Machine Learning Engineering

Bias is error from overly simple assumptions (underfitting); variance is error from sensitivity to the training data (overfitting). Increasing model complexity lowers bias but raises variance, so you aim for the balance that minimises validation error.

Use metrics like precision/recall or PR-AUC instead of accuracy, try class weights, resampling (over/under-sampling, SMOTE), adjust the decision threshold, and collect more minority-class data if possible.

Techniques that discourage overly complex models to reduce overfitting — e.g. L1/L2 penalties, dropout, early stopping and data augmentation.

Precision matters when false positives are costly (spam filters). Recall matters when missing a positive is costly (fraud, disease screening). F1 balances both.

When information that would not be available at prediction time sneaks into training — e.g. using future data or target-derived features — leading to great offline scores and poor real-world performance.

It uses self-attention so each token can weigh every other token in the sequence, stacked with feed-forward layers and positional information, enabling parallel training and long-range context.

Starting from a model pretrained on a large dataset and fine-tuning it on your smaller task, which saves data and compute and often improves accuracy.

Track input data distributions for drift, prediction distributions, latency and errors, and — when labels arrive — real accuracy. Alert on thresholds and trigger retraining or rollback.

Low-Rank Adaptation freezes the base model and trains small low-rank matrices added to certain layers, making fine-tuning large models far cheaper in memory and storage.

Batch scores many records on a schedule (cheap, high latency, e.g. nightly recommendations). Online serves single requests in real time (low latency, needs scalable APIs, e.g. fraud checks at payment).