AI Bot
Core AI Concepts & Terminology Flashcards
Microsoft Azure AI Fundamentals AI-901 Flashcards
Study our Core AI Concepts & Terminology flashcards for the Microsoft Azure AI Fundamentals AI-901 exam with 80+ flashcards. View as flashcards, a searchable table, or as a fun matching game.
| Front | Back |
| Accuracy | Proportion of correct predictions over all predictions |
| Activation function | Nonlinear function applied to neuron output enabling complex representations |
| Annotations | Labeling of data used for supervised learning |
| Anomaly detection | Identifying unusual observations that differ from normal patterns |
| AUC | Area under the ROC curve representing ranking ability |
| Bagging | Ensemble technique that builds models on bootstrap samples and averages predictions |
| Batch size | Number of samples processed before model parameters are updated |
| Bias in data | Systematic errors in data that lead to unfair outcomes |
| Bias variance tradeoff | Balancing model complexity to minimize error from bias and variance |
| Boosting | Sequentially building models that focus on previous errors |
| Classification | Predicting a discrete class label |
| Clustering | Grouping similar data points without labels |
| Confounding variable | External factor that affects both features and target causing spurious associations |
| Confusion matrix | A table showing true positives false positives true negatives and false negatives |
| Cross validation | Evaluating model performance by rotating train and validation splits |
| Curse of dimensionality | Problems that arise when number of features is very high relative to samples |
| Data leakage | When information from test data leaks into training causing inflated performance |
| Deterministic model | Model that outputs a single fixed prediction for a given input |
| Dimensionality reduction | Reducing number of features while preserving structure |
| Discriminative model | Model that learns to map inputs to labels or outputs directly |
| Dropout | Randomly disabling neurons during training to prevent overfitting |
| Early stopping | Stopping training when validation performance stops improving |
| Embedding | Low dimensional dense vector that represents discrete items |
| Ensemble methods | Combining multiple models to improve performance |
| Epoch | One full pass over the entire training dataset |
| Explainability | Ability to understand how a model makes decisions |
| F1 score | Harmonic mean of precision and recall |
| False negative | Incorrectly predicted negative case |
| False positive | Incorrectly predicted positive case |
| Feature engineering | Creating or transforming inputs to improve model performance |
| Fine tuning | Adjusting a pretrained model on task specific data |
| Generative model | Model that learns to generate data resembling the training distribution |
| Gradient descent | Optimization method that updates parameters using gradients of loss |
| Hyperparameter | Configurable setting that is not learned during training |
| Hyperparameter tuning | Searching for best hyperparameters often with grid or random search |
| Inference phase | Phase where a trained model makes predictions on new data |
| Interpretability | Ease of understanding model mechanisms or predictions |
| Iteration | One parameter update step using a batch of data |
| L1 regularization | Sparse penalty that encourages zero coefficients |
| L2 regularization | Punishes large weights and promotes small distributed weights |
| Latency | Time taken for a model to produce a prediction |
| Learning rate | Step size used by optimization algorithm to update parameters |
| Log loss | Cross entropy loss measuring confidence of probabilistic predictions |
| Loss function | Function that quantifies error between predictions and targets |
| Mean absolute error | Average of absolute differences between predictions and targets |
| Mean squared error | Average of squared differences between predictions and targets |
| Model calibration | When predicted probabilities reflect true likelihoods |
| Model versioning | Tracking model artifacts and configurations for reproducibility |
| Normalization | Scaling features to a common range often 0 to 1 |
| One hot encoding | Binary vector representation for categorical variables |
| Overfitting | Model fits training data too closely and fails on new data |
| Parameter | Model weights that are learned during training |
| Precision | Proportion of positive predictions that are correct |
| Precision recall curve | Plot of precision versus recall across thresholds |
| Principal component analysis | Linear method to reduce dimensionality by orthogonal projections |
| Probabilistic model | Model that outputs probability distributions over outcomes |
| R squared | Proportion of target variance explained by the model |
| Recall | Proportion of actual positives that are detected |
| Regression | Predicting a continuous numeric value |
| Regularization | Techniques that penalize complexity to reduce overfitting |
| Reinforcement learning | Learning a policy through trial and reward signals |
| ROC curve | Plot of true positive rate versus false positive rate at different thresholds |
| Self supervised learning | Using part of data as pseudo labels to learn representations |
| Semi supervised learning | Learning from a mix of labeled and unlabeled data |
| Sigmoid | Activation that maps values to a 0 to 1 range useful for binary outputs |
| Softmax | Function that converts logits to a probability distribution over classes |
| Specificity | Proportion of actual negatives that are correctly identified |
| Stacking | Ensemble method that trains a meta model on base model outputs |
| Standardization | Scaling features to zero mean and unit variance |
| Stochastic gradient descent | Gradient descent using single or small batches of samples |
| Supervised learning | Learning from labeled examples to predict labels for new data |
| Throughput | Number of predictions a system can produce per unit time |
| Tokenization | NLP process of splitting text into tokens |
| Train validation test split | Using separate sets for training tuning and final evaluation |
| Training phase | Phase where model parameters are learned from data |
| Transfer learning | Using a pretrained model as a starting point for a related task |
| True negative | Correctly predicted negative case |
| True positive | Correctly predicted positive case |
| Underfitting | Model is too simple to capture underlying patterns |
| Unsupervised learning | Finding patterns or structure in unlabeled data |
About the Flashcards
Flashcards for the Microsoft Azure AI Fundamentals exam provide a compact review of core machine learning terminology and model types. Cards cover learning paradigms such as supervised, unsupervised, semi-supervised, self-supervised and reinforcement learning, distinctions between generative and discriminative approaches, and common tasks like classification, regression, clustering, and anomaly detection.
They also review training and evaluation concepts used on exams: loss functions and metrics, optimization and hyperparameters (learning rate, gradient descent, SGD), epochs and batch sizes, cross-validation and train/validation/test splits. Complementary cards cover preprocessing and features (one-hot, embeddings, normalization, PCA), regularization and ensembles, data leakage and bias, interpretability, calibration, latency and throughput.
Topics covered in this flashcard deck:
- Supervised, unsupervised, reinforcement
- Classification and regression
- Evaluation metrics and loss
- Optimization and training
- Feature engineering and preprocessing
- Regularization and ensembles