Towards an AI Course Based on Neural Networks

Michael WollowskiAAAI 2025 · DOI

TL;DR

This paper proposes a new approach to teaching Artificial Intelligence (AI) by grounding the curriculum in Neural Networks (NNs) and feature learning, rather than traditional symbol-based AI. The authors present a detailed course outline and learning materials, arguing this shift is crucial for understanding modern AI and preparing students for the evolving field.

What Problem Does This Paper Address?

The authors identify a disconnect between current AI curricula and the reality of modern AI development. They argue that the standard AI curriculum, often emphasizing symbol manipulation and heuristic search, lags behind the advancements driven by neural networks, pattern recognition, and feature learning. This traditional approach, they contend, fails to equip students with the foundational understanding needed to grasp the power and limitations of contemporary AI systems like LLMs. The paper aims to bridge this gap by offering a proof-of-concept AI course that prioritizes NN-centric learning.

Methods

The core of the paper is the detailed outline of a novel AI course. This course is structured around "Five Big Ideas in Artificial Intelligence" (Perception, Representation & Reasoning, Learning, Natural Interaction, Impact) but approaches them through the lens of neural networks. Key pedagogical methods include:

  • Focus on Feature Learning: Students learn how NNs automatically learn features from data, contrasting this with manual feature engineering.
  • Pattern Recognition Emphasis: The course highlights how NNs excel at pattern matching, a fundamental aspect of their power.
  • Hands-on Programming Assignments: Students implement and experiment with various NN architectures (feed-forward networks, CNNs, transformers) and fine-tune LLMs.
  • Reading and Discussion: The curriculum incorporates seminal papers and contemporary articles on AI, NNs, and their societal impact, fostering critical discussion.
  • Contrast with Classical AI: While NN-focused, the course includes discussions of classical AI approaches (e.g., Deep Blue, Sudoku solvers) to highlight the differences and the evolution of the field.

Datasets

The paper mentions the use of the MNIST dataset for training and testing feed-forward networks and convolutional neural networks. The TinyStories dataset is also mentioned in the context of fine-tuning LLMs.

Key Insights and Ideas

  • The Power of Feature Learning: The paper strongly advocates that feature learning by machines is a more powerful paradigm for developing AI than manual feature engineering by humans.
  • Motivation through Modern AI: Students are highly motivated by learning about systems like AlphaGo and ChatGPT, making NN-centric learning a natural fit for engagement.
  • Bridging the Gap to Societal Impact: By understanding the inner workings of modern AI, students are better equipped to inform public debate and contribute to the responsible development of AI.
  • Evolution of AI Teaching: The authors argue that AI education needs to evolve to reflect the current state of AI research and development, moving beyond older paradigms.
  • The Importance of "Warm-up" Assignments: Student feedback indicated a need for introductory assignments to build confidence and understanding before tackling more complex tasks.

Key Results

The paper presents a detailed course structure and learning outcomes, rather than empirical results from experiments. The authors report overwhelmingly positive feedback from students and colleagues on the refocused course. Student feedback highlighted the clarity of learning, the relevance of software development assignments, and the value of class discussions.

Limitations and Future Work

The authors acknowledge that the course is "very much in flux" and undergoes significant revisions. They mention that while they cover some aspects of search and probability, these are limited. They also note that the software development assignments, while improved, could benefit from further scaffolding. Future work includes refining the course materials, potentially collaborating with other institutions, and further developing assignments, particularly those related to LLMs. They express a desire to continue modifying the LLM assignment and hope to have more time for it in the future.

Summary updated

BibTeX
@inproceedings{wollowski2025towards,
  author = {Michael Wollowski},
  title = {{Towards an AI Course Based on Neural Networks}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i28.35179},
}
Figure