This post shares Mohammed Houbid’s journey through the Oracle ML interview process, providing insights into the questions he was asked, the challenges he faced, and how he approached them.
Introduction
The interview began with the interviewer introducing herself, sharing her educational background, role at Oracle, and the exciting projects she’s working on. Then, she asked Mohammed to introduce himself.
Mohammed talked about his journey into machine learning, why he’s passionate about it, and what drew him to Oracle, emphasizing its innovative approach and leadership in tech.
Diving Into Mohammed’s Projects
Having reviewed his resume beforehand, the interviewer asked in-depth questions about two key projects he worked on:
1. Lung Cancer Detection in CT Scan Images
Questions:
Where did you source the data?
How did you preprocess it?
Which model architectures did you use?
Did you apply transfer learning? If so, can you explain it?
How did you measure performance?
2. Fine-Tuning LLAMA 2
Questions:
What data did you use?
Where did you perform the fine-tuning?
What is the core concept of fine-tuning?
Can you explain quantization (symmetric vs. asymmetric) and its mathematical foundation?
What’s the difference between full fine-tuning and half fine-tuning?
Can you explain LoRA (Low-Rank Adaptation), QLoRA, and how matrix decomposition reduces fine-tuning costs?
The questions were highly technical and aimed at assessing the depth of Mohammed’s knowledge and understanding.
Explaining ML Projects
The interviewer asked Mohammed how he’d tackle a machine learning project end-to-end. He outlined his approach:
Defining the problem.
Collecting and preparing data.
Choosing the right model.
Training and testing the model.
Deploying the solution in the real world.
Testing Analytical Skills
The interviewer presented a dataset with details like income, age, gender, relationship status, and a target variable (e.g., whether a customer would leave a service).
She asked Mohammed to:
Analyze the “relationship status” column to find trends or patterns.
Perform calculations on the “age” column, like finding the average, median, and standard deviation.
This part was all about showing how Mohammed could make sense of raw data.
Imbalanced Data Scenarios
The conversation then shifted to imbalanced data—when one class (like fraud cases) is much rarer than the other.
She asked:
What is imbalanced data, and how do you handle it?
Can you give real-world examples? Mohammed mentioned fraud detection and spam email classification.
The interviewer presented a fun scenario: a spam detection model that always predicts “not spam.”
She challenged him to:
Create a confusion matrix.
Calculate metrics like accuracy and F1-score.
Interpret the results and suggest improvements.
Mohammed proposed using AI to generate synthetic spam emails. The interviewer was intrigued and asked him to explain how Transformer models and attention mechanisms work. He gave a simple breakdown of how these models focus on the most relevant parts of input data to improve predictions.
The Coding Challenge
The final part of the interview was a coding challenge on LeetCode. Mohammed solved the problem, explaining his approach and analyzing the time and space complexity of his solution.
Takeaways
Mohammed’s experience highlights the key skills needed for success in a machine learning interview:
Deep knowledge of your projects and ability to explain them clearly.
Strong grasp of ML concepts, metrics, and techniques.
Analytical problem-solving skills and coding proficiency.
If you’re preparing for an ML interview, let Mohammed’s experience inspire you to approach it with confidence and clarity. Good luck, and go crush it!