AI Lexicon

Abstraction is the process of extracting the essential features of a complex system or object and representing them in a simplified form.

Source: ollama:llama3.1:8b

Adversarial refers to a situation where a model or algorithm is intentionally fed false data to test the system's accuracy.

Source: ollama:llama3.1:8b

An adversarial example is an input that causes a machine learning model to misbehave, produce incorrect output, or fail.

Source: ollama:llama3.1:8b

Adversarial robustness refers to the ability of an algorithm to withstand unwanted or misleading input without significant performance degradation.

Source: ollama:llama3.1:8b

A training method where models are trained to resist adversarial attacks.

Source: ollama:llama3.1:8b

A form of AI designed to act independently with a degree of agency, plan, and invoke tools on its own.

A theoretical level of AI that equals or exceeds human intelligence across all cognitive tasks.

An autonomous software program that perceives its environment, makes decisions, and takes actions to achieve specific goals using AI models.

Bias refers to a distortion or inaccuracy in the data or algorithm that leads to an incorrect conclusion.

Source: ollama:llama3.1:8b

A cascading failure is a situation in which a series of events amplifies each other, leading to an exponential increase in damage or problems.

Source: ollama:llama3.1:8b

Catastrophic forgetting is a problem that arises in some learning algorithms, where they quickly forget learned information and fail to develop long-term memory.

Source: ollama:llama3.1:8b

A causal graph is a given that represents a set of relationships between objects.

Source: ollama:llama3.1:8b

Causal reasoning is a method for determining the cause of an event by means of logical conclusions and inductive estimation.

Source: ollama:llama3.1:8b

The relationship between a cause and its effects.

Source: ollama:llama3.1:8b

Cognitive architectures are theoretical frameworks for structuring and modeling memory systems, including the human brain.

Source: ollama:llama3.1:8b

Cognitive bias refers to the way our perception, thinking, and behavior are influenced by preconceptions, misconceptions, or selective use of information.

Source: ollama:llama3.1:8b

Cognitive Load refers to the amount of mental effort required by a task or situation.

Source: ollama:llama3.1:8b

Cognitive mismatch refers to the difference between the intended and actual execution of a cognitive task.

Source: ollama:llama3.1:8b

The Cold Start problem occurs when a machine learning model (MLM) lacks sufficient data to make an accurate prediction or classification.

Source: ollama:llama3.1:8b

Concept drift is a phenomenon where the underlying distribution of data in a machine learning system changes over time, leading to decreased accuracy.

Source: ollama:llama3.1:8b

Context caching is a technique used to store data in memory and retrieve it quickly when needed, improving system performance.

Source: ollama:llama3.1:8b

The maximum amount of text (in tokens) an AI model can hold in its 'short-term memory' during a conversation.

Credibility Weight is the amount of credibility that a model or algorithm has to interpret data and make predictions.

Source: ollama:llama3.1:8b

A critic is a person who expresses an opinion on the quality or value of something, such as a work of art or a product.

Source: ollama:llama3.1:8b

A curriculum is a predetermined and recorded program of lessons, courses or activities that aims to achieve the desired learning outcomes.

Source: ollama:llama3.1:8b

The Curse of Dimensionality is a phenomenon that occurs when datasets with an increasing number of variables are processed.

Source: ollama:llama3.1:8b

An advanced form of ML inspired by the human brain, utilizing deep multi-layered neural networks.

Deterrence is the act of preventing a danger or threat, especially by offering consequences for those who take the risk.

Source: ollama:llama3.1:8b

A dissimilarity metric is a measure of the difference between two datasets.

Source: ollama:llama3.1:8b

Machine learning is a method for computers to discover and predict patterns in large datasets.

Source: kble:kble

An edge case is a situation that does not conform to the standard rules or general patterns, but can occur under specific circumstances.

Source: ollama:llama3.1:8b

The mathematical translation of words or documents into vectors, allowing an AI to measure semantic meaning.

Emergent behavior refers to the unpredictable and complex behavior that arises from the interactions of individual components in a system.

Source: ollama:llama3.1:8b

Evaluation is the process of assessing data or a situation to form an opinion about their value, quality, or suitability.

Source: ollama:llama3.1:8b

An evaluation metric is a measure used to assess the performance of a system or model.

Source: ollama:llama3.1:8b

Explainability is the property of a system to make its operations and decisions transparent and understandable.

Source: ollama:llama3.1:8b

Exploratory Play refers to the concept where a system is freely encouraged to explore and discover within an environment.

Source: ollama:llama3.1:8b

A situation in which a model or system is exposed to unusual or unexpected data, making it difficult for the model to function.

Source: ollama:llama3.1:8b

The process where the output of a system is circled back as input to improve future actions and refine performance.

Guiding an AI by first providing 2 or 3 concrete examples ('shots') of the exact format and tone expected.

Training an existing general AI model further with specific examples so it excels in a specific task.

An umbrella term for AI systems capable of creating new, original content rather than just analyzing data.

A gradient is a measurable change in a parameter, such as the value of a model.

Source: ollama:llama3.1:8b

Gradient Descent is an optimization algorithm used to adjust the parameters of a model.

Source: ollama:llama3.1:8b

The Hessian matrix of gradients is a two-dimensional square matrix that computes the second derivative of a cost function or log-likelihood function.

Source: ollama:llama3.1:8b

When an AI model confidently presents factually incorrect, irrelevant, or entirely fabricated information.

A hybrid approach is a combination of different methods or techniques used to solve a complex problem.

Source: ollama:llama3.1:8b

A hyperparameter is a setting or parameter that is set by the user before training the model.

Source: ollama:llama3.1:8b

The phase in production where an already trained AI model is actively used to generate predictions or answers based on new input.

Interpretability refers to the extent to which a complex system or model can be understood and explained.

Source: ollama:llama3.1:8b

Invariance is a property of a system to remain unaffected by changes in the input or environment.

Source: ollama:llama3.1:8b

A latent space is an abstract vector space where data points are represented as vectors.

Source: ollama:llama3.1:8b

A latent variable is an unobservable variable in a statistical model that affects the values of other variables.

Source: ollama:llama3.1:8b

A massive neural network trained on vast amounts of text to understand and generate human language.

A branch of AI where computers learn from data and discover patterns without being explicitly programmed.

An AI model capable of understanding and generating multiple forms of media simultaneously (text, image, audio, video).

A mathematical computing model roughly inspired by the functioning of connected neurons in the human brain.

The field of AI concerned with how computers can understand, interpret, and produce human language.

Optimism bias refers to the tendency to expect a favorable outcome, even when there is no basis for assuming that it will occur.

Source: ollama:llama3.1:8b

The overarching system component that coordinates sub-agents, delegates tasks, and monitors overall progress.

Overfitting occurs when a machine learning model is too complex and captures the noise in the training data rather than the underlying patterns.

Source: ollama:llama3.1:8b

The billions of numerical values ('weights') in a neural network tuned during training, representing the model's 'knowledge'.

A structured step-by-step plan or workflow followed by an agent to execute complex, repeatable processes.

The skill of formulating instructions or questions to get the best possible result from an AI model.

A technique where an AI first searches a database for relevant facts before generating an answer.

The process where experts attempt to hack or deceive an AI system. The goal is to find weaknesses and security flaws before the system goes live.

Reward shaping is a technique used to steer desired behavior in artificial intelligence systems by adjusting the reward system.

Source: ollama:llama3.1:8b

Robustness refers to the property of a system to be tolerant of unexpected or disruptive factors, without suffering collapse.

Source: ollama:llama3.1:8b

A saliency map is an image that depicts the locations of high value, based on their informative significance for a given task or goal.

Source: ollama:llama3.1:8b

Sensitivity analysis is a technique used to measure the sensitivity of models, allowing them to predict how they will react to changes in inputs.

Source: ollama:llama3.1:8b

Shadowing is a machine learning technique where one model perfectly replicates the speech or audio flow of another model.

Source: ollama:llama3.1:8b

A singular value is a non-negative number that determines the amount of information retained in a matrix.

Source: ollama:llama3.1:8b

A specific capability or tool assigned to an AI agent, complete with instructions and code implementations.

Societal Value Alignment refers to the agreement between an organization or system and the social norms and values that govern society.

Source: ollama:llama3.1:8b

A surrogate is a model that simulates data so that other algorithms can train on these simulations.

Source: ollama:llama3.1:8b

The invisible background instruction that defines the fundamental identity and rules of an AI chatbot.

The functional building blocks (often parts of a word) into which an AI divides language to process it mathematically.

Transfer Learning is a technique where a pre-trained model is used to tackle a new problem.

Source: ollama:llama3.1:8b

The underlying AI architecture (invented in 2017) that enables models to understand context in long texts using an attention mechanism.

A utility function is a mathematical construct that describes the objective of a decision-making process.

Source: ollama:llama3.1:8b

Utility maximization is the process of maximizing the value of a decision or action by considering all relevant factors.

Source: ollama:llama3.1:8b

Value Alignment is the matching of human values with the goals of an AI system to ensure that it functions ethically.

Source: ollama:llama3.1:8b

Value drift refers to the situation where an AI model gradually deviates from its original performance or objectives, typically due to changes in the training data or the model itself.

Source: ollama:llama3.1:8b

A Value Neuron is a type of neural network designed to identify and prioritize the useful difference between various variables.

Source: ollama:llama3.1:8b

A special type of database that stores data as vectors, allowing AI to search by meaning and context rather than exact keywords.

Having an AI perform a task without providing any previous examples in the prompt.