AI Factory Glossary

jailbreaking attempts, ai safety

Try to bypass safety measures.

jit compilation, jit, model optimization

Just-in-time compilation translates models to optimized machine code at runtime.

jit manufacturing, jit, supply chain & logistics

Just-In-Time manufacturing minimizes inventory by synchronizing material delivery with production needs reducing carrying costs.

jitter, llm optimization

Jitter adds randomness to retry timing preventing synchronized retries.

jodie, jodie, graph neural networks

JODIE learns embeddings for dynamic graphs through coupled recurrent models predicting future interactions.

joint distribution adaptation, domain adaptation

Align both marginal and conditional.

joint energy-based models, jem, generative models

Hybrid discriminative-generative model.

jsma, jsma, ai safety

Target specific features to perturb.

json mode, json, llm optimization

JSON mode ensures outputs are valid JSON objects or arrays.

jt-vae, jt-vae, graph neural networks

Junction Tree VAE generates valid molecular graphs by constructing molecules from chemical substructures.

junction depth control,diffusion

Precisely control depth of doped regions via implant and diffusion.

junction tree vae, chemistry ai

Hierarchical molecular generation.

k-anonymity, training techniques

K-anonymity ensures each record is indistinguishable from at least k-1 others.

k-wl test, graph neural networks

k-dimensional Weisfeiler-Lehman test extends standard WL providing stronger graph isomorphism testing.

kaizen event, manufacturing operations

Kaizen events are focused improvement workshops targeting specific problems over short duration.

kaizen suggestion, quality & reliability

Kaizen suggestions propose incremental improvements from frontline workers.

kaizen, manufacturing operations

Kaizen is continuous incremental improvement philosophy engaging all employees.

kalman filter, time series models

Kalman filter provides optimal linear state estimation for Gaussian state space models through recursive prediction and update steps.

kanban, supply chain & logistics

Kanban is a pull-based inventory management system using visual signals to trigger material replenishment based on consumption.

kernel fusion, model optimization

Kernel fusion merges adjacent GPU operations eliminating intermediate memory writes.

kirkendall voids, failure analysis advanced

Kirkendall voids form at bond interfaces from differential diffusion rates weakening interconnections.

knn-lm (k-nearest neighbor language model),knn-lm,k-nearest neighbor language model,llm architecture

Augment LM with nearest neighbor retrieval.

knowledge distillation for edge, edge ai

Distill large model to small for edge.

knowledge distillation variants, model compression

Different distillation approaches.

knowledge distillation, model optimization

Knowledge distillation trains compact student models to mimic larger teacher model predictions.

knowledge distillation,model optimization

Train small student model to mimic large teacher model.

knowledge editing, model editing

Modify specific facts in models.

knowledge editing,model training

Update model's factual knowledge without full retraining.

knowledge graph embedding, graph neural networks

Knowledge graph embeddings map entities and relations to vectors for link prediction and reasoning.

knowledge graph embeddings (advanced),knowledge graph embeddings,advanced,graph neural networks

Learn representations of KG entities and relations.

knowledge localization, explainable ai

Find where knowledge is stored.

knowledge neurons, explainable ai

Neurons storing factual knowledge.

kolmogorov-arnold networks (kan),kolmogorov-arnold networks,kan,neural architecture

Networks using learnable activation functions on edges.

kosmos,multimodal ai

Multimodal large language model from Microsoft.

kv cache, kv, llm optimization

KV cache stores key-value attention states avoiding recomputation during generation.

kv cache,llm architecture

Cached key-value pairs from previous tokens to speed up autoregressive generation.

l-diversity, training techniques

L-diversity requires at least l diverse values for sensitive attributes within groups.

l-infinity attacks, ai safety

Bound maximum per-pixel change.

l0 attacks, l0, ai safety

Limit number of pixels changed.

l2 attacks, l2, ai safety

Bound Euclidean distance.

label flipping, ai safety

Change labels in training data.

label propagation on graphs, graph neural networks

Semi-supervised learning via graph structure.

label smoothing, machine learning

Soften hard labels for regularization.

lagrangian neural networks, scientific ml

Learn dynamics from Lagrangian mechanics.

lamda (language model for dialogue applications),lamda,language model for dialogue applications,foundation model

Google's conversational AI model.

landmark attention, llm architecture

Efficient attention for long context.

landmark attention,llm architecture

Attend to selected landmark tokens for efficiency.

langchain, ai agents

LangChain provides framework for building LLM applications with chains and agents.

langchain,framework

Framework for building LLM applications with chains and agents.

langchain,framework,llm

LangChain is framework for LLM applications. Chains, agents, tools. Popular ecosystem.