MIT researchers have developed a machine learning model that can solve college-level math problems in seconds at a human level

Unlike humans, machine learning models find it extremely difficult to handle problems involving differential equations, linear algebra, and multivariate calculus. Even the most advanced models can only answer elementary or high school math problems, and they don’t always provide the right answers. A multidisciplinary research team at MIT has created a neural network model that can quickly and accurately answer college-level arithmetic problems. The model can also automatically explain solutions in college math courses and quickly produce new problems. University students were then given the computer-generated questions to test, and they could not determine whether an algorithm or a human had produced the questions. The study was also published in the National Academy of Sciences Proceedings.

The researchers believe their work can be used to accelerate the creation of course content for in-depth residential courses and massive open online courses (MOOCs) with thousands of students. The program could also serve as an automated tutor that shows students how to solve college-level math problems. The team believes that by helping teachers understand the connection between courses and their prerequisites, their approach has the potential to improve higher education. For more than two years, the model has continued to evolve. At first, researchers saw that pre-trained models using only text could not provide high accuracy on high school math problems. In contrast, those using graphical neural networks might require longer training periods.

The scientists then experienced a “eureka” moment. They used program synthesis and hit-and-miss learning to convert undergraduate math class questions from well-known universities that the model had never encountered before into programming tasks. The researchers added an extra “fine-tuning” step before passing these programming tasks to a neural network. The pre-trained neural network used, Codex, has been “tuned” on both text and code. The pre-trained model was trained on data containing millions of lines of code and natural language words, which allowed it to understand the connection between text and code. With just a few question code examples, the template can now convert a text question into code and then run the code to provide an answer, as it can recognize different relationships between text and code. This method showed a huge improvement in accuracy – from 8 to 80%. By giving the neural network a set of arithmetic problems about a topic and then asking it to come up with a new challenge, the researchers also used their model to generate queries. We also reviewed these computer-generated questions by displaying them to students. Students gave comparable ratings to human-generated and machine-generated questions for level of difficulty and appropriateness to the course, as they could not distinguish between human-generated and algorithm-generated questions.

The team says their efforts are aimed at paving the way for using machine learning to solve more difficult problems rather than replacing human teachers. Although the team is pleased with the results of their strategy, they must overcome several drawbacks. Due to computational complexity, the model cannot answer questions with a visual component and cannot solve unsolvable computational problems. In addition to overcoming these hurdles, they want to scale the model up to hundreds of courses so it can improve automation and offer insight into course design and curriculum.

This Article is written as a research summary article by Marktechpost Staff based on the research paper 'A neural network solves, explains, and generates university math problems by program synthesis and few-shot learning at human level'. All Credit For This Research Goes To Researchers on This Project. Checkout the paper and reference article.

Please Don't Forget To Join Our ML Subreddit


Khushboo Gupta is an intern consultant at MarktechPost. She is currently pursuing her B.Tech from Indian Institute of Technology (IIT), Goa. She is passionate about the fields of machine learning, natural language processing and web development. She likes to learn more about the technical field by participating in several challenges.


Sherry J. Basler