privacy-preserving training,privacy
Techniques to protect data privacy during training.
212 technical terms and definitions
Techniques to protect data privacy during training.
For sensitive data, run models on-prem or in a VPC. No logs to third-party clouds, strict access control, encryption, and auditing of all requests.
Use proprietary data for pre-training.
Use extra info during training only.
Deterministic ODE with same marginals as SDE.
Probe card repair replaces damaged needles cleans contamination and realigns tips to restore measurement accuracy.
Train classifiers on representations.
Train classifiers on internal representations to see what information is encoded.
AI-assisted procedural content creation.
Process optimization reduces energy by improving efficiency cycle times and yields.
Product carbon footprint quantifies greenhouse gas emissions attributable to specific products.
Product quantization decomposes vectors into subvectors quantized independently for compression.
Product stewardship extends manufacturer responsibility to entire product lifecycle including design use and end-of-life management.
Production scheduling sequences manufacturing operations optimizing throughput and resource utilization.
Analyze performance bottlenecks.
Automatically generate programs from specifications.
Generate code to solve reasoning tasks.
Program-aided language models offload computation to code interpreters.
Gradually reduce diffusion steps.
Gradually increase resolution.
Progressive growing gradually increases resolution during training stabilizing GAN training.
Gradually increase resolution during GAN training.
Add capacity for new tasks.
Add new capacity for each task.
Progressive shrinking trains once-for-all networks by gradually incorporating smaller sub-networks starting from the largest architecture.
Prompt caching reuses processed prompts reducing latency and cost for repeated prefixes.
Use output of one prompt as input to next.
Prompt chaining sequences multiple prompts passing outputs as subsequent inputs.
Split long prompts.
Vector representation of prompts.
Adversarial inputs to subvert model.
Techniques to prevent injection.
Prompt injection inserts malicious instructions into prompts attempting to hijack model behavior.
Attack where user input tricks model into ignoring instructions or leaking info.
Trick model into revealing system prompts or instructions.
Check prompts before processing.
Cut off excess tokens.
Emphasize prompt parts differently.
Edit images by modifying text prompts.
Prompt-to-prompt editing modifies images by adjusting text prompts while preserving structure.
Generate property tests.
Prophet is an additive time series model decomposing signals into trend seasonality and holiday components with automatic changepoint detection.
Proprietary models have restricted access to weights and implementation.
Identify sensitive medical data.
Design proteins with desired properties.
Infer protein function from descriptions.
Predict 3D protein structures (AlphaFold).
Interaction between protein and small molecule.
Learn representative prototypes for each class.
Search directly on target hardware.