Solving Machine Learning Problems
Project Overview
This document discusses the implementation of a generative AI system tailored for solving machine learning problems in MIT's undergraduate Introduction to Machine Learning course. By leveraging advanced technologies such as Transformers and graph neural networks, the system achieves high accuracy in answering various types of questions, including open-response and multiple-choice formats. Additionally, it generates helpful hints to support student learning, highlighting its effectiveness as an educational tool. The findings suggest that this generative AI not only enhances student engagement but also improves comprehension in STEM subjects, indicating significant potential for transforming educational practices and outcomes in higher education.
Key Applications
Machine Learning Assistance System
Context: MIT's Introduction to Machine Learning course targeting undergraduate students, including those struggling with machine learning questions.
Implementation: Utilizes a Transformer architecture combined with graph neural networks to process course-related questions and generate hints based on student understanding, providing partial solutions and examples.
Outcomes: Achieved an accuracy of 96% for open-response questions and 97% for multiple-choice questions, surpassing the average performance of MIT students. Helps students fill knowledge gaps and learn by example, enhancing the overall learning experience.
Challenges: Limited to non-coding questions and cannot process image-based problems; dependence on the quality of the underlying model for hint relevance and usefulness.
Implementation Barriers
Technical Limitations
The model cannot handle coding questions or questions that require image processing.
Proposed Solutions: Plans to expand the model capabilities to include coding and visual questions in future iterations.
Understanding Complexity
The model struggles with questions that require deep comprehension of mathematical concepts.
Proposed Solutions: Improving the training dataset and augmenting it with a broader range of problem types.
Project Team
Sunny Tran
Researcher
Pranav Krishna
Researcher
Ishan Pakuwal
Researcher
Prabhakar Kafle
Researcher
Nikhil Singh
Researcher
Jayson Lynch
Researcher
Iddo Drori
Researcher
Contact Information
For information about the paper, please contact the authors.
Authors: Sunny Tran, Pranav Krishna, Ishan Pakuwal, Prabhakar Kafle, Nikhil Singh, Jayson Lynch, Iddo Drori
Source Publication: View Original PaperLink opens in a new window
Project Contact: Dr. Jianhua Yang
LLM Model Version: gpt-4o-mini-2024-07-18
Analysis Provider: Openai