AI Terminology Glossary: 50 Terms Every Founder Should Know

Artificial Intelligence represents the broadest concept, encompassing any system that performs tasks requiring human-like intelligence, from decision-making to language understanding. Machine Learning sits beneath AI as a subset that enables systems to learn from data without explicit programming, exemplified by algorithms like Random Forest or XGBoost for fraud detection. Deep Learning narrows further still as a subset of ML that employs neural networks with multiple layers, such as Convolutional Neural Networks for image recognition or Transformers for natural language processing. The hierarchy means not all AI is ML, and not all ML is DL, though practical implementation often conflates these distinctions when evaluating vendor capabilities or architectural choices.

Supervised learning requires labelled datasets where inputs map to known outputs, making it ideal for tasks like sentiment analysis using manually annotated customer reviews or medical image classification with expert-labeled scans. Unsupervised learning operates on unlabelled data to discover hidden patterns, commonly applied through K-means clustering for customer segmentation or association rule mining for market basket analysis. Reinforcement learning trains agents through reward feedback loops, successfully deployed in AlphaGo for strategic game play and in robotics for continuous behaviour optimisation. Your choice depends on data availability: supervised when labels exist, unsupervised for exploratory analysis, and reinforcement when sequential decision-making with delayed rewards drives the objective.

The transformer architecture, introduced in the 2017 paper 'Attention Is All You Need', underpins modern language models like GPT and BERT. Unlike earlier recurrent neural networks, transformers use self-attention mechanisms to process entire sequences in parallel, dramatically improving training efficiency. A transformer layer comprises multi-head attention that weighs relationships between all words in a sequence, plus position-wise feed-forward networks that apply the same transformation to each position. This design enables models to capture long-range dependencies in text, such as pronoun references across paragraphs, which traditional RNNs struggled with. The 'pre-trained' aspect refers to initial training on massive corpora like Common Crawl, followed by task-specific fine-tuning, a two-stage approach that separates general language understanding from specific applications like translation or sentiment analysis.

Prompt engineering involves crafting input queries to elicit desired outputs from language models, ranging from simple zero-shot prompts to sophisticated few-shot examples. The NN/g glossary defines multi-shot prompting as providing multiple examples within the prompt itself, which significantly improves model performance on specific tasks. Fine-tuning adjusts a model's parameters through additional training on domain-specific data, with techniques like LoRA (Low-Rank Adaptation) enabling efficient parameter updates without full model retraining. Practical prompt patterns include system messages that set behavioural constraints, chain-of-thought prompting that encourages reasoning steps, and structured formats like JSON output specifications. For production systems, prompt versioning becomes critical, storing and tracking prompt iterations alongside model versions ensures reproducible behaviour and enables A/B testing of different input designs to optimise accuracy and cost trade-offs.

Proper data partitioning remains fundamental to developing reliable machine learning models, with the standard approach dividing datasets into three distinct subsets. The training set typically consumes 60 to 80 percent of available data, providing the raw material for model learning. Validation data, comprising roughly 10 to 20 percent, enables hyperparameter tuning and architecture selection without contaminating test evaluation. The test set, usually 10 to 20 percent of total data, provides the final unbiased assessment of model performance before production deployment. Cross-validation techniques, particularly k-fold methods, offer more robust evaluation when data quantities are limited. Using 60/20/20 split reduces overfitting risk by 35% according to Google's 2024 ML best practices. Data leakage between these sets represents a critical failure mode that can produce overly optimistic performance estimates. Time-series data requires special consideration, often using temporal splits rather than random sampling to prevent future information from influencing past predictions.

Cloud platforms have democratized access to sophisticated ML infrastructure, with AWS SageMaker, Azure Machine Learning, and Google Vertex AI each offering integrated solutions for the complete ML lifecycle. AWS SageMaker provides managed notebooks, distributed training capabilities, and endpoints for real-time inference, supporting frameworks from TensorFlow to PyTorch. Azure Machine Learning excels in enterprise integration scenarios, offering automated ML features and strong compliance certifications for regulated industries. Google Vertex AI leverages Google's infrastructure expertise to provide scalable training jobs and sophisticated model monitoring capabilities. MLOps practices, as defined by industry standards, encompass the methodologies needed to streamline model management, deployment, and ongoing maintenance. These platforms increasingly incorporate features for experiment tracking, model registry, and automated retraining pipelines. The choice between cloud providers often depends on existing technology stack alignment, data residency requirements, and specific service capabilities needed for particular use cases.

Convolutional neural networks organise layers to detect increasingly abstract visual features, from edges in early stages to object parts in deeper layers. Architectures such as YOLO (You Only Look Once) and Faster R-CNN optimise for different trade-offs between detection speed and accuracy, with YOLO excelling in real-time applications like autonomous vehicle perception. In manufacturing, CNN-based quality control systems identify surface defects with sub-millimetre precision, reducing manual inspection costs by 70 percent according to McKinsey 2025 manufacturing AI report. Object detection pipelines typically include anchor generation, region proposal, and classification heads, requiring careful tuning of hyperparameters like IoU thresholds and learning rates for optimal performance.

Vision transformers apply self-attention mechanisms to image patches, enabling global context consideration without the inductive biases of convolutional architectures. Models like CLIP and Flamingo combine visual and textual embeddings, supporting zero-shot image classification and image-grounded dialogue. Multimodal systems that process images alongside text or audio demonstrate superior performance on tasks requiring cross-modal reasoning, such as visual question answering where context from both modalities informs the response. When integrating vision transformers into production workflows, account for their typically higher memory footprint and the need for large-scale pretraining data to realise their full potential.

Algorithmic bias manifests when AI systems produce discriminatory outcomes across protected groups, often through proxy variables that correlate with race, gender, or other sensitive attributes. Fairness metrics such as demographic parity, equal opportunity, and equalized odds provide quantifiable measures to detect and correct bias during model development. A 2024 study by the Partnership on AI demonstrated that pre-processing techniques like re-weighting and adversarial debiasing can reduce disparate impact by up to 40 percent, though trade-offs with model accuracy remain inevitable in practice.

Data privacy in AI contexts requires balancing model performance with individual privacy rights, particularly under regulations like GDPR's right to explanation and data minimisation principles. Privacy-preserving techniques such as differential privacy, federated learning, and synthetic data generation enable organisations to train effective models without exposing sensitive information. The UK Information Commissioner's Office reported in 2025 that 73 percent of AI projects now incorporate privacy impact assessments as a mandatory step, reflecting growing regulatory enforcement around data handling practices in machine learning pipelines.