Glossary

What is AI Inference?
AI Inference is the phase where a trained model generates predictions or outputs from new input data. It’s often the most compute-intensive and cost-sensitive part of deployment.

What is AI Training?
AI Training is the process where a model learns from data by adjusting internal weights to improve prediction accuracy through repeated optimization.

What is a GPU compared to a TPU?
A GPU (Graphics Processing Unit) is optimized for parallel processing and widely used in AI training. A TPU (Tensor Processing Unit) is a custom chip by Google designed specifically to accelerate machine learning.

What is Backpropagation?
Backpropagation is an algorithm that adjusts neural network weights by calculating and distributing errors backward through the model after each prediction.

What is Gradient Descent?
Gradient Descent is an optimization technique used to minimize a model’s error by iteratively updating weights based on the slope of the loss function.

What is the Cost of Delivery in AI?
The Cost of Delivery refers to the energy, infrastructure, and compute resources required to serve AI outputs in real-time, including inference compute, bandwidth, and storage.

What is Quantization
Quantization reduces the precision of model weights (e.g., from 32-bit to 8-bit) to shrink models and accelerate inference without major accuracy loss.

What is a Vector Database?
A Vector Database stores high-dimensional embeddings, enabling fast similarity searches for AI applications like recommendation, semantic search, and RAG.

What is Latency in AI?
Latency refers to the time it takes for an AI system to respond with an output after receiving an input—critical for real-time applications.

What is an AI Accelerator?
An AI Accelerator is a specialized chip (like a GPU, TPU, or ASIC) designed to optimize the performance of AI workloads, especially matrix-heavy operations.

What is Embedding in AI?
An Embedding is a numerical representation of text, image, or other data that preserves semantic meaning—used to power search, recommendations, and similarity comparisons.

What is Explainability in AI?
Explainability refers to understanding how and why an AI made a decision, increasing transparency and trust—especially in regulated or high-stakes settings.

What is Bias in AI?
Bias in AI refers to unfair patterns in model outputs caused by skewed or incomplete training data.

What is Attribution?
Attribution tracks the influence of specific inputs, data sources, or model components on a given AI output—used in auditing, compliance, and model debugging.

What is a Neural Network?
A Neural Network is a layered model inspired by the human brain, built from artificial neurons that process data and learn patterns.

What is a Large Language Model (LLM)?
An LLM is an AI system trained on vast text corpora to understand, predict, and generate human-like language for tasks like chat, summarization, and Q&A.

What is an Activation Function?
An Activation Function controls whether a neuron in a neural network should ‘fire’ based on its input, enabling non-linear decision-making.

What is Prompt Engineering?
Prompt Engineering is the art and science of crafting inputs to guide AI models toward better, more accurate, or more creative outputs.

What is a Parameter in an AI Model?
A Parameter is a learned value in a model (like a weight) that gets updated during training and influences how inputs are transformed into outputs.

What is an Epoch in AI Training?
An Epoch is one complete pass through the entire training dataset. Multiple epochs are often used to improve model accuracy.

What is a Loss Function?
A Loss Function measures the gap between predicted and actual values, guiding model optimization during training.

What is a Token in NLP?
A Token is the smallest unit of text processed by a language model—usually a word, subword, or character depending on the tokenizer used.

What is Multi-Modality in AI?
Multi-Modality refers to models that process and understand multiple types of data—text, images, audio, video—often at once.

What is a Language Tokenizer?
A Tokenizer breaks down text into smaller chunks (tokens) so an AI model can process and understand them efficiently during training or inference.

What is Pre-Processing?
Pre-Processing is the stage where raw data is cleaned and structured before training. It includes normalization, tokenization, and removing duplicates.

What is Model Compression?
Model Compression reduces the size of an AI model to improve efficiency, speed up inference, and enable deployment on smaller devices.

What is Fine-Tuning?
Fine-Tuning adjusts a pre-trained model using new, domain-specific data to improve accuracy for a specific task without retraining from scratch.

What is Transfer Learning?
Transfer Learning reuses knowledge from a model trained on one task and applies it to a new task—saving time, compute, and data.

What is Overfitting vs. Underfitting?
Overfitting happens when a model learns training data too precisely and fails to generalize. Underfitting occurs when a model is too simplistic and misses important patterns.

What is Supervised Learning vs. Unsupervised Learning?
Supervised Learning uses labeled data to train a model, while Unsupervised Learning finds patterns in unlabeled data, often for clustering or anomaly detection.

What is a CNN compared to an RNN?
CNNs are designed for spatial data like images; RNNs process sequential data like text or time series.

What is Named Entity Recognition (NER) vs. Sentiment Analysis?
NER identifies and categorizes names (people, places, orgs); Sentiment Analysis determines emotional tone (positive, negative, neutral) in text.

What is Speech-to-Text vs. Text-to-Speech?
Speech-to-Text converts spoken audio into written words. Text-to-Speech does the reverse, turning text into synthesized voice.

What is a Pre-Trained Model vs. a Custom Model?
A Pre-Trained Model is trained on general data and can be adapted for new use cases. A Custom Model is built and trained from scratch for a specific task.