Results for "preference optimization"

Optimizing policies directly via gradient ascent on expected reward.

Optimization under uncertainty.

Measure of vector magnitude; used in regularization and optimization.

Model optimizes objectives misaligned with human values.

Lowest possible loss.

A subfield of AI where models learn patterns from data to make predictions or decisions, improving with experience rather than explicit rule-coding.

A branch of ML using multi-layer neural networks to learn hierarchical representations, often excelling in vision, speech, and language.

Training with a small labeled dataset plus a larger unlabeled dataset, leveraging assumptions like smoothness/cluster structure.

Learning where data arrives sequentially and the model updates continuously, often under changing distributions.

Methods that learn training procedures or initializations so models can adapt quickly to new tasks with little data.

Automatically learning useful internal features (latent variables) that capture salient structure for downstream tasks.

The learned numeric values of a model adjusted during training to minimize a loss function.

A scalar measure optimized during training, typically expected loss over data, sometimes with regularization terms.

Average of squared residuals; common regression objective.

Iterative method that updates parameters in the direction of negative gradient to minimize loss.

Popular optimizer combining momentum and per-parameter adaptive step sizes via first/second moment estimates.

A gradient method using random minibatches for efficient training on large datasets.

One complete traversal of the training dataset during training.

A parameterized function composed of interconnected units organized in layers with nonlinear activations.

Gradients grow too large, causing divergence; mitigated by clipping, normalization, careful init.

Nonlinear functions enabling networks to approximate complex mappings; ReLU variants dominate modern DL.

Constraining outputs to retrieved or provided sources, often with citation, to improve factual reliability.

An RNN variant using gates to mitigate vanishing gradients and capture longer context.

Model trained to predict human preferences (or utility) for candidate outputs; used in RLHF-style pipelines.

Ensuring model behavior matches human goals, norms, and constraints, including reducing harmful or deceptive outputs.

Ordering training samples from easier to harder to improve convergence or generalization.

Training across many devices/silos without centralizing raw data; aggregates updates, not data.

Practices for operationalizing ML: versioning, CI/CD, monitoring, retraining, and reliable production management.

Removing weights or neurons to shrink models and improve efficiency; can be structured or unstructured.

Converts logits to probabilities by exponentiation and normalization; common in classification and LMs.