias and solution is algorithm - understanding of behavior **behind** race gap (need pre-work that points to the problem) - > insight behind right algorithm - feasible and impactful paper is very scarce (a thinks tool-based paper can overcome this; JSS) - decisions are being increasingly algorithim (playbook: find consequential distortion that's hard to fix, decisions are increasinlgy algorithmic - suggest algorithm to improve distortion) ### why ml can't live wo be - set of automation tasks (images, label) dataset, predict label from image - adoption by clinician - distribution shift issues (urban, country) - doctor's label (labels are gold stnadard, but for human quality degrade over time) - algorithm is being trained on cnosensus of the field (jury, biased to not see (enemy is coming in the dessert), where's waldo) - need model of physician judgment - noise iid = universal south of ignorance (1. ; 2.don't believe in prob.) - enormous psychology in judgement (error is not noise, but systematic object) - must make assujmpion about meaning of that data, be models give structure to those assumptions - disentangle expertise from bias and noise - adjudicated vs clinical settings - automate as they are but "at their best" - "breaktrhough on how to incorporate formal be model with supervised learning" is hard; adjudicated jus - clinical decision making is a large field in psychology - smell test is funny as when sendil did the resume, employers predicted the opposite (need to show they hire black) ### fundus and resume commons: - outsized positivity: improve hiring and diagnostics - algorithm is qualitatively different but solves biggest problems in this space: scale - train data include human's behavior fingerprint (learning from human to outperform them) - resume example is for algorithm - nudges (default setting) is specific, algorithms is more on choice architecture - decide where to turn the ray bug, how to build it in first place so far was normative (establish hypotheses that can be empirically tested; understanding should be behind algorithm) positive (based on opinion or subjective values) - how can algorithms help us with understanding