AI and Machine Learning Terms
88 flashcards by MentisLibrary
The vocabulary of modern AI, from gradient descent and overfitting to transformers, embeddings, and the terms behind large language models.
- What is machine learning?
- Building systems that learn patterns from data instead of following hand-written rules
- What is supervised learning?
- Learning from labeled examples: inputs paired with correct outputs
- What is unsupervised learning?
- Finding structure in unlabeled data
- What is reinforcement learning?
- An agent learns by acting in an environment and receiving rewards or penalties
- What is a training set?
- The portion of data the model learns its parameters from
- What is a validation set?
- Held-out data used to tune hyperparameters and detect overfitting during development
- What is a test set?
- Data touched only once, at the end, to estimate real-world performance
- What is overfitting?
- The model memorizes training data (including noise) and fails to generalize
- What is underfitting?
- The model is too simple to capture the underlying pattern
- What is the bias-variance tradeoff?
- Balancing error from oversimplified assumptions (bias) against error from sensitivity to training data (variance)
- What is regularization?
- Techniques that discourage model complexity to improve generalization
- What is dropout?
- Randomly deactivating a fraction of neurons during each training step
- What is early stopping?
- Halting training when validation performance stops improving
- What is gradient descent?
- Iteratively adjusting parameters in the direction that reduces the loss
- What is stochastic gradient descent (SGD)?
- Gradient descent using a small random batch of examples per update instead of the full dataset
- What is the learning rate?
- The step size for each parameter update during training
- What is backpropagation?
- The algorithm that computes loss gradients for every weight by applying the chain rule backward through the network
- What is a loss function?
- A number measuring how wrong the model's predictions are; training minimizes it
- What is cross-entropy loss?
- A loss that penalizes confident wrong predictions, comparing predicted probabilities to true labels
- What is a neural network?
- Layers of weighted connections and nonlinear activations that transform inputs into outputs
- What is an activation function?
- The nonlinearity applied after a neuron's weighted sum
- What is ReLU?
- Rectified Linear Unit: outputs max(0, x)
- What is softmax?
- A function that turns a vector of scores (logits) into a probability distribution summing to 1
- What is a hyperparameter?
- A setting chosen before training rather than learned from data
- What is an epoch?
- One complete pass through the entire training dataset
- What is a batch (and batch size)?
- The set of examples processed together in one forward/backward pass; its count is the batch size
- What is feature engineering?
- Manually crafting input representations that make patterns easier to learn
- What is a confusion matrix?
- A table of predicted vs actual classes showing where a classifier errs
- What is precision?
- Of everything the model flagged positive, the fraction that truly is positive: TP / (TP + FP)
- What is recall?
- Of all true positives, the fraction the model found: TP / (TP + FN)