Business Context

FitVision trains a neural network to predict daily calorie burn from wearable sensor data for 8 million active users. The ML team wants to standardize optimizer selection and needs a clear, empirical comparison of Adam vs SGD under realistic training constraints.

Dataset

You are given a supervised regression dataset built from 12 months of wearable telemetry and profile features.

• Size: 240K user-day records, 36 input features
• Target: Daily calorie burn (continuous)
• Data quality: 6% missing in recovery metrics, 2% missing in sensor aggregates due to sync failures
• Outliers: Some extreme workout days and device glitches

Success Criteria

A strong solution should:

• Compare Adam and SGD on the same model architecture and preprocessing pipeline
• Reach test RMSE < 115 kcal and show a justified optimizer choice
• Report convergence speed, final validation performance, and training stability
• Explain when Adam is preferable and when SGD may generalize better

Constraints

• Training must finish within 30 minutes on a single CPU or modest GPU
• The final model will be retrained weekly
• Inference latency must stay under 20 ms per example
• The approach should be simple enough for another engineer to maintain

Deliverables

1. Build a neural-network regression pipeline with identical architecture for both optimizers
2. Implement preprocessing for missing values, categorical encoding, and feature scaling
3. Train and compare Adam vs SGD using the same train/validation/test split
4. Evaluate with RMSE, MAE, and convergence behavior across epochs
5. Recommend one optimizer for production and explain the tradeoffs in learning rate sensitivity, convergence speed, and generalization