All 68 questions
What are overfitting and underfitting, and how do you detect and prevent them?
Asked in 15 of 320 interviews · last asked Oct 23 · Concept · Junior
Show answer
Overfitting means learning training data too closely, including noise, and performing poorly on new data. Underfitting means missing useful patterns, often because the model is too simple. A large training-validation gap suggests overfitting; poor results on both suggest underfitting. Reduce overfitting with regularization or early stopping; reduce underfitting with better features, more model complexity, or enough training.
- Use a separate validation set or cross-validation to check performance on unseen data.
- Learning curves show how training and validation errors change with training time or data size.
- More representative training data can reduce overfitting; stronger regularization can worsen underfitting.
- Keep the test set untouched until final evaluation.
What the interviewer checks: The interviewer checks whether you can recognize each problem and choose changes that improve performance on unseen data.
Other ways it was asked
- “What are overfitting and underfitting?”
- “Describe overfitting, how you detect it, and how you prevent it.”
- “What is overfitting in machine learning?”
What is the difference between L1 and L2 regularization?
Asked in 4 of 320 interviews · last asked Aug 11 · Concept · Junior
Show answer
L1 regularization adds a penalty based on the sum of absolute weight values; L2 uses the sum of squared weights. Both help reduce overfitting by discouraging large weights. L1 can make weights exactly zero, which helps select features. L2 usually shrinks weights without making them exactly zero.
- L1 encourages sparse models: some features get zero weights.
- L2 spreads shrinkage across weights and can help when features are strongly correlated.
- A larger regularization strength means a stronger penalty.
- Scale features so their units do not unfairly affect the penalty.
What the interviewer checks: The interviewer is checking whether you understand how each penalty affects weights, feature selection, and overfitting.
Other ways it was asked
- “What is L2 regularization?”
- “What are L1 and L2 regularization?”
- “What are L1 and L2 regularization, and what are they used for?”
What is gradient descent?
Asked in 4 of 320 interviews · last asked Sep 4 · Concept · Junior
Show answer
Gradient descent is a method for reducing a model’s loss by repeatedly adjusting its parameters. It calculates the loss gradient and moves the parameters in the opposite direction. The learning rate scales each step, while the gradient determines its direction and also affects its size.
- The gradient points toward the steepest local increase in loss.
- Each update subtracts the learning rate times the gradient from the parameters.
- A large learning rate can cause overshooting; a small one can make training slow.
- Gradient descent does not guarantee finding the global minimum.
What the interviewer checks: The interviewer checks whether you understand how gradients guide parameter updates and how the learning rate affects training.
Other ways it was asked
- “What do you mean by gradient descent?”
- “What is gradient descent, how does it work, and what determines its pace?”
- “What happens under the hood in gradient descent?”
What are bias and variance, how do they relate to RMSE, and what is their tradeoff?
Asked in 3 of 320 interviews · last asked Sep 7 · Concept · Junior
Other ways it was asked
- “Tell me about the bias-variance trade-off and how bias and variance are related to RMSE.”
- “How are bias and variance defined and calculated, and how are they related to RMSE?”
- “What is the equation relating RMSE, bias, and variance?”
How would you improve the accuracy of a predictive model?
Asked in 2 of 320 interviews · last asked Mar 27 · Scenario · Mid-level
Other ways it was asked
- “How would you improve a machine learning model?”
Describe some common machine learning architectures and when you would use them.
Asked in 2 of 320 interviews · last asked Oct 19 · Concept · Mid-level
Other ways it was asked
- “Describe common machine learning architectures, their typical use cases, and briefly note any limitations.”
How do you train a machine-learning model?
Asked in 2 of 320 interviews · last asked Sep 10 · Concept · Junior
Other ways it was asked
- “How would you train a machine learning model?”
What is the learning rate, and how does increasing or decreasing it affect training?
Asked in 2 of 320 interviews · last asked Oct 9 · Concept · Mid-level
Other ways it was asked
- “Can you explain learning rate in simple language?”
- “How do the trade-offs between a low and a high learning rate affect model training?”
- “What does the learning rate mean in machine learning?”
What does it mean to freeze the model head during transfer learning?
Asked once · last asked Sep 16 · Concept · Mid-level
Other ways it was asked
- “When should you freeze the model head during transfer learning?”
How would you decide whether to spend limited time collecting more data or improving the model’s algorithm to achieve the best results?
Asked once · last asked Oct 20 · Scenario · Senior
Other ways it was asked
- “How would you decide whether to collect more data or improve a machine learning algorithm when time is limited?”
What initial parameters would you use to fit a linear regression model to data in three-dimensional space, and how would you optimize them?
Asked once · last asked Mar 4 · Concept · Junior
Could you explain how structured feedback can optimize the performance of machine learning models?
Asked once · last asked Apr 8 · Concept · Mid-level
How do data augmentation and transfer learning help when labeled data is scarce in supervised learning?
Asked once · last asked Apr 21 · Concept · Mid-level
What is hyperparameter tuning, and when would you use it?
Asked once · last asked May 22 · Concept · Mid-level
What strategies would you use for hyperparameter tuning in object detection models?
Asked once · last asked May 22 · Scenario · Mid-level
How do you decide between GridSearchCV and RandomizedSearchCV, and for imbalanced classification, which scoring metric (such as PR AUC, ROC AUC, F1, or balanced accuracy) and cross-validation splitter would you use to preserve class ratios?
Asked once · last asked May 24 · Concept · Senior
What is the difference between stochastic gradient descent and gradient descent?
Asked once · last asked May 30 · Concept · Mid-level
Why do we divide the data into batches for stochastic gradient descent?
Asked once · last asked May 30 · Concept · Mid-level
If a dataset already contains noise, why does dividing it into batches help? Does batching introduce more noise, and why use stochastic gradient descent rather than batch gradient descent?
Asked once · last asked May 30 · Concept · Mid-level
In Python, how would you make ML experiment results reproducible across runs, and what are common sources of nondeterminism even after setting random seeds?
Asked once · last asked Jun 3 · Scenario · Senior
When would you choose to fine-tune a pretrained model versus using it as a frozen feature extractor, and what trade-offs drive that choice?
Asked once · last asked Jun 15 · Concept · Mid-level
How do gradient descent and its variants help find minima or maxima during optimization?
Asked once · last asked Jun 22 · Concept · Junior
Can you explain L1 and L2 regularization and how they affect model coefficients and feature selection?
Asked once · last asked Jun 26 · Concept · Mid-level
What is the loss function in logistic regression?
Asked once · last asked Jul 16 · Concept · Junior
Suppose you are training a deep learning model for product classification. The training loss steadily decreases, but the validation loss begins to rise after the tenth epoch. What is the most likely cause, and why?
Asked once · last asked Aug 1 · Scenario · Junior
Other ways it was asked
- “Suppose you are training a deep learning model for product classification on Amazon. The training loss steadily decreases, but the validation loss begins to rise after the tenth epoch. Which of the following is the most likely cause, and why?”
How can concepts such as attention and perception in human intelligence be incorporated into AI systems, and how could they improve model performance?
Asked once · last asked Aug 13 · Design · Senior
When fine-tuning a pretrained transformer on a small domain dataset, how do you prevent catastrophic forgetting while still adapting meaningfully to the new domain?
Asked once · last asked Aug 19 · Scenario · Senior
In AI data annotation and model evaluation, how do you differentiate between bias and variance when assessing a model's performance?
Asked once · last asked Aug 25 · Concept · Mid-level
How have recent advancements in AI affected bias and variance in models, and are there any trends emerging in the field that impact these issues?
Asked once · last asked Aug 25 · Concept · Senior
How do you account for data variability when building predictive analytics models?
Asked once · last asked Aug 31 · Scenario · Senior
What factors would guide your decision between using a pretrained transformer model and developing a custom model from scratch for a unique domain-specific application?
Asked once · last asked Sep 4 · Scenario · Mid-level
How would the availability and quality of domain-specific data affect your choice between a pretrained model and developing a custom one?
Asked once · last asked Sep 4 · Scenario · Mid-level
Do you design the model architecture, or mainly evaluate different architectures, and what is the biggest challenge or interesting finding?
Asked once · last asked Sep 8 · Behavioral · Mid-level
How do you tune hyperparameters for a machine learning model?
Asked once · last asked Sep 15 · Concept · Mid-level
Why do we use squared error instead of cubes, fourth powers, or other higher powers? How would those powers affect sensitivity, and what are outliers?
Asked once · last asked Dec 16 · Concept · Mid-level
How do you know that higher powers make the error metric more sensitive? On what basis are you saying that, and can you explain it in layman's terms?
Asked once · last asked Dec 16 · Concept · Mid-level
When would you train a face recognition model from scratch rather than use an existing model?
Asked once · last asked Sep 23 · Scenario · Mid-level
How do you train a custom model in AI Center?
Asked once · last asked Sep 24 · Tooling · Mid-level
If you multiply the loss for the cat class by a very large factor, such as 10,000, which class will the model learn faster: cats or dogs?
Asked once · last asked Sep 29 · Scenario · Mid-level
Do you think the features learned for one task are broadly reusable across other tasks, with only minimal changes in the representation for a given input?
Asked once · last asked Dec 20 · Concept · Mid-level
Can you explain L1 and L2 regularization mathematically, from a computer vision perspective?
Asked once · last asked Oct 9 · Concept · Senior
What is the biggest challenge in reinforcement learning, and how do you approach it?
Asked once · last asked Oct 17 · Scenario · Mid-level
How do you choose the regularization strength?
Asked once · last asked Oct 17 · Concept · Mid-level
What are common limitations of machine learning architectures, and when would you choose a different architecture?
Asked once · last asked Oct 19 · Concept · Senior
How might you approach improving the generalization of a machine learning model trained on a biomedical dataset?
Asked once · last asked Oct 20 · Scenario · Mid-level
In gradient-boosted trees, how do you decide whether to tune for lower bias or lower variance, and what diagnostics tell you you've gone too far either way?
Asked once · last asked Oct 23 · Scenario · Senior
When training XGBoost or LightGBM, how would you choose between tuning max_depth or num_leaves versus min_child_weight or min_data_in_leaf, and what failure mode does each primarily control?
Asked once · last asked Oct 23 · Concept · Senior
How are the parameters of a linear regression model initialized?
Asked once · last asked Mar 4 · Concept · Junior
What initial values would you use for the slope and intercept in a linear regression model?
Asked once · last asked Mar 4 · Concept · Junior
Can you manually initialize the slope and intercept in a linear regression model, or does initialization depend on the model?
Asked once · last asked Mar 4 · Concept · Junior
What is optimization?
Asked once · last asked Mar 4 · Concept · Junior
What mechanism do we use to optimize a linear regression model?
Asked once · last asked Mar 4 · Concept · Junior
What optimization techniques can you think of?
Asked once · last asked Mar 4 · Concept · Junior
What are transfer learning and hyperparameter tuning?
Asked once · last asked Apr 15 · Concept · Junior
How does bias affect a model's performance?
Asked once · last asked Apr 20 · Concept · Junior
What is the difference between bias and variance?
Asked once · last asked May 26 · Concept · Junior
What is zero-shot learning, and what is zero-shot classification?
Asked once · last asked Jul 16 · Concept · Junior
What is regularization?
Asked once · last asked Jul 16 · Concept · Junior
What do you know about hyperparameter optimization?
Asked once · last asked Jul 15 · Concept · Mid-level
What does gradient descent do during model training, and why is it useful?
Asked once · last asked Jul 31 · Concept · Junior
What is the difference between data augmentation and regularization?
Asked once · last asked Aug 27 · Concept · Junior
Why is overfitting problematic?
Asked once · last asked Aug 27 · Concept · Junior
What is transfer learning, and when is it useful?
Asked once · last asked Sep 16 · Concept · Junior
Other ways it was asked
- “What is the benefit of transfer learning?”
Which hyperparameters would you adjust during model tuning?
Asked once · last asked Sep 16 · Concept · Mid-level
Other ways it was asked
- “What parameters would you fine-tune during hyperparameter tuning?”
What happens if the learning rate is too high?
Asked once · last asked Sep 16 · Concept · Junior
Can you explain transfer learning in simple language?
Asked once · last asked Sep 16 · Concept · Junior
Can you explain a loss function in detail?
Asked once · last asked Sep 29 · Concept · Junior
How would you tune a machine-learning pipeline to improve prediction accuracy?
Asked once · last asked Oct 16 · Scenario · Mid-level
More Machine Learning interview questions
- All Machine Learning interview questions
- Machine Learning Fundamentals Interview Questions
- Data Preprocessing and Feature Engineering Interview Questions
- Machine Learning Algorithms and Model Selection Interview Questions
- Cross-validation and Data Leakage Interview Questions
- Model Evaluation Metrics Interview Questions
- Applied Machine Learning Interview Questions
- Search and Recommendation System Interview Questions
- MLOps and Model Deployment Interview Questions
How this list is made
Every question here was asked by a real interviewer in a real job interview that a candidate ran MeetAssist in. After the call, the interviewer's questions are pulled from the transcript. Practice sessions, tests, exams and anything that was not a job interview are dropped.
Questions are rewritten so they make sense on their own: names of people, employers and clients are removed, and speech-to-text mistakes are fixed. Small talk, logistics and questions about one candidate's own résumé are left out. Rephrasings of the same question are merged, and the count is the number of different interviews it came up in.
The interviews come from candidates who use MeetAssist, so they lean towards software, data and AI roles: the shares describe these interviews, not every Machine Learning interview. Answers to the most asked questions are written and checked by an editor; a "common mistake" is shown only when reviews of real answers in these interviews show it, with the count.
320 interviews, 221 candidates, Nov 2025 to Oct 2026. Updated every month with new interviews.
Or add the Chrome extension