Model AI Assignments 2018

Todd W. Neller, Zack Butler, Nate Derbinsky, Heidi Furey, Fred Martin, Michael Guerzhoy, Ariel Anders, Joshua EckrothAAAI 2018 · DOI

TL;DR

This paper presents a collection of seven "Model AI Assignments" designed to be easily adopted and engaging for undergraduate AI courses. These assignments aim to provide students with hands-on experience in core AI concepts like pathfinding, search, neural networks, and ethical considerations, while also being flexible for diverse teaching needs.

What Problem Does This Paper Address?

The authors address the challenge of effectively teaching Artificial Intelligence by providing readily implementable and engaging project assignments. While many AI courses cover theoretical concepts, there's a need for practical, well-defined projects that allow students to apply these concepts, experiment with different techniques, and develop a deeper understanding of AI's complexities and trade-offs. The paper aims to bridge this gap by curating and presenting a set of high-quality, reusable assignments that can be integrated into existing AI curricula.

Methods

The paper doesn't present a single novel algorithm or method. Instead, it showcases seven distinct AI assignments, each focusing on different areas of AI:

  • Go For a Walk! Pedestrian-Friendly A* with Learned Cost Functions: This assignment uses real-world data (OpenStreetMap, digital elevation data) to teach A* path planning. Students learn to derive cost functions and heuristics, and then use collected data to perform regression and learn new cost functions, comparing individual vs. class data.
  • Solve a Maze via Search: This project guides students in developing an image-based maze solver. It involves formulating maze-solving as a search problem and using basic computer vision techniques (OpenCV) to convert images into a solvable representation, projecting the solution path back onto the image.
  • Understanding How Neural Networks Recognize Faces: Students build and visualize feedforward neural networks for face recognition using TensorFlow, exploring concepts like template matching and overfitting through weight visualization.
  • Understanding How Recurrent Neural Networks Model Text: This assignment introduces students to Recurrent Neural Networks (RNNs) by having them generate "fake" English text from a trained RNN model (on Shakespeare) and explain the model's behavior based on its weights.
  • A Module on Ethical Thinking about Autonomous Vehicles in an AI Course: A collaborative effort between computer science and philosophy faculty, this module integrates ethics and utilitarianism into an AI course, using the "Trolley Problem" to engage students in thinking about the ethical implications of AI, particularly autonomous vehicles.
  • Robot Juggling: Students program a velocity-controlled robot in a 2D physics simulator to juggle a ball, implementing a hybrid controller with a mirror control law. This assignment focuses on translating theoretical control concepts into real-world practice, highlighting the challenges of nonlinear dynamics.
  • Biductive Computing: Several Variants of a Universal Paradigm: This assignment introduces students to logic programming (Prolog, ProbLog) through "biductive computing," which supports both deductive and abductive inference. It offers four variants: database querying, planning, parsing, and probabilistic reasoning.

Datasets

The assignments utilize a variety of data sources:

  • OpenStreetMap and digital elevation data: Used in the "Go For a Walk!" assignment for path planning.
  • Real-world collected pedestrian data: Students collect this data in the "Go For a Walk!" assignment to train cost functions.
  • Input images of mazes: Used in the "Solve a Maze via Search" assignment.
  • Face recognition datasets: Implied for training neural networks in the "Understanding How Neural Networks Recognize Faces" assignment.
  • A corpus of Shakespeare plays: Used to train the RNN model for text generation in "Understanding How Recurrent Neural Networks Model Text."
  • 2D physics simulator: Used in the "Robot Juggling" assignment, simulating the dynamics of juggling.

No new benchmarks are introduced; rather, existing real-world data and simulated environments are leveraged.

Key Insights and Ideas

  • The power of hands-on learning: The assignments emphasize that practical application is crucial for understanding complex AI concepts.
  • Flexibility in AI education: The collection demonstrates that a variety of AI topics can be taught effectively through well-designed, adaptable projects.
  • Bridging theory and practice: Assignments like "Go For a Walk!" and "Robot Juggling" explicitly connect theoretical knowledge to real-world data and physical systems.
  • The importance of visualization: Assignments like the face recognition and maze solver projects highlight how visualizing intermediate steps and model internals aids understanding and debugging.
  • Integrating ethics into AI: The module on autonomous vehicle ethics shows a successful model for incorporating crucial societal considerations into technical AI courses.
  • The value of reusable resources: By sharing these assignments, the authors aim to reduce the burden on instructors and promote best practices in AI education.

Key Results

The paper focuses on the pedagogical value and adoptability of the assignments rather than quantitative performance metrics. However, it does report on the effectiveness of the ethics module:

  • Ethics Module Effectiveness: Nearly all students (62 out of 63) in the course understood the "Trolley Problem" after the one-week module.

The other assignments are presented as having specifications and supporting resources available online, implying their successful use in teaching.

Limitations and Future Work

  • Adoptability: While designed for easy adoption, the success of these assignments will depend on instructors' ability to integrate them into their specific course structures and available resources.
  • Scalability: The paper doesn't explicitly discuss the scalability of these assignments for very large class sizes, though the "Go For a Walk!" assignment mentions class-wide data sharing as a feature.
  • Further Development: The authors provide links to assignment specifications and resources, suggesting that these are living documents that may be updated and expanded upon. The "Biductive Computing" assignment explicitly mentions suggested enhancements for students. The "Go For a Walk!" assignment also lists extensions for bicycles and cars.

Summary updated

BibTeX
@inproceedings{neller2018model,
  author = {Todd W. Neller and Zack Butler and Nate Derbinsky and Heidi Furey and Fred Martin and Michael Guerzhoy and Ariel Anders and Joshua Eckroth},
  title = {{Model AI Assignments 2018}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.1609/aaai.v32i1.11398},
}
Figure