AI concepts

28 explained· 300 coming soon · See alsoEngineering Product Research

28Explained · every one with checked sources

Retrieval-Augmented Generation
A method that answers a question by searching a document collection first, then having a language model write the answer from the passages it found.
Mixture of Experts
A model design that stores many sub-networks, called experts, and runs only a few of them per input, so capacity grows faster than compute cost.
Prompt Injection Attacks
Prompt injection is an attack that hides instructions inside text an AI system reads, causing it to follow the attacker's commands instead of its own.
Reinforcement Learning from Human Feedback
Reinforcement Learning from Human Feedback (RLHF) trains a language model to produce the responses human raters prefer, using a reward model built from their comparisons.
Pruning
Pruning is removing weights, neurons, or layers from a trained neural network to shrink it while keeping accuracy close to the original.
Quantization
Quantization stores a neural network's weights and activations using fewer bits, shrinking memory use and speeding up inference at some accuracy cost.
Distillation
A technique where a small model is trained to copy the behavior of a large one, keeping most of the accuracy while being cheaper and faster to run.
Bias-Variance Tradeoff
The pattern that a model's prediction error splits into bias from being too simple and variance from being too sensitive to its training sample, which move in opposite directions.
Data Leakage
In machine learning, information reaching a model during training that it would not have at prediction time, producing test scores that do not survive deployment.
CNN (Convolutional Neural Network)
A neural network that learns image features by sliding small reusable filters across the input, making it efficient at recognizing patterns like edges, shapes, and objects.
Activation Function
An activation function is the rule a single neuron in a neural network applies to its combined input before passing a signal onward, turning a plain weighted sum into a nonlinear value so the network can learn curves rather than only straight lines.
Active Learning
A machine learning approach where the model selects which unlabeled examples to have labeled next, usually the ones it is least certain about, to reach higher accuracy with fewer labels.
Fine-Tuning
Continuing to train a model that already learned general patterns during pretraining on a smaller, task-specific dataset, so it adapts to a narrower job without starting from scratch.
Self-Supervised Learning
Training a model on unlabeled data by hiding part of each example and having the model predict the hidden part from the rest.
Hallucination
When an AI model generates fluent, confident text that is factually wrong or unsupported by its source material, such as a fabricated citation or invented statistic.
Transformer
A neural network architecture that processes a whole sequence at once, letting every token draw information directly from every other token through attention.
Overfitting
When a machine learning model matches its training data so closely, including its noise, that it fails to make accurate predictions on new, unseen data.
Large Language Model (LLM)
A neural network trained on very large amounts of text to predict what comes next, then adjusted with human feedback so it follows instructions.
Chain of Thought
A prompting technique that asks a large language model to write out its intermediate reasoning steps before giving a final answer, instead of jumping straight to the answer.
Backpropagation
The algorithm that trains a neural network by computing how much each weight contributed to a wrong output, then adjusting every weight to reduce that error.
Beam Search
Beam search is a decoding method that keeps the k highest-scoring partial outputs at each step, so a model can compare several candidate sequences without scoring every possible one.
Self-Attention
The mechanism inside a Transformer that lets every element of a sequence weigh and combine information from every other element in that sequence.
Zero-Shot Learning
A model correctly handling a class or task it saw zero labeled training examples of, by transferring from a description or related task.
Few-Shot Learning
Few-shot learning is giving a model a handful of example input-output pairs inside the prompt itself, so it infers the task and applies it to a new input without any retraining.
Online Learning
In machine learning, updating a model's parameters incrementally on each new example as it streams in, rather than training once on a fixed batch, so the model keeps adapting to new data without being retrained from scratch.
Context Window
The total amount of text a language model can consider in a single request, measured in tokens and shared between the input and the response.
Loss Function
A loss function is the single number a training algorithm tries to minimize, measuring exactly how wrong a model's prediction is for one example.
Uncanny Valley
The dip in comfort people feel toward a robot, animated character, or synthetic face as its human likeness approaches, but does not quite reach, fully realistic.

28Explained · every one with checked sources

Retrieval-Augmented Generation

Mixture of Experts

Prompt Injection Attacks

Reinforcement Learning from Human Feedback

Pruning

Quantization

Distillation

Bias-Variance Tradeoff

Data Leakage

CNN (Convolutional Neural Network)

Activation Function

Active Learning

Fine-Tuning

Self-Supervised Learning

Hallucination

Transformer

Overfitting

Large Language Model (LLM)

Chain of Thought

Backpropagation

Beam Search

Self-Attention

Zero-Shot Learning

Few-Shot Learning

Online Learning

Context Window

Loss Function

Uncanny Valley

300Coming soon · not written yet

Temporal Consistency

Meta-Learning Framework

Sparse Attention Mechanism

Neural Turing Machine

Explainable Graph Attention

Causal Discovery

Sparsity-Aware Pruning

Quantization-Aware Training

Hierarchical Temporal Convolutional Network

Explainable Graph Attention Network

Data Shapley

Temporal Knowledge Graph Embedding Attacks

Robustness Evaluation Metrics

Reinforcement Learning from Demonstrations

Neural-Symbolic Integration Risks

Large Language Model Cascading Failures

Human-AI Collaborative Filtering

Graph-Based Explainability

Generative Sparring

Explainable AI Toolkit

Diffusion-Specific Loss Functions

Adversarial Robustness Evaluation

Value Alignment Frameworks

Temporal Graph Contrastive Learning

Neural-Symbolic Integration Framework

Knowledge Graph Completion

Graph Neural Network Pruning

Explainable Multimodal Fusion

Causal Representation Learning

Neural Tangent Kernel

Diffusion Probabilities

Temporal Knowledge Graph Embeddings

Sparse Graph Learning

Robust Out-of-Distribution Generalization

Prompt Engineering Frameworks

Neural Circuit Analysis

Memory-Efficient Transformers

Knowledge Graph Refinement

Graphcore

Explainable Multimodal Learning

Data Quality Metrics

Token-Free Models

Sparsity-Aware Training

Neural Rendering

Multimodal Transfer

Knowledge Graph Attention Network

Implicit Neural Representations

Human-in-the-Loop Feedback

Graph Embedding

Frozen Pre-trained Transformers