The CAPS AI Policy and Strategy (CAPS) skills course prepares international relations students to critically evaluate AI algorithms, specifically for applications that challenge conventional modes of policy and strategic decision-making. The course focuses on providing students with a deployable toolkit for assessing the trade-offs of common AI algorithms in the fields of supervised learning, unsupervised learning and reinforcement learning.
CAPS places strong emphasis on understanding how data and the environments in which an algorithm is applied influence the algorithm’s performance and the unexpected failure modes this can create. The course addresses how domain knowledge, from the regional or strategic fields taught at SAIS, can make critical contributions towards minimizing such failure modes and ensuring an optimal application of AI to complex problems. Upon completion of the course, students will have acquired a rigorous conceptual understanding of AI algorithms. Students will have gained the foundational skills needed to evaluate and make recommendations on the application of AI algorithms to the manifold challenges of today’s complex world.
The fall 2019 iteration is the fourth taught at SAIS. Materials from previous iterations can be found at https://github.com/capsseminar/.
CAPS takes place over six weeks, with one 90 minute session each Tuesday. The first session will be on Tuesday, October 29.
📧 For enrollment or questions, contact capsseminar@gmail.com
A pdf version of the syllabus can be found here (last updated, October 7, 2019).
📘 Lecture 1 - Learning and Decision Making
-
Guiding questions: How can policymakers and engineers jointly establish the most reliable, consistent paths of decision-making? How can we assess whether these are “good” paths? What are the relevant dimensions of “good” in a given strategic setting? What are the fundamental limitations of decision-making, from both human and algorithmic perspectives? What aspects can be improved? What problems do we face in instituting decision-making practices, specifically given differences in human and machine decision-making?
-
Topics covered: Types and measures of learning, approximation, complexity, efficiency, rule-based algorithms
-
Case: Analyzing patterns of learning in a counterinsurgency team
-
[Slides] | Case
📘 Lecture 2 - Supervised Learning I
-
Guiding questions: What does “learning” mean in supervised learning? How can we measure learning? What are the key data risks? How do we manage them, both through measurement and alternative techniques? What knowledge and language must policymakers and engineers share for productive collaboration?
-
Topics covered: Learning from data, logistic regression, standard regression, tree-based models, k-nearest neighbor (KNN)
-
Case: Training human data labelers in a complex organizational ecosystem
-
[Slides] | [Case]
📘 Lecture 3 - Supervised Learning II and Unsupervised Learning
-
Guiding questions: Based on our findings, what can we and can’t we learn from unsupervised learning? How can we connect supervised and unsupervised machine learning to optimize our own understanding of policy or strategy issues? Looking ahead to our next section: how does machine learning compare to human learning?
-
Topics covered: Learning about data, principle component analysis (PCA), K-means clustering, ensemble methods, differences between human and machine learning
-
Case: Deploying an autonomous supply agent to assist a human team
-
[Slides] | [Case]
📘 Lecture 4 - Reinforcement Learning
-
Guiding questions: How can machines learn from interacting with the environment? With what limitations and risks? What are the key tradeoffs involved in reinforcement learning? How do they parallel and diverge from human learning? How can human-machine teaming be leveraged – and to what ends?
-
Topics covered: Dynamic data acquisition, goals and rewards, exploration vs exploitation, online vs offline learning, model-based vs model-free learning, human-machine teaming
-
Case: Reverse engineering the agency of an unknown aerial vehicle
-
[Slides] | [Case]
📘 Lecture 5 - Neural Networks I
-
Guiding questions: How do neural networks operate? How can they be designed to implement a range of learning types? What are the requirements for neural networks to perform well, and how does this enable or constrain them in tackling strategic and policy issues?
-
Topics covered: Understanding the design of the Perceptron
-
Case: Training a simple neural net in Python and Google Colaboratory
-
[Slides] | [Case]
📘 Lecture 6 - Neural Networks II and Operational Aspects of AI
-
Guiding questions: What key data risks do we face when implementing machine learning? What algorithmic risks? As a policy or strategy professional, what tradeoffs may be most essential for your assessment of deployed neural network systems? In what scenarios might you commission a neural network versus another machine learning system? Using the foundations learned in this course, what next steps should you take as a policy or strategy professional interested in the future of machine learning?
-
Topics covered: Feature selection, hyperparameter tuning, biasvariance trade-off, transparency, interoperability, safety
-
Case: Devising levels of stakeholder access to an advanced AI-based model
-
[Slides] | [Case]
CAPS is an official SAIS skills course. Participants who attend all sessions and pass the requirements obtain formal certification for the course on their Johns Hopkins academic transcripts.
For the course to appear on one’s trascript, they must attend all sessions. If a session is missed due to illness or an emergency, the CAPS team will arrange a make-up session.
CAPS is created and taught by SAIS MAs Leo Klenner, Henry Fung, and Cory Combs.
📧 To contact us, reach out to capsseminar@gmail.com.