/
lectern.tv
Play15s30s1x0:00/--:--
DoneWatched
Expand
lectern.tv
2:45/1:13:25
0:57
2:49
3:31
1:41
1:16
3:12
3:55
2:05
1:31
1:15
1:39
4:46
1:44
2:01
2:59
0:57
1:03
5:46
0:54
3:16
1:17
1:05
2:08
5:30
3:33
1:23
2:47
2:44
2:58
Just Machine Learning
Investigates the societal implications of algorithmically infused societies, focusing on generative AI. It addresses issues like epistemic instability, the challenge of oversight, and accountability for rapidly evolving AI technologies.
Lecture guide · Written by AI from the lecture. For the lecture’s own words, open the Transcript tab.
Tina Eliassi-Rad explores the societal implications and life cycle of machine learning systems, examining issues such as proxy variables, mathematical impossibility results in risk assessment, data bias, and the complex feedback loops between algorithms and society. The lecture concludes by analyzing generative AI, large language models, epistemic instability, and the challenges of governing these technologies within complex democratic systems.
This lecture assumes
- Large Language Modelstaught in Artificial Intelligence at 48:50
- Transformer Architecturetaught in Artificial Intelligence at 49:39
- Neural Networkstaught in Artificial Intelligence at 43:38
Key Takeaways
- Machine learning systems often learn proxies or unintended patterns (such as weather instead of camouflage) rather than the actual task defined.
- Impossibility results prove that risk assessment algorithms cannot simultaneously satisfy equal precision, equal detection probability, and equal false alarm rates under unequal base rates.
- Large language models act as counterfeit people that simulate intimacy, increasing epistemic instability and threatening shared social reality.
- Democracy functions as a complex system, and algorithmically amplified misinformation can drive societal systems toward instability and backsliding.
Concepts Explained18
Machine Learning—The field of study that gives computers the ability to learn without being explicitly programmed.
Proxy in Machine Learning—A substitute variable used to predict something that is hard to predict directly.
Risk Assessment in Machine Learning—A popular prediction task that assigns a score, such as zero to ten, representing the risk of an event like loan default.
Impossibility Results for Risk Assessment—Mathematical results showing that under certain conditions like unequal base rates and regular priors, you cannot simultaneously satisfy equal precision, equal probability of detection, and equal probability of false alarm.
Examples, Proofs & Derivations
Works & Sources Cited23
Book
What Artificial Experts Can and Cannot Do—Hubert L. Dreyfus and Stuart E. Dreyfus, published in AI & Society (1992)
Paper
Dissociating racial bias in an algorithm used to manage health populations—Ziad Obermeyer et al., Science (2019)
Paper
Fair prediction with disparate impact: A study of bias in recidivism prediction instruments—A. Chouldechova, Big Data (2017)
Paper
Inherent trade-offs in the fair determination of risk scores—J. Kleinberg, S. Mullainathan, M. Raghavan, Innovations in Theoretical Computer Science (2017)
Paper
Fairness Through Awareness—Dwork et al. [2012], introducing the definition of individual fairness.