Company Interview Questions · Telecom, Auto & Industry

Mercedes-Benz R&D India & Philips: AI / Data Scientist

❓ 6 questions🏢 Mercedes-Benz R&D & Philips🆕 Recent reports

Mercedes-Benz R&D India & Philips: AI / Data Scientist Interviews. Rounds, real reported questions and model answers, compiled from public candidate reports (sources at the end). Practice mode hides every answer: say your answer first, press Show answer, then rate yourself. Use 🔁 Review again to build your revision list and 🎲 Random question for a quick drill.

★ Must-know 145📝 Assignments
About this interview

Compiled from a public Mercedes-Benz Research and Development India (MBRDI, Bengaluru) report with ML-focused rounds (resume deep dive → domain questions → a scenario on modelling a CNN, with no coding round) and a public Philips Data Scientist process guide. The questions are paraphrased, and the model answers are written by EasyLearnSpace.

Mercedes-Benz R&D India Philips
Role AI / Applied Data Scientist (ML-focused report) Data Scientist
Rounds 3: resume deep dive → general domain → scenario: how would you model a CNN? (no coding) Recruiter → technical assessment (MCQ + coding: maths, programming, ML) → technical interviews → managerial → sometimes a presentation + Q&A
Reported questions What a CNN is, backpropagation, how to choose the batch size Supervised vs unsupervised, overfitting, feature engineering, confusion matrix, CLT, Type I/II, normality checks, longitudinal data, logistic coefficients for categorical variables
Domain Driver assistance, perception, in-car voice assistants, manufacturing analytics Medical imaging, patient monitoring, healthcare analytics

01Rounds at a glance

02Questions and model answers

Q1
Scenario: design a CNN to classify road signs from dash-cam images.
  • Data: cropped sign images across lighting, weather and occlusion; class balance; augmentation (brightness, blur, small rotations, no horizontal flips for asymmetric signs).
  • Model: a pre-trained lightweight backbone (MobileNet/EfficientNet) for edge deployment; replace the head with N classes; fine-tune with AdamW and cosine learning-rate decay.
  • Evaluation: per-class recall (safety-critical classes such as STOP are weighted), a confusion matrix, robustness tests (night, rain), latency on the target hardware.
  • Deployment: quantize to INT8 and validate; monitor field data for drift (new sign designs).
Q2
How do you choose the batch size?
  • Memory: the largest batch that fits on the GPU (activations scale with batch × resolution); gradient accumulation can simulate bigger batches.
  • Optimisation: large batches give smoother gradients and better throughput, but may generalise worse and need learning-rate scaling (the linear scaling rule plus warm-up). Small batches add noise that can help generalisation but train more slowly.
  • BatchNorm needs reasonably large batches (or use Group/LayerNorm).
  • In practice: tune the batch size together with the learning rate; common choices are 32–256 for vision fine-tuning.
Q3
Philips: how do you check whether data is normally distributed?
  • Visual: a histogram/KDE and a Q-Q plot (points on the line suggest normality).
  • Tests: Shapiro–Wilk (small n), D'Agostino K², Anderson–Darling. With large samples, tiny deviations become "significant", so rely more on plots and effect size.
  • If not normal: transform (log, Box-Cox), use non-parametric tests (Mann–Whitney, Wilcoxon) or bootstrap.
Q4
How do you interpret logistic regression coefficients for a categorical variable?
  • One-hot encode with a reference category. Each coefficient β is the change in log-odds relative to the reference, holding other variables constant.
  • exp(β) is the odds ratio. For example, β = 0.69 for "smoker" vs "non-smoker" means about 2× the odds of the outcome.
  • Check the confidence intervals; changing the reference category changes the coefficients, but not the model's predictions.
Q5
How do you handle longitudinal (repeated-measures) patient data?
  • Don't treat visits as independent (that underestimates the variance and leaks across splits). Split by patient.
  • Models:
  • Mixed-effects models (random intercepts and slopes per patient).
  • GEE.
  • Survival models for time-to-event.
  • Sequence models (RNN/Transformer) for event histories.
  • Missing visits: distinguish missing-at-random from dropout due to health status; use multiple imputation or models that handle irregular time.
Q6
What does a confusion matrix tell you, and which metrics come from it?
  • A table of actual vs predicted classes: TP, FP, FN, TN (N×N for multi-class).
  • Metrics:
  • Accuracy.
  • Precision.
  • Recall / sensitivity.
  • Specificity (TN/(TN+FP)), which is important in medicine.
  • F1.
  • Positive and negative predictive values, which depend on prevalence.
  • Use it to spot which classes get confused and to set thresholds based on clinical cost.

03How to prepare

Sources

Created by Ajeetkumar