self-critiquing, ai safety
Model critiques its own outputs.
9,967 technical terms and definitions
Model critiques its own outputs.
Distill model into itself.
Self-distillation uses model's own predictions as soft targets improving generalization.
Distill model into itself.
Ensemble predictions for consistency.
Model evaluates its own outputs.
Architectures designed for interpretability.
Use gates controlled by input.
Temperature rise due to buried oxide.
Self-heating in interconnects couples electrical and thermal effects influencing reliability.
Model temperature rise in device.
Self-inspection has operators check their own work immediately.
Generate instructions from model.
Self-Instruct generates training data by prompting models to create instructions and responses.
Self-monitoring tracks progress evaluating whether actions advance toward goals.
Self-paced learning automatically selects training samples based on current model capability starting with easier examples.
Let model select training order.
Model decides when and what to retrieve.
Self-RAG adaptively retrieves only when needed using learned retrieval decisions.
Model decides when it needs to retrieve documents vs generate from memory.
Iteratively improve generations.
Self-refine iteratively generates critiques and improvements without external feedback.
Learn without ground truth.
Learn disentangled representations without labels.
Self-supervised graph neural networks learn representations through pretext tasks without labeled data.
Learn normal patterns to detect anomalies.
Learn depth without labels.
DINO MAE BEiT methods.
Asynchronous circuits less sensitive to variation.
Self-training is a semi-supervised method where models generate pseudo-labels for unlabeled data to augment training sets iteratively.
Use model predictions on unlabeled data as training labels.
Model checks its own answers.
Alternative molecular representation.
Self-normalizing activation.
Semantic attention in heterogeneous GNNs learns importance weights for different metapaths or relation types.
Semantic caching matches semantically similar queries to cached results.
Split based on semantic coherence.
Semantic chunking creates segments based on topic boundaries rather than fixed size.
Search by meaning not keywords.
Semantic segmentation maps guide generation controlling regional content.
Find meaningful directions in latent space.
Semantic editing manipulates specific attributes in latent space for targeted modifications.
Object boundaries emerge without labels.
Encode meaning.
Microsoft's SDK for orchestrating AI plugins.
Semantic Kernel is Microsoft SDK for AI orchestration. Plugins, planners. Enterprise focus.
Semantic memory stores factual knowledge and learned concepts without temporal tags.
Convert natural language to formal representations.
Identify who did what to whom.
Search based on meaning similarity using embeddings rather than keyword matching.