--- title: "AI glossary, in plain English" description: "192 AI and ML terms defined in plain English — from ablation study to zero-shot — organized by fundamentals, architecture, training, and research." canonical_url: "https://fanout.sh/ai/glossary" md_url: "https://fanout.sh/ai/glossary.md" access: "public" --- # AI glossary, in plain English 192 AI and ML terms defined in plain English — from ablation study to zero-shot — organized by fundamentals, architecture, training, and research. ## Terms - **Ablation Study:** An experiment where components of a model are systematically removed or disabled to measure each part's contribution to overall performance. _(Research)_ - **Activation Function:** A non-linear function applied to a neuron's output that allows neural networks to learn complex patterns. Common examples include ReLU, sigmoid, and tanh. _(Fundamentals)_ - **Adam Optimizer:** An adaptive learning rate optimizer that combines momentum and RMSProp. It maintains per-parameter learning rates and is the default choice for most deep learning tasks. _(Training)_ - **Adversarial Examples:** Inputs deliberately crafted with small perturbations to fool a model into making incorrect predictions, exposing vulnerabilities in learned representations. _(Research)_ - **API:** Application Programming Interface — the contract through which software systems communicate. In ML deployment, APIs expose model inference as HTTP endpoints. _(Deployment)_ - **Attention:** A mechanism that lets a model dynamically focus on relevant parts of the input when producing each element of the output, weighting tokens by their contextual importance. _(Architecture)_ - **Backpropagation:** The algorithm used to compute gradients of the loss with respect to every weight by applying the chain rule backward through the network. _(Fundamentals)_ - **Batch Inference:** Processing multiple inputs together in a single forward pass rather than one at a time, improving throughput and hardware utilization during deployment. _(Deployment)_ - **Batch Size:** The number of training examples processed together in one forward and backward pass. Larger batches give more stable gradients but require more memory. _(Fundamentals)_ - **Benchmark:** A standardized dataset and evaluation protocol used to compare model performance across different approaches, such as GLUE, ImageNet, or MMLU. _(Research)_ - **Bias:** A learnable constant added to a neuron's weighted sum before the activation function. It shifts the activation, allowing the model to fit data that doesn't pass through the origin. _(Fundamentals)_ - **BLEU Score:** Bilingual Evaluation Understudy — a metric that evaluates machine-generated text by measuring n-gram overlap with reference translations. _(Research)_ - **CNN:** Convolutional Neural Network — an architecture that uses learnable spatial filters to detect local patterns like edges and textures, widely used in computer vision. _(Architecture)_ - **Cross-Entropy:** A loss function that measures the difference between two probability distributions. It is the standard loss for classification tasks and penalizes confident wrong predictions heavily. _(Fundamentals)_ - **Data Augmentation:** Techniques that create modified copies of training data (flipping, cropping, noise injection) to increase dataset diversity and reduce overfitting. _(Training)_ - **Dataset:** A structured collection of examples used for training, validation, or testing a machine learning model. Each example typically consists of input features and an optional label. _(Fundamentals)_ - **Decoder:** The part of a sequence model that generates the output sequence token-by-token, typically using masked self-attention and cross-attention to the encoder. _(Architecture)_ - **Distributed Training:** Splitting model training across multiple GPUs or machines using data parallelism or model parallelism to reduce wall-clock time for large models. _(Training)_ - **Docker:** A containerization platform that packages an application and its dependencies into a portable image, ensuring consistent environments from development to production. _(Deployment)_ - **Dropout:** A regularization technique that randomly sets a fraction of neuron activations to zero during training, forcing the network to learn redundant representations. _(Fundamentals)_ - **Early Stopping:** A training strategy that halts optimization when validation performance stops improving, preventing the model from overfitting to the training set. _(Training)_ - **Edge Deployment:** Running ML models directly on end-user devices (phones, IoT sensors, browsers) rather than in the cloud, reducing latency and bandwidth requirements. _(Deployment)_ - **Embedding:** A dense, low-dimensional vector representation that captures semantic meaning of discrete inputs like words, tokens, or categories. _(Fundamentals)_ - **Emergent Abilities:** Capabilities that appear in large language models only above a certain scale and are not present in smaller models, such as chain-of-thought reasoning. _(Research)_ - **Encoder:** The component of a model that processes the full input sequence and produces contextual representations used by downstream layers or a decoder. _(Architecture)_ - **Epoch:** One complete pass through the entire training dataset. Models are typically trained for multiple epochs until convergence. _(Fundamentals)_ - **Feature:** An individual measurable property of the input data used by the model to make predictions, such as pixel values, word frequencies, or sensor readings. _(Fundamentals)_ - **Feed-Forward Network:** A neural network where information flows strictly from input to output with no cycles. In transformers, a two-layer FFN is applied to each token after attention. _(Architecture)_ - **Few-Shot Learning:** The ability of a model to perform a new task given only a handful of examples, typically provided as in-context demonstrations in the prompt. _(Research)_ - **Fine-Tuning:** Continuing the training of a pre-trained model on a smaller, task-specific dataset to adapt its learned representations for a particular downstream task. _(Training)_ - **GAN:** Generative Adversarial Network — two networks (generator and discriminator) trained in competition, where the generator learns to produce realistic data that the discriminator cannot distinguish from real samples. _(Architecture)_ - **GGUF:** A binary file format for storing quantized LLM weights optimized for fast CPU and GPU inference, commonly used with llama.cpp and local model runners. _(Deployment)_ - **Gradient:** The vector of partial derivatives of the loss function with respect to model parameters. It indicates the direction and magnitude of steepest increase in loss. _(Fundamentals)_ - **Gradient Accumulation:** A technique that sums gradients over multiple mini-batches before updating weights, effectively simulating a larger batch size when GPU memory is limited. _(Training)_ - **Gradient Descent:** The core optimization algorithm that iteratively adjusts parameters in the direction opposite to the gradient to minimize the loss function. _(Fundamentals)_ - **GRU:** Gated Recurrent Unit — a simplified RNN variant that uses reset and update gates to control information flow, offering similar performance to LSTMs with fewer parameters. _(Architecture)_ - **Hallucination:** When a language model generates text that sounds plausible but is factually incorrect or fabricated, not grounded in its training data or provided context. _(Research)_ - **Hyperparameter:** A configuration value set before training begins (learning rate, batch size, number of layers) that controls the training process but is not learned from data. _(Fundamentals)_ - **Inception:** A CNN architecture that applies multiple filter sizes in parallel within each layer and concatenates the results, capturing patterns at different spatial scales. _(Architecture)_ - **Kubernetes:** An open-source container orchestration platform that automates deploying, scaling, and managing containerized ML services across clusters of machines. _(Deployment)_ - **KV Cache:** Key-Value Cache — a memory optimization for autoregressive transformers that stores previously computed key and value tensors so they are not recomputed at each generation step. _(Architecture)_ - **Label:** The ground-truth answer or target value associated with a training example, used to compute the loss and guide the model's learning. _(Fundamentals)_ - **Learning Rate:** A scalar that controls the step size of each parameter update during gradient descent. Too high causes instability; too low causes slow convergence. _(Fundamentals)_ - **LoRA:** Low-Rank Adaptation — a parameter-efficient fine-tuning method that freezes the original weights and injects small trainable low-rank matrices into each layer. _(Training)_ - **Loss Function:** A mathematical function that quantifies the difference between a model's predictions and the true labels. Training aims to minimize this value. _(Fundamentals)_ - **LSTM:** Long Short-Term Memory — an RNN variant with gated cells that can learn long-range dependencies by selectively remembering or forgetting information over many time steps. _(Architecture)_ - **Mini-batch:** A small subset of the training dataset used in one iteration of gradient descent, balancing the noise of single-example updates with the cost of full-batch computation. _(Fundamentals)_ - **Mixed Precision:** A training technique that uses 16-bit floats for most computations while keeping a 32-bit master copy of weights, cutting memory usage and speeding up training on modern GPUs. _(Training)_ - **Mixture of Experts (MoE):** An architecture where multiple specialist sub-networks (experts) exist but only a subset is activated per input via a gating network, enabling massive model capacity with lower compute cost. _(Architecture)_ - **MLP:** Multi-Layer Perceptron — the simplest deep network: stacked fully connected layers with non-linear activations. It forms the feed-forward blocks inside transformers. _(Architecture)_ - **Multi-Head Attention:** Running multiple attention operations in parallel with different learned projections, then concatenating the results. This allows the model to attend to information from different representation subspaces. _(Architecture)_ - **Neuron:** The basic computational unit of a neural network that computes a weighted sum of its inputs, adds a bias, and applies an activation function. _(Fundamentals)_ - **Normalization:** Rescaling data or intermediate activations to a standard range (e.g., zero mean, unit variance) to stabilize and accelerate training. Common variants include batch norm and layer norm. _(Fundamentals)_ - **ONNX:** Open Neural Network Exchange — an open format for representing ML models that enables interoperability between frameworks like PyTorch, TensorFlow, and inference engines. _(Deployment)_ - **Overfitting:** When a model memorizes the training data (including noise) and performs well on training examples but poorly on unseen data, indicating a failure to generalize. _(Fundamentals)_ - **Perplexity:** A metric that measures how well a language model predicts a sample, defined as the exponentiated average cross-entropy. Lower perplexity indicates better prediction. _(Research)_ - **Pre-training:** The initial phase of training a large model on a broad dataset (e.g., next-token prediction on web text) to learn general representations before task-specific fine-tuning. _(Training)_ - **QLoRA:** Quantized LoRA — combines 4-bit quantization of the base model with LoRA adapters, enabling fine-tuning of very large models on a single consumer GPU. _(Training)_ - **Quantization:** Reducing the numerical precision of model weights (e.g., from 32-bit to 4-bit) to shrink model size and speed up inference with minimal accuracy loss. _(Deployment)_ - **Regularization:** Techniques (L1/L2 penalties, dropout, weight decay) that constrain model complexity to prevent overfitting and improve generalization to unseen data. _(Fundamentals)_ - **Residual Connection:** A skip connection that adds a layer's input directly to its output, enabling gradient flow through very deep networks and making optimization easier. _(Architecture)_ - **RLHF:** Reinforcement Learning from Human Feedback — a training method that aligns language model outputs with human preferences using a reward model trained on human comparisons. _(Training)_ - **RNN:** Recurrent Neural Network — a network that processes sequences by maintaining a hidden state that gets updated at each time step, enabling it to model temporal dependencies. _(Architecture)_ - **Scaling Laws:** Empirical power-law relationships showing that model performance improves predictably as compute, data, and parameter count increase, guiding resource allocation decisions. _(Research)_ - **Self-Attention:** An attention mechanism where queries, keys, and values all come from the same sequence, allowing each token to attend to every other token in the input. _(Architecture)_ - **Serving:** The infrastructure and process of making a trained model available for real-time or batch predictions in production, handling request routing, batching, and scaling. _(Deployment)_ - **Softmax:** A function that converts a vector of raw scores (logits) into a probability distribution where all values are positive and sum to one. _(Architecture)_ - **Tensor:** A multi-dimensional array that generalizes scalars, vectors, and matrices. Tensors are the fundamental data structure in deep learning frameworks like PyTorch and TensorFlow. _(Fundamentals)_ - **TensorRT:** NVIDIA's high-performance deep learning inference optimizer and runtime that applies graph optimizations, kernel fusion, and precision calibration for GPU deployment. _(Deployment)_ - **Tokenizer:** A component that splits raw text into tokens (subwords, words, or characters) and maps them to integer IDs that a language model can process. _(Deployment)_ - **Transfer Learning:** Reusing a model trained on one task (often a large general task) as the starting point for a different but related task, dramatically reducing data and compute requirements. _(Training)_ - **Transformer:** The dominant neural network architecture built on self-attention mechanisms rather than recurrence, enabling parallel processing of sequences and forming the backbone of modern LLMs. _(Architecture)_ - **U-Net:** An encoder-decoder CNN with skip connections between corresponding layers, originally designed for biomedical image segmentation and now widely used in diffusion models. _(Architecture)_ - **vLLM:** A high-throughput LLM serving engine that uses PagedAttention to efficiently manage GPU memory, enabling faster inference and higher concurrency for large language models. _(Deployment)_ - **Warm-up:** A training schedule that starts with a very small learning rate and gradually increases it over the first few hundred or thousand steps to stabilize early optimization. _(Training)_ - **Weight:** A learnable parameter in a neural network that scales the input signal at each connection. The collection of all weights defines what the model has learned. _(Fundamentals)_ - **Zero-Shot Learning:** A model's ability to perform a task it was never explicitly trained on by leveraging general knowledge acquired during pre-training, without any task-specific examples. _(Research)_ - **Accuracy:** The fraction of predictions a model gets correct. Simple but misleading on imbalanced datasets — a spam detector that always says 'not spam' gets 99% accuracy if 99% of emails are legitimate. _(Fundamentals)_ - **Bias-Variance Tradeoff:** The tension between a model that is too simple (high bias, underfitting) and too complex (high variance, overfitting). The sweet spot minimizes total error on unseen data. _(Fundamentals)_ - **Broadcasting:** Automatic expansion of tensor dimensions during element-wise operations so that tensors of different shapes can be combined without explicit copying. _(Fundamentals)_ - **Confusion Matrix:** A table showing true positives, false positives, true negatives, and false negatives for a classifier — the foundation for computing precision, recall, and F1. _(Fundamentals)_ - **Convolution:** A mathematical operation that slides a small learnable filter across an input (e.g., an image) to produce a feature map that detects local patterns like edges or textures. _(Fundamentals)_ - **Cosine Similarity:** A measure of similarity between two vectors based on the cosine of the angle between them, ranging from -1 (opposite) to 1 (identical direction). Widely used in embeddings and retrieval. _(Fundamentals)_ - **Cross-Validation:** A technique that splits data into k folds, trains on k-1 folds and validates on the remaining one, rotating through all folds to get a robust performance estimate. _(Fundamentals)_ - **Curse of Dimensionality:** The phenomenon where data becomes increasingly sparse as the number of features grows, making distance-based methods unreliable and requiring exponentially more data. _(Fundamentals)_ - **Dimensionality Reduction:** Techniques (PCA, t-SNE, UMAP) that project high-dimensional data into fewer dimensions while preserving structure, used for visualization and preprocessing. _(Fundamentals)_ - **Dot Product:** The sum of element-wise products of two vectors. Measures how aligned they are — the core operation behind attention scores, similarity, and linear layers. _(Fundamentals)_ - **F1 Score:** The harmonic mean of precision and recall, providing a single metric that balances both. Useful when classes are imbalanced and accuracy is misleading. _(Fundamentals)_ - **Feature Engineering:** The process of creating, selecting, or transforming input features to improve model performance. Often the highest-leverage activity in classical ML. _(Fundamentals)_ - **Inference:** Using a trained model to make predictions on new data. Distinct from training — inference is typically faster and doesn't update weights. _(Fundamentals)_ - **Kernel / Filter:** A small matrix of learnable weights used in convolution operations to detect specific patterns. Multiple kernels detect different features like edges, corners, and textures. _(Fundamentals)_ - **Latent Space:** A compressed, abstract representation space learned by a model where similar inputs map to nearby points. Used in autoencoders, VAEs, and diffusion models. _(Fundamentals)_ - **Logits:** Raw, unnormalized output scores from a model's final layer before applying softmax. Higher logits correspond to higher predicted probabilities. _(Fundamentals)_ - **Matrix Multiplication:** The fundamental operation in neural networks — multiplying input matrices by weight matrices. Every linear layer, attention head, and embedding lookup is a matmul. _(Fundamentals)_ - **One-Hot Encoding:** Representing a categorical variable as a binary vector with a single 1 and all other positions 0. Used for class labels before softmax and in sparse input representations. _(Fundamentals)_ - **Padding:** Adding zeros (or other values) around the border of an input before convolution to control the output spatial dimensions, typically preserving the original size. _(Fundamentals)_ - **Pooling:** A downsampling operation (max pooling or average pooling) that reduces spatial dimensions of feature maps while retaining the most important information. _(Fundamentals)_ - **Precision:** Of all the positive predictions the model made, how many were actually correct. High precision means few false positives. _(Fundamentals)_ - **Recall:** Of all the actual positives in the data, how many did the model correctly identify. High recall means few false negatives. _(Fundamentals)_ - **Scalar:** A single number — a zero-dimensional tensor. Loss values, learning rates, and individual weights are all scalars. _(Fundamentals)_ - **Stride:** The step size a convolutional filter moves across the input. Stride of 1 moves one pixel at a time; stride of 2 halves the spatial dimensions. _(Fundamentals)_ - **Temperature:** A scaling factor applied to logits before softmax that controls output randomness. Low temperature (< 1) makes predictions sharper; high temperature (> 1) makes them more uniform. _(Fundamentals)_ - **Underfitting:** When a model is too simple to capture the underlying patterns in the data, performing poorly on both training and test sets. _(Fundamentals)_ - **Vanishing Gradients:** A problem in deep networks where gradients become exponentially small as they propagate backward through many layers, causing early layers to stop learning. _(Fundamentals)_ - **Vector:** An ordered list of numbers — a one-dimensional tensor. In ML, inputs, weights, embeddings, and gradients are all vectors or collections of vectors. _(Fundamentals)_ - **Weight Initialization:** The strategy for setting initial parameter values before training. Poor initialization (all zeros, too large) can cause vanishing/exploding gradients. Xavier and He initialization are standard. _(Fundamentals)_ - **Autoencoder:** A network trained to compress input into a low-dimensional latent representation (encoder) and reconstruct it (decoder). Used for denoising, anomaly detection, and representation learning. _(Architecture)_ - **Autoregressive Model:** A model that generates output one token at a time, using all previously generated tokens as context for the next prediction. GPT and all decoder-only LLMs are autoregressive. _(Architecture)_ - **Batch Normalization:** Normalizing layer activations across the batch dimension to stabilize and accelerate training. Adds learnable scale and shift parameters after normalization. _(Architecture)_ - **Beam Search:** A decoding strategy that maintains the top-k most probable partial sequences at each step, balancing between greedy search and exhaustive search for better output quality. _(Architecture)_ - **BERT:** Bidirectional Encoder Representations from Transformers — a pre-trained encoder model that reads text in both directions simultaneously, excelling at understanding tasks like classification and NER. _(Architecture)_ - **Bottleneck Layer:** A layer with fewer neurons than surrounding layers that forces the network to compress information, learning a compact representation of the input. _(Architecture)_ - **Causal Masking:** A triangular mask applied in self-attention that prevents each position from attending to future positions, enforcing the autoregressive property in decoder models. _(Architecture)_ - **Cross-Attention:** An attention mechanism where queries come from one sequence (e.g., decoder) and keys/values from another (e.g., encoder), enabling the model to align and combine information across sequences. _(Architecture)_ - **Depthwise Separable Convolution:** A factorized convolution that applies a separate filter per input channel followed by a 1x1 pointwise convolution, drastically reducing parameters and compute compared to standard convolution. _(Architecture)_ - **Diffusion Model:** A generative model that learns to reverse a gradual noising process — trained to denoise, then generates new data by iteratively removing noise from random noise. Powers Stable Diffusion, DALL-E 3, etc. _(Architecture)_ - **Flash Attention:** An IO-aware exact attention algorithm that tiles the computation to minimize memory reads/writes, achieving 2-4x speedups without approximation. _(Architecture)_ - **GPT:** Generative Pre-trained Transformer — a decoder-only architecture trained with next-token prediction. The foundation of ChatGPT, GPT-4, and most modern LLMs. _(Architecture)_ - **Grouped Query Attention (GQA):** A variant of multi-head attention where groups of query heads share the same key/value heads, reducing KV cache memory at inference time with minimal quality loss. _(Architecture)_ - **Layer Normalization:** Normalizing activations across the feature dimension (per sample) rather than across the batch. Standard in transformers because it works with variable-length sequences and small batches. _(Architecture)_ - **Positional Encoding:** Information added to token embeddings that encodes each token's position in the sequence. Without it, transformers have no notion of word order since attention is permutation-invariant. _(Architecture)_ - **RMSNorm:** Root Mean Square Normalization — a simpler, faster alternative to Layer Norm that skips mean-centering and just divides by the RMS of activations. Used in LLaMA, Mistral, and Qwen. _(Architecture)_ - **RoPE:** Rotary Positional Embedding — encodes position by rotating query and key vectors in 2D subspaces, providing relative position information that generalizes well to longer sequences than seen during training. _(Architecture)_ - **Seq2Seq:** Sequence-to-Sequence — an encoder-decoder framework that maps an input sequence to an output sequence of potentially different length. Used in translation, summarization, and speech-to-text. _(Architecture)_ - **Speculative Decoding:** An inference optimization where a smaller draft model generates candidate tokens and a larger model verifies them in parallel, speeding up generation without changing output quality. _(Architecture)_ - **SwiGLU:** A gated activation function combining Swish and GLU that has become standard in modern LLMs (LLaMA, PaLM, Mistral). Outperforms ReLU and GELU in practice. _(Architecture)_ - **Top-k / Top-p Sampling:** Decoding strategies that restrict token selection: top-k picks from the k most likely tokens; top-p (nucleus) picks from the smallest set whose cumulative probability exceeds p. _(Architecture)_ - **Variational Autoencoder (VAE):** An autoencoder that learns a smooth, continuous latent space by regularizing the encoder to output a probability distribution, enabling generation of new samples by sampling from the latent space. _(Architecture)_ - **Vision Transformer (ViT):** A transformer applied to images by splitting them into fixed-size patches, linearly embedding each patch, and processing them with standard transformer blocks — proving transformers work beyond NLP. _(Architecture)_ - **AdamW:** A corrected version of Adam that decouples weight decay from the adaptive learning rate, providing better regularization. The default optimizer for most LLM training. _(Training)_ - **Catastrophic Forgetting:** The tendency of a neural network to lose previously learned knowledge when trained on a new task, a key challenge in continual learning and fine-tuning. _(Training)_ - **Checkpointing:** Periodically saving model weights and optimizer state during training so that training can be resumed from a saved point after interruptions or crashes. _(Training)_ - **Contrastive Learning:** A self-supervised technique that learns representations by pulling similar examples closer and pushing dissimilar examples apart in embedding space. Powers CLIP, SimCLR, etc. _(Training)_ - **Convergence:** The point during training where the loss stabilizes and additional updates no longer meaningfully improve performance. A model that hasn't converged needs more training. _(Training)_ - **Cosine Annealing:** A learning rate schedule that decreases the learning rate following a cosine curve from its initial value to near zero, often with periodic warm restarts. _(Training)_ - **Curriculum Learning:** A training strategy that presents examples in order of increasing difficulty, mimicking how humans learn. Can improve convergence speed and final performance. _(Training)_ - **DPO:** Direct Preference Optimization — an alternative to RLHF that directly optimizes a language model to match human preferences without training a separate reward model, using a simpler classification-style loss. _(Training)_ - **Gradient Checkpointing:** A memory optimization that recomputes intermediate activations during the backward pass instead of storing them all, trading compute time for significantly less GPU memory usage. _(Training)_ - **Gradient Clipping:** Capping gradient values to a maximum norm during training to prevent exploding gradients that cause unstable updates and loss spikes. _(Training)_ - **Instruction Tuning:** Fine-tuning a pre-trained LLM on a dataset of (instruction, response) pairs to make it follow user instructions rather than just completing text. _(Training)_ - **Knowledge Distillation:** Training a smaller student model to mimic the outputs (soft probabilities) of a larger teacher model, compressing knowledge into a more efficient architecture. _(Training)_ - **Learning Rate Scheduler:** A rule that adjusts the learning rate during training (e.g., step decay, cosine annealing, warm-up + decay). Proper scheduling is often as important as the optimizer choice. _(Training)_ - **Momentum:** An optimization technique that accumulates a moving average of past gradients to smooth updates and accelerate convergence, helping escape shallow local minima. _(Training)_ - **PEFT:** Parameter-Efficient Fine-Tuning — a family of methods (LoRA, prefix tuning, adapters) that update only a small fraction of model parameters, making fine-tuning of large models feasible on limited hardware. _(Training)_ - **PPO:** Proximal Policy Optimization — a policy gradient RL algorithm that uses clipped objective functions to prevent destructively large updates. The standard algorithm for RLHF in LLMs. _(Training)_ - **Pruning:** Removing unnecessary weights or neurons from a trained model to reduce size and inference cost. Can be unstructured (individual weights) or structured (entire channels/heads). _(Training)_ - **Reward Model:** A model trained on human preference data to predict which outputs humans would prefer. Used in RLHF to provide reward signals for policy optimization. _(Training)_ - **Self-Supervised Learning:** Training on unlabeled data by creating supervision from the data itself (e.g., predicting masked words, next tokens, or rotations). Powers pre-training of LLMs and vision models. _(Training)_ - **SGD:** Stochastic Gradient Descent — the simplest optimizer that updates parameters using the gradient of a randomly sampled mini-batch. The foundation all other optimizers build upon. _(Training)_ - **Supervised Learning:** Training a model on labeled data where each input has a known target output. The model learns to map inputs to outputs by minimizing prediction error. _(Training)_ - **Unsupervised Learning:** Learning patterns from data without explicit labels — includes clustering, dimensionality reduction, and density estimation. Discovers hidden structure in the data. _(Training)_ - **Weight Decay:** A regularization technique that adds a penalty proportional to the magnitude of weights to the loss, discouraging large weights and reducing overfitting. Equivalent to L2 regularization in SGD. _(Training)_ - **A/B Testing:** Deploying two model versions simultaneously to different user segments and comparing metrics to determine which performs better in production. _(Deployment)_ - **CI/CD:** Continuous Integration / Continuous Deployment — automated pipelines that test, build, and deploy code changes. In MLOps, extends to model training, validation, and serving. _(Deployment)_ - **Continuous Batching:** An inference optimization that dynamically adds new requests to a running batch as slots free up, dramatically improving GPU utilization compared to static batching. _(Deployment)_ - **Data Drift:** A shift in the distribution of incoming production data compared to the training data, potentially degrading model performance over time. Requires monitoring and retraining. _(Deployment)_ - **Feature Store:** A centralized repository that stores, manages, and serves pre-computed features for both training and inference, ensuring consistency between offline and online environments. _(Deployment)_ - **Latency:** The time between sending a request and receiving a response. In LLM serving, measured as time-to-first-token (TTFT) and inter-token latency. _(Deployment)_ - **llama.cpp:** A C/C++ library for running LLM inference on CPUs and GPUs with aggressive quantization support. Enables running large models on consumer hardware. _(Deployment)_ - **MLOps:** Machine Learning Operations — practices for deploying and maintaining ML models in production, bridging the gap between model development and reliable software engineering. _(Deployment)_ - **Model Compression:** Umbrella term for techniques (quantization, pruning, distillation) that reduce model size and inference cost while preserving as much accuracy as possible. _(Deployment)_ - **Model Drift:** Gradual degradation of a deployed model's performance as the real world changes. Caused by data drift, concept drift, or upstream data pipeline changes. _(Deployment)_ - **Model Registry:** A centralized store for versioning, tracking, and managing trained models through their lifecycle stages (development → staging → production). _(Deployment)_ - **PagedAttention:** A memory management technique used by vLLM that stores KV cache in non-contiguous pages (like OS virtual memory), enabling efficient memory sharing and higher throughput. _(Deployment)_ - **Prefix Caching:** Reusing previously computed KV cache entries for shared prompt prefixes across requests, avoiding redundant computation for system prompts or common prefixes. _(Deployment)_ - **Throughput:** The number of requests or tokens a serving system can process per second. High throughput is critical for cost-effective deployment at scale. _(Deployment)_ - **Triton Inference Server:** NVIDIA's open-source inference serving platform supporting multiple frameworks and model types with dynamic batching, model ensembling, and GPU scheduling. _(Deployment)_ - **Agent / Agentic AI:** An AI system that can autonomously plan, use tools, execute multi-step tasks, and interact with external systems to achieve goals — beyond simple prompt-response interaction. _(Research)_ - **Alignment:** The challenge of making AI systems behave in accordance with human values and intentions. Includes techniques like RLHF, constitutional AI, and red teaming. _(Research)_ - **AUC-ROC:** Area Under the Receiver Operating Characteristic Curve — a metric measuring a classifier's ability to distinguish between classes across all probability thresholds. _(Research)_ - **BPE:** Byte Pair Encoding — a tokenization algorithm that iteratively merges the most frequent character pairs into subword tokens, balancing vocabulary size with representation efficiency. _(Research)_ - **Chain-of-Thought (CoT):** A prompting technique where the model is encouraged to show its reasoning step-by-step before giving a final answer, significantly improving performance on complex reasoning tasks. _(Research)_ - **Context Window:** The maximum number of tokens a model can process in a single forward pass. Determines how much text the model can 'see' at once — ranges from 2K to 1M+ tokens in modern LLMs. _(Research)_ - **Data Contamination:** When benchmark test data accidentally appears in a model's training set, inflating evaluation scores and giving a misleading picture of true generalization ability. _(Research)_ - **Distribution Shift:** When the data a model encounters in production differs from its training distribution, causing degraded performance. A fundamental challenge in real-world ML deployment. _(Research)_ - **FLOPS:** Floating Point Operations Per Second — a measure of computational throughput. Used to estimate training cost and compare hardware (e.g., H100 = ~1 petaFLOP for FP16). _(Research)_ - **Generalization:** A model's ability to perform well on unseen data that wasn't part of training. The ultimate goal of machine learning — memorizing training data is not learning. _(Research)_ - **GPTQ:** A post-training quantization method for LLMs that uses approximate second-order information to minimize quantization error, enabling 3-4 bit models with minimal quality loss. _(Research)_ - **GRPO:** Group Relative Policy Optimization — DeepSeek's RL algorithm that replaces the critic model with group-based advantage estimation, simplifying RLHF training while maintaining performance. _(Research)_ - **In-Context Learning:** An LLM's ability to learn new tasks from examples provided directly in the prompt, without any weight updates. A key emergent capability of large models. _(Research)_ - **Interpretability:** The ability to understand and explain why a model makes specific predictions. Ranges from attention visualization to mechanistic interpretability of individual circuits. _(Research)_ - **Loss Landscape:** The surface formed by the loss function over all possible parameter values. Its shape (smooth vs rugged, flat vs sharp minima) determines how easily optimization can find good solutions. _(Research)_ - **Multimodal:** Models that process and generate multiple data types (text, images, audio, video) in a unified architecture. Examples: GPT-4V, Gemini, Claude's vision capabilities. _(Research)_ - **Out-of-Distribution (OOD):** Data that falls outside the distribution the model was trained on. OOD detection identifies when a model is being asked to extrapolate rather than interpolate. _(Research)_ - **Prompt Engineering:** The practice of crafting input prompts to elicit desired behavior from language models — including system prompts, few-shot examples, chain-of-thought, and structured output formats. _(Research)_ - **RAG:** Retrieval-Augmented Generation — a technique that retrieves relevant documents from an external knowledge base and includes them in the prompt, grounding LLM responses in factual sources. _(Research)_ - **Red Teaming:** Systematically probing an AI model to discover failure modes, biases, harmful outputs, and security vulnerabilities before deployment. _(Research)_ - **Reproducibility:** The ability to replicate an experiment's results given the same code, data, and configuration. A cornerstone of scientific research and trustworthy ML. _(Research)_ - **ROUGE Score:** Recall-Oriented Understudy for Gisting Evaluation — a family of metrics comparing generated text to references using n-gram overlap. Standard for evaluating summarization. _(Research)_ - **SOTA:** State of the Art — the best performance achieved on a benchmark or task at a given time. 'Achieving SOTA' means beating all previously published results. _(Research)_ - **Synthetic Data:** Artificially generated training data created by models, simulations, or rules. Increasingly used to scale training data for LLMs and address data scarcity in specialized domains. _(Research)_ - **Tool Use / Function Calling:** An LLM's ability to invoke external tools (calculators, APIs, databases, code execution) by generating structured function calls, extending its capabilities beyond text generation. _(Research)_ --- This representation contains public Fanout content only. Protected Pro lessons, account data, billing, checkout, and pricing are not included. Browse the public content map: https://fanout.sh/sitemap.md