AI Case Study: Abstract Pattern Recognition
Welcome to the final stage of our hiring process. This project is designed to test your practical skills in a scenario similar to what you will face at ReasonEra.
If you need more time (e.g., for a master's thesis or current job), please email us to request an extension.
The Problem
ReasonEra's core AI must solve abstract reasoning matrices (like Raven's Progressive Matrices) instantly. Existing general-purpose vision models often fail to grasp the subtle logical rules governing these patterns (e.g., rotation, addition/subtraction of elements, XOR operations).
Your task is to build a small proof-of-concept model that can solve a simplified version of this problem better than a standard off-the-shelf ResNet or ViT.
Project Requirements
1. Dataset
We recommend using a publicly available dataset like RAVEN (Relational and Analogical Visual rEasoNing) or PGM (Procedurally Generated Matrices). You do not need to use the whole dataset; a small subset is sufficient for this POC.
2. The Model
Design and train a model that takes a matrix puzzle as input and outputs the correct answer. We are interested in your architecture choices. How do you explicitly model the relationships between panels?
3. Deliverables
- Code Repository: A link to a private GitHub repo (invite user:
[YOUR_GITHUB_USERNAME]) OR a zipped folder of your code. - Technical Report (2-3 pages PDF): Explain your approach. Why did you choose this architecture? what were the results? what would you improve if you had more time?
- Video Walkthrough (Optional but recommended): A 5-minute Loom or YouTube video explaining your code and thought process.
How to Submit
When you are ready, please email your deliverables directly to us.
Email Your Submission