Metrics and losses

Notes on evaluation metrics and loss functions used across ML domains.

Loss functions

How models learn — the optimization signal during training.

• General losses — Cross-Entropy, MSE/MAE, KL Divergence, L1/L2/Elastic Net regularization.
• NLP losses — NLL, Perplexity, CTC, Triplet, Contrastive, RLHF (PPO/DPO).
• Computer vision losses — Focal, Dice, IoU/Jaccard, Perceptual, Adversarial, SSIM.

Evaluation metrics

How models are measured — performance assessment after training.

• f1 score — harmonic mean of precision and recall, with macro/micro/weighted variants.
• Confusion matrix — true/false positives and negatives for classification.
• Regression metrics — MSE, RMSE, MAE, R-squared and variants.
• Recommendation system metrics — NDCG, MAP, MRR, Hit Rate, coverage, diversity.
• Computer vision metrics — IoU, mAP, pixel accuracy, FID, SSIM.
• NLP metrics — BLEU, ROUGE, METEOR, perplexity, BERTScore.
• Clustering metrics — silhouette score, Davies-Bouldin, Calinski-Harabasz, purity.