AI・機械学習英語：必須ボキャブラリーとフレーズ

"training data"

The labeled dataset used to teach a machine learning model to make predictions.

"The quality of the training data directly impacts the model's accuracy on unseen examples."

"overfitting"

When a model learns the training data too well and fails to generalize to new data.

"We noticed overfitting when the validation loss started increasing while training loss kept decreasing."

"feature engineering"

The process of selecting, transforming, and creating input variables to improve model performance.

"Good feature engineering often matters more than choosing the right algorithm."

"gradient descent"

An optimization algorithm that iteratively adjusts model parameters to minimize the loss function.

"We switched from batch gradient descent to stochastic gradient descent to speed up training."

"hyperparameter tuning"

Searching for the optimal configuration settings that control how a model is trained.

"After hyperparameter tuning, we reduced the validation error by nearly 12 percent."

"loss function"

A mathematical function that measures how far a model's predictions are from the true values.

"For multi-class classification tasks, we typically use cross-entropy as the loss function."

"inference"

Running a trained model on new data to generate predictions or outputs.

"The model achieves sub-10ms inference latency when deployed on our GPU cluster."

"ground truth"

The correct, verified labels or outcomes used to evaluate a model's predictions.

"We hired domain experts to annotate the ground truth labels for our medical imaging dataset."

"transformer architecture"

A neural network design based on self-attention mechanisms, widely used for language and vision tasks.

"The transformer architecture revolutionized NLP by enabling models to capture long-range dependencies efficiently."

"fine-tuning"

Adapting a pre-trained model to a specific downstream task by continuing training on task-specific data.

"We fine-tuned a pre-trained BERT model on our customer support dataset to improve intent classification."

"embedding"

A dense numerical vector representation of discrete objects such as words, sentences, or images.

"Semantic search relies on embeddings to find documents that are conceptually similar to a query."

"attention mechanism"

A technique that allows a model to focus on the most relevant parts of the input when producing an output.

"Visualizing the attention mechanism helped us understand which tokens the model relied on most."

"backpropagation"

The algorithm used to compute gradients and update weights by propagating errors backward through a network.

"Vanishing gradients during backpropagation were a major problem before residual connections were introduced."

"batch normalization"

A technique that normalizes layer inputs to stabilize and accelerate neural network training.

"Adding batch normalization layers allowed us to train with a much higher learning rate."

"dropout"

A regularization technique that randomly deactivates neurons during training to prevent overfitting.

"We applied a dropout rate of 0.3 to the fully connected layers to improve generalization."

"convolutional neural network"

A neural network architecture particularly effective for processing grid-like data such as images.

"The convolutional neural network automatically learned spatial features without manual feature extraction."

"tokenization"

The process of splitting text into smaller units (tokens) that a model can process.

"The tokenizer splits the sentence into subword units, which helps handle rare vocabulary effectively."

"prompt engineering"

Designing and optimizing input prompts to elicit desired outputs from a language model.

"Careful prompt engineering reduced hallucinations and improved the factual accuracy of our chatbot."

"context window"

The maximum number of tokens a model can process at one time as input.

"We need to chunk long documents because they exceed the model's 8,000-token context window."

"hallucination"

When a language model generates factually incorrect or fabricated information with apparent confidence.

"We implemented a retrieval-augmented generation pipeline to reduce hallucinations in the product FAQ bot."

"semantic similarity"

A measure of how closely the meaning of two pieces of text relates to each other.

"We use cosine similarity between sentence embeddings to compute semantic similarity for duplicate detection."

"named entity recognition"

An NLP task that identifies and classifies named entities such as people, organizations, and locations in text.

"Named entity recognition was used to automatically extract company names from financial news articles."

"retrieval-augmented generation"

A technique that combines a retrieval system with a generative model to ground outputs in external documents.

"Retrieval-augmented generation significantly improved the accuracy of our internal knowledge base assistant."

"zero-shot learning"

A model's ability to perform a task it was not explicitly trained on, using only a description or examples in the prompt.

"The large language model classified customer sentiment with zero-shot learning and no labeled training examples."

"model drift"

The degradation of a deployed model's performance over time as real-world data patterns change.

"We set up automated alerts to detect model drift before it affects customer-facing predictions."

"A/B testing"

A controlled experiment comparing two model versions on real traffic to determine which performs better.

"We ran an A/B test for two weeks before rolling out the new recommendation model to all users."

"feature store"

A centralized repository for storing, sharing, and serving machine learning features across teams.

"The feature store eliminated duplicate computation and ensured consistency between training and serving pipelines."

"model registry"

A centralized repository for tracking, versioning, and managing trained ML models throughout their lifecycle.

"Every experiment is logged to the model registry so the team can reproduce any previous result."

"data pipeline"

An automated workflow that ingests, transforms, and delivers data for model training or serving.

"The data pipeline processes over 10 million events per day and feeds the real-time recommendation engine."

"canary deployment"

A strategy for gradually rolling out a new model version to a small percentage of traffic before full release.

"We used canary deployment to expose 5% of users to the new model and monitored error rates closely."

"latency"

The time it takes for a deployed model to return a prediction after receiving a request.

"We optimized the inference pipeline to bring latency down from 200ms to under 50ms."

"model serving"

The infrastructure and processes that expose a trained model as a service for real-time or batch predictions.

"We containerized our model serving layer using Docker to ensure consistent environments across deployments."

"algorithmic bias"

Systematic and unfair discrimination in model outputs caused by biased training data or flawed design choices.

"Auditing the hiring model revealed algorithmic bias against candidates from certain demographic groups."

"explainability"

The degree to which humans can understand and interpret why an AI model produced a specific output.

"Regulators required the bank to provide explainability reports for every automated credit decision."

"fairness metric"

A quantitative measure used to evaluate whether a model treats different groups equitably.

"The team evaluated multiple fairness metrics to ensure the model did not disadvantage minority classes."

"data governance"

The policies, processes, and standards that ensure data is accurate, secure, and used appropriately.

"Strong data governance practices were required before the healthcare AI could receive regulatory approval."

"consent"

The explicit permission given by individuals for their data to be collected and used for model training.

"Users must provide informed consent before their interaction data can be used to improve the recommendation system."

"responsible AI"

A framework for developing and deploying AI systems that are safe, fair, transparent, and accountable.

"Our responsible AI policy requires every model to undergo a bias audit before going to production."

"model card"

A document accompanying a published AI model that describes its intended use, limitations, and performance metrics.

"Publishing a model card helped external researchers understand the model's training conditions and known failure modes."

"differential privacy"

A mathematical technique that adds statistical noise to datasets so individual records cannot be identified.

"We applied differential privacy when training on sensitive user data to comply with GDPR requirements."

"ablation study"

An experiment that systematically removes components of a model to measure each component's contribution.

"The ablation study showed that removing the pre-training step reduced accuracy by 8 percentage points."

"benchmark"

A standardized evaluation dataset or task used to compare different models on a common scale.

"The new model achieves state-of-the-art results on the GLUE benchmark across all nine tasks."

"baseline"

A simple reference model used to measure whether a more complex approach actually provides improvement.

"Our baseline was a logistic regression model; the deep learning approach improved F1 by 15 points over baseline."

"state of the art"

The highest level of performance currently achieved on a given task or benchmark.

"Our architecture achieves state-of-the-art performance on three out of four standard computer vision benchmarks."

"reproducibility"

The ability to replicate an experiment's results using the same data, code, and methodology.

"The reproducibility crisis in ML motivated us to release all training scripts and random seeds alongside the paper."

"cross-validation"

A technique that evaluates model performance by splitting data into multiple training and validation folds.

"Five-fold cross-validation gave us a more reliable estimate of generalization performance than a single train-test split."

"statistical significance"

A measure indicating that an observed result is unlikely to have occurred by chance alone.

"The improvement in accuracy was statistically significant with a p-value below 0.01 across all test conditions."

"peer review"

The process by which research papers are evaluated by independent experts in the field before publication.

"The paper passed peer review at NeurIPS after we addressed the reviewers' concerns about evaluation methodology."