| No. | Item | Definition |
|---|---|---|
| 1. | accuracy | overall proportion predicted correctly |
| 2. | activation | function applied to neuron output |
| 3. | Adam | adaptive gradient optimization method |
| 4. | algorithm | step-by-step problem-solving method |
| 5. | attention | mechanism weighting relevant inputs |
| 6. | AUC | area under ROC curve |
| 7. | augmentation | creating varied training examples |
| 8. | backpropagation | gradient computation through layers |
| 9. | baseline | simple reference performance |
| 10. | batch | subset processed together |
| 11. | benchmark | standard test for comparison |
| 12. | bias | added offset term |
| 13. | boosting | sequentially improving weak learners |
| 14. | checkpoint | saved training state |
| 15. | classification | predicting a category label |
| 16. | clustering | grouping similar unlabeled items |
| 17. | confusion matrix | table of prediction outcomes |
| 18. | convergence | approaching a stable solution |
| 19. | corpus | large body of text |
| 20. | cross-validation | repeated train-test splitting |
| 21. | dataset | collection of training examples |
| 22. | decision tree | rule-splitting predictive model |
| 23. | deep learning | machine learning with many layers |
| 24. | distance | measure of separation |
| 25. | drift | data distribution changing over time |
| 26. | dropout | randomly disabling units during training |
| 27. | embedding | dense numeric representation |
| 28. | encoding | turning categories into numbers |
| 29. | ensemble | combination of multiple models |
| 30. | epoch | one full pass through data |
| 31. | F1 score | precision-recall harmonic mean |
| 32. | fairness | equitable behavior across groups |
| 33. | feature | input variable used by a model |
| 34. | few-shot | performing with few examples |
| 35. | fine-tuning | adapting a pretrained model |
| 36. | generalization | performance on unseen data |
| 37. | generation | producing new content |
| 38. | gradient | direction of steepest change |
| 39. | ground truth | correct reference answer |
| 40. | guardrail | constraint for safer behavior |
| 41. | hallucination | confident but false output |
| 42. | hyperparameter | setting chosen before training |
| 43. | imputation | filling missing values |
| 44. | inference | using a trained model |
| 45. | interpretability | ease of understanding model behavior |
| 46. | k-means | centroid-based clustering method |
| 47. | kernel | similarity function or filter |
| 48. | label | target value to predict |
| 49. | latent | hidden underlying representation |
| 50. | layer | group of model units |
| 51. | learning rate | step size during updates |
| 52. | logit | pre-sigmoid score value |
| 53. | loss | measure of model error |
| 54. | metric | numerical performance measure |
| 55. | MLOps | operations for machine learning systems |
| 56. | model | learned system making predictions |
| 57. | monitoring | tracking system behavior over time |
| 58. | n-gram | contiguous token sequence |
| 59. | naive Bayes | probabilistic classifier with independence assumption |
| 60. | nearest neighbor | most similar stored example |
| 61. | neural network | layered function approximator |
| 62. | neuron | single computational unit |
| 63. | noise | random unwanted variation |
| 64. | one-hot | binary vector category representation |
| 65. | optimizer | method updating model parameters |
| 66. | outlier | unusually extreme data point |
| 67. | overfitting | memorizing training data too closely |
| 68. | parameter | learned value inside a model |
| 69. | PCA | linear dimensionality reduction method |
| 70. | pipeline | ordered data-processing workflow |
| 71. | precision | share of predicted positives correct |
| 72. | prediction | model output for an input |
| 73. | preprocessing | cleaning and preparing inputs |
| 74. | pretraining | initial large-scale model training |
| 75. | probability | estimated likelihood of an outcome |
| 76. | quantization | reducing numeric precision |
| 77. | recall | share of actual positives found |
| 78. | regression | predicting a continuous value |
| 79. | regularization | technique reducing overfitting |
| 80. | reinforcement learning | learning through rewards and actions |
| 81. | ReLU | zeroes negative values |
| 82. | ROC curve | tradeoff across classification thresholds |
| 83. | sampling | selecting examples from data |
| 84. | SGD | basic stochastic gradient descent |
| 85. | sigmoid | S-shaped squashing function |
| 86. | similarity | degree of likeness |
| 87. | softmax | turns scores into probabilities |
| 88. | support vector machine | margin-based classifier |
| 89. | test set | held-out data for final evaluation |
| 90. | threshold | cutoff for decision making |
| 91. | token | basic unit of text |
| 92. | train set | data used to fit model |
| 93. | training | process of fitting a model |
| 94. | transfer learning | reusing knowledge across tasks |
| 95. | transformer | attention-based neural architecture |
| 96. | underfitting | failing to capture patterns |
| 97. | versioning | tracking changes to artifacts |
| 98. | vocabulary | set of known tokens |
| 99. | weight | connection strength in a model |
| 100. | zero-shot | performing without task examples |
