This glossary covers essential terms for understanding natural language and artificial intelligence technologies. Although these terms are common in articles and resources written about AI, some may be used in other fields or contexts. all definitions in this glossary are written within the context of AI.
#
3D Autoencoders: A specialized form of two-part neural network that includes an “encoder” and a “decoder.” The encoder transforms the initial data into a smaller depiction. The decoder tries to reconstitute the original data from the depiction, restoring it to its original state.
3D GAN (Generative Adversarial Network): A distinctive architectural framework within the GAN paradigm, specialized for the generation of three-dimensional shapes.
A
A3C (Asynchronous Advantage Actor-Critic): A robust reinforcement learning algorithm in which a policy and value function coexist. Operating within the forward trajectory, A3C leverages multi-step returns for policy and value-function updates, exemplifying sophisticated learning techniques.
A2C (Advantage Actor-Critic): An advanced fusion of policy gradient and learned value function within reinforcement learning. This hybrid algorithm is characterized by two interdependent components: the “Actor,” which learns a parameterized policy, and the “Critic,” which assimilates a value function for the evaluation of state-action pairs. These components collectively contribute to a refined learning process.
Accuracy: Accuracy is a scoring system in binary classification (i.e., determining if an answer or output is correct or not) and is calculated as (True Positives + True Negatives) / (True Positives + True Negatives + False Positives + False Negatives).
Actionable Intelligence: Information you can leverage to support decision-making.
Activation Function: A crucial element within artificial neural networks, responsible for modifying input signals. This function adjusts the output magnitude based on the input magnitude: Inputs over a predetermined threshold result in larger outputs. It acts like a gate that selectively permits values over a certain point.
Active Learning: A form of reinforcement learning from human feedback where an algorithm actively engages with a user to obtain labels for data. It refines its performance by getting labels for desired outputs.
Actor-Critic Model: A two-part algorithmic structure employed in reinforcement learning. Within this model, the “Actor” determines optimal actions based on the state of its environment. At the same time, the “Critic” evaluates the quality of state-action pairs, improving them over time.
Adapter: Adapters are an advanced method for making pre-trained AI models adaptable to new tasks without complete retraining. These modules save time, money, and resources by efficiently repurposing existing models for different tasks in areas like natural language processing, computer vision, and robotics.
Adversarial Attack: An attempt to damage a machine learning model by giving it misleading or deceptive data during its training phase, or later exposing it to maliciously-engineered data, with the intent to induce, degrade, or manipulate the model’s output.
Adversarial Examples: The building blocks of an adversarial attack: Inputs deliberately constructed to provoke errors in machine learning models. These are typically deviations from valid inputs included in the data set that involve subtle alterations that an attacker introduces to exploit vulnerabilities in the model.
Agent: Any software program or autonomous entity (such as a machine learning model) capable of acting or making decisions in pursuit of specific goals.
Agent-Based Modeling: A method employed for simulating intricate systems, focusing on interactions between individual agents to glean insights into emergent system behaviors.
AGI (Artificial General Intelligence): A concept that suggests a more advanced version of AI than we know today, one that can perform tasks much better than humans while also teaching and advancing its own capabilities.
AI (Artificial Intelligence): The simulation of human intelligence in machines that are programmed to think and learn like humans.
Example: A self-driving car that can navigate and make decisions on its own using AI technology.
AI Copilot: An AI copilot is a conversational interface that uses large language models to support users in various tasks and decision-making processes across multiple domains within an enterprise environment.
AI Ethics: Principles aimed at preventing AI from harming humans, achieved through means like determining how AI systems should collect data or deal with bias.
AI Music: A musical composition made by or with AI-based audio generation.
AI Plugin: AI plugins are specialized software components that allow AI systems to interface with external applications and services.
AI Policy and Regulation: The formulation of public sector frameworks and legal measures aimed at steering and overseeing artificial intelligence technologies. This facet of regulation extends into the broader realm of algorithmic governance.
AI Safety: An interdisciplinary field that’s concerned with the long-term impacts of AI and how it could progress suddenly to a superintelligence that could be hostile to humans.
AI Writer: A software application that uses artificial intelligence to produce written content, mimicking human-like text generation. AI writing tools can be invaluable for businesses engaged in content marketing.
AI Writing: Text written by, or with the assistance of, an AI writer.
ALBERT (A Lite BERT): A cloud-centric artificial intelligence platform that helps integrate and manage an existing digital marketing tech stack.
Algorithm: A series of instructions that allows a computer program to learn and analyze data in a particular way, such as recognizing patterns, to then learn from it and accomplish tasks on its own.
AlphaGo Alpha Zero: A specialized computer program devised by Google DeepMind to play the intricate Chinese strategy game Go. It showcases the potential of narrow AI by engaging in strategic gameplay akin to chess but with a far broader range of possible outcomes.
Alignment: Tweaking an AI to better produce the desired outcome. This can refer to anything from moderating content to maintaining positive interactions toward humans.
ANN (Artificial Neural Network): A key element of machine learning that serves as the cornerstone of deep learning. It uses an intricate network structure that mirrors the neural connections of the human brain.
Anaphora: In linguistics, an anaphora is a reference to a noun by way of a pronoun.
Example: In the sentence, “While John didn’t like the appetizers, he enjoyed the entrée,” the word “he” is an anaphora.
Annotation: The process of labeling data with additional information to help machine learning algorithms understand and learn.
Anthropomorphism: When humans tend to give nonhuman objects humanlike characteristics. In AI, this can include believing a chatbot is more humanlike and aware than it actually is, like believing it’s happy, sad, or even sentient altogether.
Associative Memory: Associative memory refers to a system’s ability to store, retrieve, and process related information based on connections between elements, enabling it to efficiently identify and use relevant data for decision-making.
ASR (Automatic Speech Recognition): A technology that transcribes spoken language into text.
Attention Map: A visual representation accentuating sections of an image pertinent to a specific target class. This offers interpretable insights into the inner workings of deep neural networks.
Attention Mechanisms: An aspect of machine learning models enabling them to prioritize certain data segments during predictions. This mimics human cognitive focus by assigning different weights to distinct data elements, allowing the model to pay more attention to those elements.
AUC-ROC (Area Under the Receiver Operating Characteristic Curve): A quantitative evaluation metric for classification tasks, charting the performance curve of various threshold settings. The Area Under the Curve (AUC) signifies the degree of separability within the classification model.
Audio Generation: The process of generating raw audio content such as speech or AI music by using artificial intelligence.
Auto-classification: The application of machine learning, natural language processing (NLP), and other AI-guided techniques to automatically classify text in a faster, more cost-effective, and more accurate manner.
Auto-complete: Auto-complete is a search functionality used to suggest possible queries based on the text being used to compile a search query.
Autoencoder: A specialized variant of artificial neural networks employed for unsupervised learning. Autoencoders master the dual functions of data encoding and decoding, facilitating efficient data representation and reconstruction.
Automation: Automation refers to the use of technology to perform tasks with minimal human intervention.
B
Backdoor Attack: A strategy employed to maliciously access computer systems or encrypted data by using covert pathways to evade conventional security measures.
Backpropagation: Short for “backward propagation of errors,” this is an algorithmic bedrock of supervised learning for artificial neural networks. Backpropagation computes weight adjustments based on the gradient of an error function to help refine a model.
Batch Size: The number of samples processed within a single iteration of model training. Batch size governs the pace of model updates and optimization.
Bayesian Optimization: A methodology using probabilistic modeling to guide hyperparameter selection, enhancing the efficiency of optimization procedures in machine learning.
Bellman Equation: A pivotal dynamic programming equation integral to discrete-time optimization problems, facilitating optimal decision-making within sequential contexts.
Benchmarking: Benchmarking is the process of evaluating and comparing products or systems using standardized tests to gauge performance and capabilities.
BERT (Bidirectional Encoder Representations from Transformers): Google’s technology. A large-scale pretrained model that is first trained on very large amounts of unannotated data. The model is then transferred to an NLP task where it is fed another smaller task-specific dataset which is used to fine-tune the final model.
BERT (Bidirectional Encoder Representations from Transformers): An influential deep learning approach employed in natural language processing, enabling artificial intelligence programs to unravel contextual nuances in ambiguous text.
Bias: The presence of systematic and undesired preferences or imbalances in the output generated by an AI model. Bias can emerge in various forms, such as in the content, language, or perspectives generated by the AI system.
Bias: In regards to large language models, errors resulting from the training data. This can result in falsely attributing certain characteristics to certain races or groups based on stereotypes.
Big data: Big data refers to the large data sets that can be studied to reveal patterns and trends to support business decisions. It’s called “big” data because organizations can now gather massive amounts of complex data using data collection tools and systems. Big data can be collected very quickly and stored in a variety of formats.
Black Box: A system or model whose internal mechanisms or workings are not transparently understandable from its inputs and outputs. The inner processes are concealed, making it challenging to discern how the system arrives at specific decisions or predictions. While black box models such as deep neural networks can achieve high performance, their lack of interpretability can hinder understanding and trust.
BLEU Score (Bilingual Evaluation Understudy Score): A metric for automatic evaluation of machine-translated text. The BLEU score gauges the similarity between machine-generated text and a set of high-quality reference translations, yielding a value between zero and one.
Burstiness: The abrupt shifts in quality, coherence, or relevance often observed in AI-generated content, particularly in writing. It refers to the inconsistencies in style, tone, or factual accuracy that can occur within a short span. Identifying burstiness helps distinguish AI-generated content from human-created content.
C
Capsule Network: A form of artificial neural networks that model intricate hierarchical relationships. Drawing inspiration from biological neural organization, they aim to emulate more closely the structure of human neural connections.
Catastrophic Forgetting: A problem that occurs when two similar game states yield dramatically divergent outcomes, causing confusion in the Q-function’s learning process.
Cataphora: In linguistics, a cataphora is a reference placed before any instance of the noun it refers to.
Example: In the sentence, “Though he enjoyed the entrée, John didn’t like the appetizers,” the word “he” is a cataphora.
Category: A category is a label assigned to a document in order to describe the content within said document.
Category Trees: Enables you to view all of the rule-based categories in a collection. Used to create categories, delete categories, and edit the rules that associate documents with categories. Is also called a taxonomy, and is arranged in a hierarchy.
Chatbot¹: A user-friendly interface that allows the user to ask questions and receive answers. Depending on the backend system that fuels the chatbot, it can be as basic as pre-written responses to a fully conversational AI that automates issue resolution.
Chatbot²: An interactive software application that imitates human conversation via text or voice interactions, frequently found in online environments.
ChatGPT: A chat interface built on top of GPT-3.5. GPT-3.5 is a large language model developed by OpenAI that is trained on a massive amount of internet text data and fine-tuned to perform a wide range of natural language tasks. Example: GPT-3.5 has been fine-tuned for tasks such as language translation, text summarization, and question answering.
Class: A distinct category or group that objects, entities, or data points can be classified into based on shared characteristics or features. They are used in various machine learning and pattern recognition tasks such as image classification or text categorization in which algorithms learn to differentiate and sort input data.
Classification: Techniques that assign a set of predefined categories to open-ended text to be used to organize, structure, and categorize any kind of text – from documents, medical records, emails, files, within any application and across the web or social media networks.
CNN (Convolutional Neural Network): A deep learning class of neural networks with one or more layers used for image recognition and processing.
Co-occurrence: A co-occurrence commonly refers to the presence of different elements in the same document. It is often used in business intelligence to heuristically recognize patterns and guess associations between concepts that are not naturally connected (e.g., the name of an investor often mentioned in articles about startups successfully closing funding rounds could be interpreted as the investor is particularly good at picking his or her investments).
Cognitive computing: Cognitive computing is essentially the same as AI. It’s a computerized model that focuses on mimicking human thought processes such as pattern recognition and learning. Marketing teams sometimes use this term to eliminate the sci-fi mystique of AI.
Cognitive Map: A mental representation (otherwise known as a mental palace) which serves an individual to acquire, code, store, recall, and decode information about the relative locations and attributes of phenomena in their environment.
Collective Learning: Collective learning is an AI training approach that leverages diverse skills and knowledge across multiple models to achieve more powerful and robust intelligence.
Completions: The output from a generative prompt.
Composite AI: The combined application of different AI techniques to improve the efficiency of learning in order to broaden the level of knowledge representations and, ultimately, to solve a wider range of business problems in a more efficient manner.
Computational Creativity: A multidisciplinary domain that aims to replicate human creativity through computational methods by combining approaches from fields including artificial intelligence, philosophy, cognitive psychology, and the arts.
Computational Linguistics (Text Analytics, Text Mining): Computational linguistics is an interdisciplinary field concerned with the computational modeling of natural language.
Computational Semantics (Semantic Technology): Computational semantics is the study of how to automate the construction and reasoning of meaning representations of natural language expressions.
Computer Vision: Computer vision is an interdisciplinary field of science and technology that focuses on enabling computers to interpret, understand, and process visual information from the world. It has a broad range of applications including object detection, image classification, facial recognition, and autonomous vehicles.
Computer Vision: A subfield of artificial intelligence (AI) and computer science that focuses on enabling computers to interpret, understand, and process visual information from the world. It has a broad range of applications including object detection, image classification, facial recognition, and autonomous vehicles.
Confusion Matrix: A pivotal tool for evaluating model performance by identifying misclassified objects, providing insights into model accuracy.
Content: Individual containers of information — that is, documents — that can be combined to form training data or generated by Generative AI.
Content Enrichment or Enrichment: The process of applying advanced techniques such as machine learning, artificial intelligence, and language processing to automatically extract meaningful information from your text-based documents.
Contextual Bandit: An extension of the multi-armed bandit approach that considers contextual information to optimize decision-making in multi-action scenarios.
Continual Learning: An approach that enables machine learning models to accumulate knowledge from a sequence of unrelated tasks, preserving previously acquired insights and applying them to new challenges. Also known as incremental learning or lifelong learning.
Controlled Vocabulary: A controlled vocabulary is a curated collection of words and phrases that are relevant to an application or a specific industry. These elements can come with additional properties that indicate both how they behave in common language and what meaning they carry, in terms of topic and more.
While the value of a controlled vocabulary is similar to that of taxonomy, they differ in that the nodes in taxonomy are only labels representing a category, while the nodes in a controlled vocabulary represent the words and phrases that must be found in a text.
Controllability: Controllability is the ability to understand, regulate, and manage an AI system’s decision-making process, ensuring its accuracy, safety, and ethical behavior, and minimizing the potential for undesired consequences.
Convergence: Convergence in machine learning refers to the state during training where a model’s loss stabilizes within a certain error range around its final value. It signifies that further training will not significantly enhance the model’s performance.
Conversational AI: Used by developers to build conversational user interfaces, chatbots, and virtual assistants for a variety of use cases. They offer integration into chat interfaces such as messaging platforms, social media, SMS, and websites. A conversational AI platform has a developer API so third parties can extend the platform with their own customizations.
Conversational AI: A subfield of AI that focuses on developing systems that can understand and generate human-like language and conduct a back-and-forth conversation.
Example: A chatbot that can understand and respond to customer inquiries in a natural and human-like manner.
Corpus: The full set of data utilized to train an artificial intelligence model. More specifically, a corpus is a balanced collection of documents that should be representative of the documents an NLP solution will face in production, both in terms of content as well as distribution of topics and concepts.
Cost of Large Language Models: The cost of large language models primarily stems from their size and complexity, which demand significant computational power, storage, and resources for training and deployment. These factors can result in substantial expenses for building, maintaining, and using such models, sometimes amounting to several dollars per conversation or thousands of dollars per month.
Cross-Validation: A robust technique for assessing machine learning models by training them on only a subset of the data and evaluating them on a different subset.
Curriculum Learning: An approach in machine learning that mimics classical human education by introducing progressively more complex aspects of a problem. This enhances a model’s learning trajectory by ensuring it remains optimally challenged.
CycleGAN: An image-to-image translation methodology using unpaired datasets to learn mappings between the input and output images.
D
Data Annotation: The process of labeling and annotating data to facilitate supervised learning, enhancing a model’s understanding of the inputs.
Data Augmentation: A technique where users artificially enrich the training set by adding modified copies of the same data. It involves making minor changes such as flipping, resizing, or adjusting the brightness of images, to enhance the dataset and prevent models from overfitting.
Data Discovery: The process of uncovering data insights and getting those insights to the users who need them, when they need them.
Data Drift: Data Drift occurs when the distribution of the input data changes over time; this is also known as covariant shift.
Data Extraction: Data extraction is the process of collecting or retrieving disparate types of data from a variety of sources, many of which may be poorly organized or completely unstructured.
Data Imbalance: An uneven distribution of classes within a dataset. This can challenge model performance and accuracy.
Data Ingestion: The process of obtaining disparate data from multiple sources, restructuring it, and importing it into a common format or repository to make it easy to utilize.
Data Labelling: A technique through which data is marked to make objects recognizable by machines. Information is added to various data types (text, audio, image, and video) to create metadata used to train AI models.
Data Leakage: A phenomenon where external information inadvertently influences model training, compromising its integrity.
Data Poisoning: A type of adversarial attack involving the manipulation of training data by deliberately introducing contaminated samples to skew model behavior and outputs.
Data Scarcity: The lack of data that could possibly satisfy the need of the system to increase the accuracy of predictive analytics.
DDPG (Deep Deterministic Policy Gradient): A reinforcement learning algorithm employing deep neural networks to learn optimal policies in continuous action spaces to maximize the expected long-term reward.
Deep Learning (DL): A type of machine learning that relates to intricate neural network architectures capable of hierarchical feature extraction.
Example: A deep learning model that can recognize objects in an image by processing the image through multiple layers of neural networks.
Deep Learning: Deep learning is part of a broader family of machine learning methods based on artificial neural networks with representation learning. In other words, deep learning models can learn to classify concepts from images, text, or sound.
Deep Q-Learning: A pivotal approach in reinforcement learning that employs deep neural networks to approximate the Q-function, which it uses to determine the optimal course of action.
Deepfake: Synthetic media (typically images or videos) produced through digital manipulation by convincingly replacing one individual’s likeness with another’s.
Depth Estimation: The task of predicting the depth of objects within an image. This is essential for various computer vision applications like self-driving vehicles.
Deterministic Model: A deterministic model follows a specific set of rules and conditions to reach a definite outcome, operating on a cause-and-effect basis.
Differentiable Neural Computers: Advanced and typically recurrent neural network architectures enhanced with memory modules for complex learning and reasoning tasks.
Differentiable Rendering: A process enabling gradients of 3D objects to be calculated and propagated through 2D images.
Dimensionality Reduction: A technique for simplifying datasets by reducing their feature dimensions while preserving critical information.
Disambiguation: Disambiguation, or word-sense disambiguation, is the process of removing confusion around terms that express more than one meaning and can lead to different interpretations of the same string of text.
Discriminative Model: Discriminative models are algorithms designed to directly model and learn the boundary between different classes or categories in a dataset.
Discriminative Models: Models designed for supervised machine learning that focus on learning class boundaries and decision boundaries.
Domain Knowledge: The experience and expertise your organization has acquired over time.
Double DQN: A DQN technique that uses double Q-learning to mitigate overestimation biases and improve Q-value approximations.
DQN (Deep Q-Network): A framework employing deep neural networks for Q-learning in reinforcement learning tasks.
Dropout: A regularization technique involving the temporary exclusion of randomly selected nodes during neural network training.
DTW (Dynamic Time Warping): A method for measuring similarity between time series data, commonly used in time series analysis and pattern recognition.
Dueling DQN: A reinforcement learning algorithm incorporating both value and advantage functions for improved Q-value estimation.
DYM (Did You Mean): “Did You Mean” is an NLP function used in search applications to identify typos in a query or suggest similar queries that could produce results in the search database being used.
E
Edge Learning: A decentralized approach to machine learning where processing occurs on user devices, enhancing privacy and efficiency. This required model compression because of the complexity of AI programs.
Edge model: A model that includes data typically outside centralized cloud data centers and closer to local devices or individuals — for example, wearables and Internet of Things (IoT) sensors or actuators.
Embeddings: A set of data structures in a large language model (LLM) of a body of content where a high-dimensional vector represents words. This is done so data is more efficiently processed regarding meaning, translation, and generation of new content.
ELMo (Embeddings from Language Models): A word embedding technique that generates context-aware word representations by considering character-level tokens.
Emergent behavior: Emergent behavior, also called emergence, is when an AI system shows unpredictable or unintended capabilities.
Emergent behavior: When an AI model exhibits unintended abilities.
Emotion AI (aka Affective Computing): AI to analyze the emotional state of a user (via computer vision, audio/voice input, sensors, and/or software logic). It can initiate responses by performing specific, personalized actions to fit the mood of the customer.
End-to-end learning (E2E): A deep learning process in which a model is instructed to perform a task from start to finish. It’s not trained to accomplish a task sequentially but instead learns from the inputs and solves it all at once.
Enterprise AI: Enterprise AI refers to the strategic integration and deployment of AI within an organizational framework to enhance various business processes, decision-making, and overall operational efficiency.
Entity: An entity is any noun, word, or phrase in a document that refers to a concept, person, object, abstract or otherwise (e.g., car, Microsoft, New York City). Measurable elements are also included in this group (e.g., 200 pounds, 14 fl. oz.).
Environmental, Social, and Governance (ESG): An acronym initially used in business and government pertaining to enterprises’ societal impact and accountability; reporting in this area is governed by a set of binding and voluntary regulatory reporting.
Epoch: An iteration in the training process where the entire dataset is presented to a machine learning model.
ESG (Environmental, Social, and Governance): An acronym initially used in business and government pertaining to enterprises’ societal impact and accountability; reporting in this area is governed by a set of binding and voluntary regulatory reporting.
Ethical considerations: An awareness of the ethical implications of AI and issues related to privacy, data usage, fairness, misuse, and other safety issues.
ETL (Entity Recognition, Extraction): Entity extraction is an NLP function that serves to identify relevant entities in a document.
Explainable AI/Explainability: An AI approach where the performance of its algorithms can be trusted and easily understood by humans. Unlike black-box AI, the approach arrives at a decision and the logic can be seen behind its reasoning and results.
Explainability: Explainability refers to techniques that make AI model decisions and predictions interpretable and understandable to humans.
Extensibility: Extensibility in AI refers to the ability of AI systems to expand their capabilities to new domains, tasks, and datasets without needing full retraining or major architectural changes.
Extraction or Keyphrase Extraction: Multiple words that describe the main ideas and essence of text in documents.
Extraction: Extraction is the ability of generative models to analyze large datasets and extract relevant patterns, trends, and specific pieces of information.
Extractive Summarization: Identifies the important information in a text and groups the text fragments together to form a concise summary.
F
F-score (F-measure, F1 measure): An F-score is the harmonic mean of a system’s precision and recall values. It can be calculated by the following formula:
2 x [(Precision x Recall) / (Precision + Recall)].
Criticism around the use of F-score values to determine the quality of a predictive system is based on the fact that a moderately high F-score can be the result of an imbalance between precision and recall and, therefore, not tell the whole story. On the other hand, systems at a high level of accuracy struggle to improve precision or recall without negatively impacting the other.
Critical (risk) applications that value information retrieval more than accuracy (i.e., producing a large number of false positives but virtually guaranteeing that all the true positives are found) can adopt a different scoring system called F2 measure, where recall is weighed more heavily. The opposite (precision is weighed more heavily) is achieved by using the F0.5 measure.
FastText: A word embedding technique that represents words as bags of character N-grams, facilitating efficient language processing.
Feature Extraction: The process of distilling relevant information from raw data to create meaningful features for machine learning.
Feature Selection: The task of identifying and retaining the most crucial features while discarding less relevant ones in a dataset to use only relevant data and ignore noise.
Federated Learning: A collaborative machine learning paradigm where models are trained across distributed devices while preserving data privacy.
Few-shot learning: Few-shot learning is a machine learning approach where models can learn concepts from a small number of training examples to generalize and produce worthwhile output, often 5 or less per category.
Fine Tuning: The process of adapting a pre-trained model to a specific task, context, category, or problem set by further training it on a smaller, more focused dataset. The process involves adjusting the model’s parameters using task-specific data to improve its performance on the desired task. Fine-tuning allows the model to leverage the knowledge gained from its initial pre-training while specializing in a particular domain.
Example: An image classification model that has been pre-trained on a large dataset of various images can be fine-tuned to detect specific instances, such as when a car runs a red light at an intersection, by training it on a smaller dataset containing relevant images.
Fitting: The process of adjusting the parameters of a model to best match observed data. Fitting involves minimizing the difference between the model’s predictions and the actual data points, typically achieved through optimization techniques like gradient descent. This process enables the model to generalize and make accurate predictions on new, unseen data.
Foom: Also known as fast takeoff or hard takeoff. The concept that if someone builds an AGI that it might already be too late to save humanity.
Foundation Model: A baseline model used for a solution set, typically pretrained on large amounts of data using self-supervised learning. Applications or other models are used on top of foundational models — or in fine-tuned contextualized versions.
Examples: BERT, GPT-n, Llama, DALL-E, etc.
FQI (Fitted Q Iteration): An algorithm in reinforcement learning used to approximate the Q-function and solve optimal control problems.
G
GANs (Generative Adversarial Networks): A generative AI model composed of two neural networks to generate new data: a generator and a discriminator. The generator creates new content, and the discriminator checks to see if it’s authentic.
GenAI (Generative AI): A content-generating technology that uses AI to create text, video, computer code, images, or music. The AI is fed large amounts of training data, finds patterns to generate its own novel responses, which can sometimes be similar to the source material.
Example: Creating an original short story based on analyzing existing, published short stories.
Generalized model: A model that does not specifically focus on use cases or information.
Generation: Generation is the ability of a generative model to create brand new, original content such as text, images, audio or video from scratch.
Generative 3D Modeling: A technique for representing three-dimensional shapes as a series of processing steps using generative algorithms, simplifying design and fabrication processes.
Generative Design: A design approach driven by AI, where algorithms generate multiple design options based on defined constraints and objectives.
Generative Model: An AI model designed to generate new data that resembles the patterns and characteristics of the training data it has been exposed to. Focused AI models can be used to predict the outcomes of hypothetical phenomena (such as how a car might crumple during a crash).
Generative Summarization (Abstractive Summarization): Using LLM functionality to take text prompt inputs like long-form chats, emails, reports, contracts, policies, etc., and distilling them down to their core content, generating summaries from the text prompts for quick comprehension. Thus using pre-trained language models and context understanding to produce concise, accurate, and relevant summaries.
Genetic Algorithm: A heuristic optimization technique inspired by natural selection and genetic processes to solve complex optimization problems.
Global AI Strategies: Coordinated AI policies and regulations between governments aimed at fostering advancements and collaborations in artificial intelligence research and development.
GLoVe (Global Vectors for Word Representation): A word embedding technique capturing semantic relationships between words based on global co-occurrence statistics.
Google Bard: An AI chatbot by Google that functions similarly to ChatGPT but pulls information from the current web, whereas ChatGPT is limited to data until 2021 and isn’t connected to the internet.
GPT (Generative Pre-Trained Transformer): Generative pre-trained transformers (GPT) are neural network models trained on large datasets in an unsupervised manner to generate text.
GPT-3: GPT-3 is the 3rd version of the GPT-n series of models. It has 175 billion parameters — knobs that can be tuned — with weights to make predictions. Chat-GPT uses GPT-3.5, which is another iteration of this model.
GPT-4: GPT-4 is the latest model addition to OpenAI’s deep learning efforts and is a significant milestone in scaling deep learning. GPT-4 is also the first of the GPT models that is a large multimodal model, meaning it accepts both image and text inputs and emits text outputs.
Gradient: The rate of change of a function concerning its input variables, indicating the direction and magnitude of the change. This is useful in optimization algorithms like gradient descent to iteratively adjust model parameters and minimize errors.
Gradient Descent: A fundamental optimization algorithm using gradients to minimize errors in machine learning models by iteratively adjusting parameters.
Graph-Based Model: A machine learning model that uses graph structures to represent and analyze relationships between data points.
Grid Search: A hyperparameter tuning technique involving systematic exploration of a predefined hyperparameter space to optimize model performance.
Grounding: Grounding is the process of anchoring artificial intelligence (AI) systems in real-world experiences, factual knowledge, or data. The objective is to improve the AI’s understanding of the world, so it can effectively interpret and respond to user inputs, queries, and tasks. Grounding helps AI systems become more context-aware, allowing them to provide better, more relatable, and relevant responses or actions.
GRU (Gated Recurrent Units): A gating mechanism in recurrent neural networks designed to address the vanishing gradient problem, enhancing learning in sequences.
Guardrails: Policies and restrictions placed on AI models to ensure data is handled responsibly and that the model doesn’t create disturbing content.
H
Hallucination: Hallucinations refer to incorrect, irrelevant, or nonsensical outputs generated by an AI system, particularly in natural language processing, due to its inability to understand the context or subject matter accurately. These outputs may include fabricated facts, inaccurate information, or misaligned references presented as truth, often resulting from the AI’s overreliance on patterns learned from its training data.
Heat Map: A visual representation that highlights important elements in the output generated by an AI model. It helps understand where the model focuses and assists in evaluating and improving the generated content.
Hierarchical RL: A reinforcement learning paradigm focused on breaking down large processes into simple subtasks to improve decision-making across multiple levels of abstraction or hierarchy.
Hindsight Experience Replay: A technique in reinforcement learning where failed experiences are replayed with alternate goals to improve learning efficiency.
Human-AI Collaboration: Cooperation between humans and artificial intelligence to collectively accomplish tasks by leveraging their respective strengths.
Human-Centered Machine Learning: An approach to machine learning that emphasizes human needs, ethics, and user experience in model development.
Human-in-the-Loop AI: A paradigm where human input is integrated into AI processes, enhancing model performance and accountability.
Hybrid AI: Hybrid AI is any artificial intelligence technology that combines multiple AI methodologies. In NLP, this often means that a workflow will leverage both symbolic and machine learning techniques.
Hyperparameter: A configurable parameter setting that influences training behavior and performance in machine learning models.
Hyperparameter Tuning: The process of optimizing hyperparameters to enhance machine learning model performance and generalization.
I
ICM (Intrinsic Curiosity Module): A type of intrinsic motivation in reinforcement learning where agents are driven by internal curiosity signals to explore their environment.
IDP (Intelligent Document Processing), also known as IDEP (Intelligent Document Extraction and Processing): This is the ability to automatically read and convert unstructured and semi-structured data, identify usable data and extract it, then leverage it via automated processes. IDP is often an enabling technology for Robotic Process Automation (RPA) tasks.
Image Recognition: The process of identifying an object, person, place, or text in an image or video.
Image Segmentation: A computer vision task involving the division of an image into distinct segments to enable object detection and localization.
Image Synthesis: The process of generating new images from existing ones using machine learning models, vital for creative applications.
Image-to-Image Translation: An advanced computer vision technique that transforms images from one domain into another to learn the mapping between input and output images.
Image-to-Text Generation: A challenging natural language processing task in which a model generates text descriptions of input images.
Imitation Learning: An approach in machine learning where models learn by imitating human behavior, often used for training autonomous systems.
Inference Engine: A component of a [expert] system that applies logical rules to the knowledge base to deduce new or additional information.
Insight Engines: An insight engine, also called cognitive search or enterprise knowledge discovery. It applies relevancy methods to describe, discover, organize, and analyze data. It combines search with AI capabilities to provide information for users and data for machines. The goal of an insight engine is to provide timely data that delivers actionable intelligence.
Instance: A single data point in a dataset, characterized by a set of features and, in supervised learning, associated with a label.
Instance Segmentation: A computer vision task that involves identifying and categorizing individual objects within an image, crucial for object detection.
Instruction-tuning: Instruction-tuning is an approach where a pre-trained model is adapted to perform specific tasks by providing a set of guidelines or directives that outline the desired operation.
Intelligence Augmentation: Intelligence augmentation refers to empowering human capabilities through synergistic combinations of AI systems and traditional tools.
Interpretability: The study of explaining and understanding the decision-making processes of machine learning models, and of designing systems whose decisions can be easily understood. Also known as explainability.
Intrinsic Motivation: A mechanism in reinforcement learning where models are driven to exhibit inherently rewarding behaviors like exploration and curiosity. Modeled after the psychological concept of the same name.
Inverse Reinforcement Learning: A technique where agents learn an underlying reward function from observed human behavior.
J
K
K-Shot Learning: K-shot learning is a machine learning approach where models learn from only k labeled examples per class, where k is a small number like 1-5.
Knowledge Distillation: A technique where a complex model transfers its knowledge to a smaller, simpler model, enhancing efficiency without losing validity.
Knowledge Engineering: A method for helping computers replicate human-like knowledge. Knowledge engineers build logic into knowledge-based systems by acquiring, modeling, and integrating general or domain-specific knowledge into a model.
Knowledge Generation: Knowledge generation involves training models on extensive datasets, allowing them to analyze data, discover patterns, and craft new insights.
Knowledge Graph: A knowledge graph is a graph of concepts whose value resides in its ability to meaningfully represent a portion of reality, specialized or otherwise. Every concept is linked to at least one other concept, and the quality of this connection can belong to different classes (see: taxonomies).
The interpretation of every concept is represented by its links. Consequently, every node is the concept it represents only based on its position in the graph (e.g., the concept of an apple, the fruit, is a node whose parents are “apple tree”, “fruit”, etc.). Advanced knowledge graphs can have many properties attached to a node including the words used in language to represent a concept (e.g., “apple” for the concept of an apple), if it carries a particular sentiment in a culture (“bad”, “beautiful”) and how it behaves in a sentence.
Knowledge Model: A process of creating a computer-interpretable model of knowledge or standards about a language, domain, or process(es). It is expressed in a data structure that enables the knowledge to be stored in a database and be interpreted by software.
Knowledge-Based AI: Knowledge-based systems (KBs) are a form of artificial intelligence (AI) designed to capture the knowledge of human experts to support decision-making and problem-solving.
L
Labelled Data: see Data Labelling.
Labeling: The process of annotating data with class labels, enabling supervised machine learning.
LangOps (Language Operations): The workflows and practices that support the training, creation, testing, production deployment and ongoing curation of language models and natural language solutions.
Language Data: Language data is data made up of words; it is a form of unstructured data. This is qualitative data and also known as text data, but simply it refers to the written and spoken words in language.
Latency: Latency refers to the time delay between when an AI system receives an input and generates the corresponding output.
Lemma: The base form of a word representing all its inflected forms.
Lemmatization: A text normalization technique in natural language processing that transforms words into their base or root form, enhancing language processing efficiency.
Lexicon: Knowledge of all of the possible meanings of words, in their proper context; is fundamental for processing text content with high precision.
Lifelong Learning: An approach where machine learning models continually learn from new data and adapt to evolving challenges. Also known as continual learning.
LIME (Local Interpretable Model-Agnostic Explanations): A technique in interpretability that uses a local, interpretable model to provide human-understandable explanations for the predictions of a black box model.
Linked Data: Linked data is an expression that informs whether a recognizable store of knowledge is connected to another one. This is typically used as a standard reference. For instance, a knowledge graph in which every concept/node is linked to its respective page on Wikipedia.
LLM (Large Language Model): A machine learning model trained to understand and generate human language, often used in natural language processing tasks. They are trained on vast datasets of documents from various sources, analyzing input data, mapping out words, and predicting the most likely word combinations based on patterns discovered during training. LLMs are characterized by their extensive capacity to process and generate human-like language.
Loss Function: A mathematical function used to measure the difference between predicted and actual values in machine learning models.
Low-code: Low-code is a visual approach to software development that enables faster delivery of applications through minimal hand-coding.
LSTM (Long Short-Term Memory): A type of recurrent neural network architecture designed to address the vanishing gradient problem, vital for sequential data processing.
M
Machine Learning (ML): A subfield of Artificial Intelligence (AI) that involves the study and development of computer algorithms and statistical models that enable computers to automatically improve their performance and make informed decisions or predictions through experience and data, without being explicitly programmed. These algorithms are built on sample data, known as “training data,” in order to make predictions or decisions.
Example: A machine learning algorithm that can predict which customers are most likely to churn based on their past behavior.
Machine Learning Algorithms: Mathematical models and techniques employed to discern and understand inherent patterns within datasets, aiding in predictions, decisions, and insights.
Machine Translation: Translating text from one language to another using machine learning models.
MDP (Markov Decision Process): A mathematical framework used to model decision-making in scenarios with sequential actions and uncertain outcomes.
Memory Networks: An AI model that combines reasoning abilities with a long-term memory component, learning to effectively utilize both. The long-term memory is read from and written to, aiming to enhance predictive capabilities. These networks are particularly explored in question-answering contexts, utilizing long-term memory as a dynamic knowledge base to generate relevant textual responses.
Meta Reinforcement Learning: A higher-level reinforcement learning paradigm where agents learn to adapt and generalize their skills across different tasks.
Meta-Learning: An approach where machine learning models learn to learn, acquiring knowledge that helps them rapidly adapt to new tasks. Meta-learning involves utilizing machine learning algorithms to effectively integrate predictions from other machine learning models.
Metadata: Data that describes or provides information about other data.
MCTS (Monte Carlo Tree Search): A decision-making algorithm often used in game-playing AI. It excels in managing intricate and strategic video games with vast search spaces, a challenge where conventional algorithms might falter due to the overwhelming number of possible actions.
Microsoft Bing: A search engine by Microsoft that can now use the technology powering ChatGPT to give AI-powered search results. It’s similar to Google Bard in being connected to the internet.
MLOps (Machine Learning Operations): The practice of streamlining the deployment, management, and monitoring of machine learning models in real-world applications.
Model: A machine learning model is the artifact produced after an ML algorithm has processed the sample data it was fed during the training phase. The model is then used by the algorithm in production to analyze text (in the case of NLP) and return information and/or predictions.
Model Chaining: Model chaining is a technique in data science where multiple machine learning models are linked in a sequence to make predictions or analyzations.
Model Compression: The process of optimizing deep learning models for deployment on resource-constrained devices without compromising performance, allowing for edge learning.
Model Deployment: The operationalization of machine learning models, making them accessible for real-world applications and interactions.
Model Drift: Model drift is the decay of models’ predictive power as a result of the changes in real-world environments. It is caused due to a variety of reasons including changes in the digital environment and ensuing changes in relationship between variables.
Example: A model that detects spam based on email content and then the content used in spam was changed.
Model Parameter: These are parameters in the model that are determined by using the training data. They are the fitted/configured variables internal to the model whose value can be estimated from data. They are required by the model when making predictions. Their values define the capability and fit of the model.
Model Quantization: A model compression technique that reduces model memory and computational requirements by representing parameters with fewer bits.
Model Robustness: The ability of a machine learning model to perform consistently across different input data and scenarios.
Model Validation: The process of assessing a machine learning model’s performance and generalization capabilities on unseen data.
Model-Based RL: A reinforcement learning approach where agents learn a model of the environment to make decisions and plan actions.
Model-Free RL (Reinforcement Learning): A method in reinforcement learning that doesn’t rely on the transition probability distribution or reward function of the problem’s decision process, collectively known as the “model.” It instead operates on trial-and-error, exploring various solutions to optimize and achieve the most favorable outcome.
Momentum: A technique in gradient-based optimization algorithms that accelerates convergence by incorporating past gradients.
Monte Carlo Method: A stochastic algorithm that estimates mathematical values by generating random samples or simulations.
Morphological Analysis: Breaking a problem with many known solutions down into its most basic elements or forms, in order to more completely understand them. Morphological analysis is used in general problem-solving, linguistics, and biology.
Motion Prediction: The task of forecasting future trajectories or actions of objects, often used in autonomous systems.
MPT (Multitask prompt tuning): An approach that configures a prompt representing a variable — that can be changed — to allow repetitive prompts where only the variable changes.
Multi-Agent RL: A subfield of reinforcement learning where multiple agents interact and learn in a shared environment.
Multi-Armed Bandit: A classic optimization problem where an agent chooses between multiple actions (“arms”) to maximize cumulative rewards.
Multi-Dimensional Scaling (MDS): A technique for visualizing high-dimensional data by projecting it onto a lower-dimensional space while preserving data similarity.
Multi-Head Attention: An attention mechanism that allows a model to focus on different aspects of input data simultaneously.
Multi-Instance Learning: A machine learning paradigm where each example in a dataset comprises multiple instances, enabling learning from groups of data.
Multi-Label Classification: A classification task where each data instance can be assigned multiple class labels.
Multi-Objective RL: A reinforcement learning scenario where agents optimize multiple objectives simultaneously, often leading to trade-offs.
Multi-hop Reasoning: Multi-hop is a term often used in natural language processing and, more specifically, machine reading comprehension tasks. It refers to the process by which an AI model retrieves answers to questions by connecting multiple pieces of information present in a given text or across various sources and systems, rather than directly extracting the information from a single passage.
Multilingual Models: Machine learning models capable of understanding and generating text in multiple languages.
Multimodal AI: A type of AI that can process multiple types of inputs, including text, images, videos and speech.
Multimodal Learning: A deep learning model trained on large datasets of both textual and non-textual data (e.g., text, image, audio) to improve model performance.
Multitask Learning: An approach where a single machine learning model is trained to perform multiple related tasks simultaneously.
Music Generation: The process of creating new AI music compositions using machine learning models.
N
N-Gram: A contiguous sequence of n items (usually words) within a block of text, often used in language modeling and text analysis.
N-Shot Learning: Zero/Single/Few shot learning are variations of the same concept – providing a model with little or no training data to classify new data and guide predictions. A “shot” represents a single training example. Fun fact: Within the GPT prompt, you can ask for “N” examples to improve the accuracy of the response.
Narrow AI: An AI model that can perform one specific task, and can’t generalize its experience to other tasks.
Natural Language Ambiguity: Natural language ambiguity refers to situations where a word, phrase, or sentence can have multiple meanings, making it challenging for both humans and AI systems to interpret correctly.
Natural Language Generation (NLG): A subfield of AI that produces human-like natural written or spoken language from structured data or other information sources.
Natural Language Processing (NLP): A subfield of artificial intelligence (AI) that focuses on enabling computers to interact with, process, and generate text or speech as humans do. It involves algorithms and models that help it understand, interpret, and generate human language in a way that is both meaningful and contextually relevant.
Natural Language Query (NLQ): A natural language input that only includes terms and phrases as they occur in spoken language (i.e., without non-language characters).
Natural Language Technology (NLT): A subfield of linguistics, computer science, and artificial intelligence (AI) dealing with Natural Language Processing (NLP), Natural Language Understanding (NLU), and Natural Language Generation (NLG).
Natural Language Understanding (NLU): A subtopic of Natural Language Processing (NLP) that focuses on enabling machines to comprehend and interpret the semantic meaning of human language by analyzing context, sentiment, and intent in unstructured text data.
Named Entity Recognition (NER): A natural language processing task involving the identification and categorization of named entities in text.
NeRF (Neural Radiance Fields): A technique for generating detailed 3D reconstructions of objects and scenes from 2D images.
Neural Architecture Search (NAS): An automated technique that explores and discovers optimal neural network architectures for specific tasks.
Neural Network: A machine learning model inspired by the human brain’s structure used to analyze and interpret complex data patterns. The model is composed of layers of interconnected nodes or “neurons.”
Example: A neural network that can recognize handwritten digits with high accuracy.
Neural Network Pruning: The process of reducing the size of a neural network by removing unnecessary connections.
Neural Rendering: A computer graphics technique that uses neural networks to generate realistic images based on existing scenes by simulating light transport.
Neural Style Transfer: A transformative technique that combines the content of one image with the artistic style of another, yielding unique visual effects.
Neural Turing Machine (NTM): A neural network controller with differentiable interactions with external memory that it interacts with via attention mechanisms, allowing optimization through gradient descent.
Neuromorphic Computing: Computer architecture that mimics the human brain’s structure using electronic circuits, aiding tasks like pattern recognition and cognitive processes.
Neurosymbolic AI: A model that fuses statistical AI and symbolic reasoning, aiming to achieve general AI capabilities by combining data-driven and logic-based approaches.
No-code: No-code is an approach to designing and using applications that doesn’t require any coding or knowledge of programming languages.
NTM (Neural Turing Machine): A neural network controller with differentiable interactions with external memory that it interacts with via attention mechanisms, allowing optimization through gradient descent.
O
Object Detection: A computer vision technique that uses instance segmentation to identify objects in images or videos, crucial for applications like autonomous driving and surveillance.
Off-Policy Learning: A reinforcement learning method that improves actions by evaluating and refining previously collected data.
One-Hot Encoding: A transformation of categorical data into binary vectors, a common technique for representing categorical variables in machine learning.
One-Shot Learning: The practice of training models with limited examples to recognize new objects or concepts. This is valuable when available data is limited.
Ontology: An ontology is similar to a taxonomy, but it enhances its simple tree-like classification structure by adding properties to each node/element and connections between nodes that can extend to other branches. These properties are not standard, nor are they limited to a predefined set. Therefore, they must be agreed upon by the classifier and the user.
OOV (Out of Vocabulary): A term absent from a language model’s training data, posing challenges in understanding new language.
OpenAI: The organization that developed ChatGPT. More broadly speaking, OpenAI is a research company that aims to develop and promote friendly AI responsibly.
Example: OpenAI’s GPT-3 model is one of the largest and most powerful language models available for natural language processing tasks.
OpenAI CLIP: A model unifying vision and language understanding to associate images with textual descriptions.
OpenAI Codex: An AI system that translates human language into code, streamlining programming tasks by generating code snippets.
OpenAI DALL-E: A deep learning model that generates images from textual descriptions, showcasing the potential of AI in content creation.
OpenAI GPT-3: A language model that generates human-like text, applicable in chatbots, content creation, and more.
OpenAI GPT-4: The successor to GPT-3, advancing natural language understanding and generation capabilities.
Open-Source AI: The practice of openly sharing AI project source code for collaborative development, enabling community contribution and innovation.
Optical Flow: A computer vision technique estimating object movement and velocity in images or videos, useful in tracking and motion analysis.
Optimization: The process of adjusting the parameters of a model to minimize a loss function that measures the difference between the model’s predictions and the true values.
Example: Optimizing a neural network’s parameters using a gradient descent algorithm to minimize the error between the model’s predictions and the true values.
Option-Critic Architecture: A reinforcement learning framework incorporating “options” to enhance agents’ decision-making flexibility.
Out-of-Distribution Detection: The task of identifying data instances in machine learning that deviate significantly from the trained model’s input distribution. This is useful to defend against adversarial attacks.
Overfitting: A problem that occurs when a model is too complex, performing well on the training data but poorly on unseen data.
Example: A model that has memorized the training data instead of learning general patterns and thus performs poorly on new data.
Overfitting: When a machine learning model performs well on training data but poorly on new, unseen data due to excessive fitting.
Output: The generated content produced by a generative AI system. It can be text, images, audio, music, video, or other data the model is designed to produce.
P
Panoptic Segmentation: An advanced computer vision task combining instance and semantic segmentation, enabling the model to comprehend an entire scene.
Paperclips: The Paperclip Maximiser theory, coined by philosopher Nick Boström of the University of Oxford, is a hypothetical scenario where an AI system will create as many literal paperclips as possible. In its goal to produce the maximum amount of paperclips, an AI system would hypothetically consume or convert all materials to achieve its goal. This could include dismantling other machinery to produce more paperclips, machinery that could be beneficial to humans. The unintended consequence of this AI system is that it may destroy humanity in its goal to make paperclips.
Parameter: A set of numerical weights representing neural connections or other aspects in an AI model with values that are determined by training. Large language models (LLMs) can have billions of parameters.
Parameter-efficient Fine-tuning (PEFT): Parameter-Efficient Fine-Tuning, also known as PEFT, is an approach that helps you improve the performance of large AI models while optimizing for resources like time, energy, and computational power. To do this, PEFT focuses on adjusting a small number of key parameters while preserving most of the pretrained model’s structure.
Parameters: Numerical values that give LLMs structure and behavior, enabling it to make predictions.
Parsing: Identifying the single elements that constitute a text, then assigning them their logical and grammatical value.
Part of Speech Tagging: Labeling words in text with their grammatical roles (e.g., noun, verb). This is important for language-understanding tasks.
Part-of-Speech (POS) Tagging: A Part-of-Speech (POS) tagger is an NLP function that identifies grammatical information about the elements of a sentence. Basic POS tagging can be limited to labeling every word by grammar type, while more complex implementations can group phrases and other elements in a clause, recognize different types of clauses, build a dependency tree of a sentence, and even assign a logical function to every word (e.g., subject, predicate, temporal adjunct, etc.).
Partially Observable Markov Decision Process (POMDP): A generalization of the Markov decision process (MDP) designed to model decision-making in uncertain environments where agents lack complete information.
Pattern Recognition: The method of using computer algorithms to analyze, detect, and label regularities in data. This informs how the data gets classified into different categories.
Pattern recognition: Pattern recognition is the method of using computer algorithms to analyze, detect, and label regularities in data. This informs how the data gets classified into different categories.
PCA (Principal Component Analysis): A dimensionality reduction technique that facilitates visualization and analysis of complex datasets.
PEFT (Parameter-efficient Fine-tuning): Parameter-Efficient Fine-Tuning, also known as PEFT, is an approach that helps you improve the performance of large AI models while optimizing for resources like time, energy, and computational power. To do this, PEFT focuses on adjusting a small number of key parameters while preserving most of the pretrained model’s structure.
PEMT (Post Edit Machine Translation): Solution allows a translator to edit a document that has already been machine translated. Typically, this is done sentence-by-sentence using a specialized computer-assisted-translation application.
Perplexity: A measure used to assess the coherence and consistency of AI-generated text. Higher perplexity values suggest the content is more likely to be AI-generated due to unusual patterns or inconsistencies. Content identification systems use perplexity to identify AI-generated content.
Pix2Pix: A generative adversarial network (GAN) that converts images from one format to another, valuable for style transfer and image enhancement.
Plugins: A software component or module that extends the functionality of an LLM system into a wide range of areas, including travel reservations, e-commerce, web browsing and mathematical calculations.
Point Cloud-Based Model: A model that uses a large collection of small data points to process and represent 3D information, important in various fields like robotics and computer graphics.
Policy: A set of rules, strategies, or instructions that guide decision-making within an autonomous agent or system. It outlines the agent’s course of action based on its current state and the surrounding environment. Policies can be learned through reinforcement learning.
Policy Gradient: A group of reinforcement learning techniques that optimize policies through gradient descent.
POMDP (Partially Observable Markov Decision Process): A generalization of the Markov decision process (MDP) designed to model decision-making in uncertain environments where agents lack complete information.
Pose Estimation: A computer vision task estimating positions and orientations of objects or human body parts in images or videos.
Positional Encoding: A technique that assigns a number to each word during training that is used to show the position (or order) of words in a sequence.
Post-processing: Procedures that can include various pruning routines, rule filtering, or even knowledge integration. All these procedures provide a kind of symbolic filter for noisy and imprecise knowledge derived by an algorithm.
PPO (Proximal Policy Optimization): A policy gradient method for reinforcement learning that optimizes policies gradually to ensure stable learning.
Pre-processing: A step in the data mining and data analysis process that takes raw data and transforms it into a format that can be understood and analyzed by computers. Analyzing structured data, like whole numbers, dates, currency, and percentages is straightforward. However, unstructured data, in the form of text and images must first be cleaned and formatted before analysis.
Pre-training: Training a model on a large dataset before fine-tuning it to a specific task. Example: Pre-training a language model like ChatGPT on a large corpus of text data before fine-tuning it for a specific natural language task such as language translation.
Precision, Recall, and F1 Score: Metrics evaluating classification model performance based on true positives, false positives, and false negatives.
Precision: Given a set of results from a processed document, precision is the percentage value that indicates how many of those results are correct based on the expectations of a certain application. It can apply to any class of a predictive AI system such as search, categorization and entity recognition.
Example: You have an application that is supposed to find all the dog breeds in a document. If the application analyzes a document that mentions 10 dog breeds but only returns five values (all of which are correct), the system will have performed at 100% precision. Even if half of the instances of dog breeds were missed, the ones that were returned were correct.
Predictive analytics: Predictive analytics is a type of analytics that uses technology to predict what will happen in a specific time frame by analyzing historical data and patterns for factors such as possible situations and scenarios, past and present performance, and other resources to help organizations make better strategic decisions.
Pretrained model: A model trained to accomplish a task — typically one that is relevant to multiple organizations or contexts. Also, a pretrained model can be used as a starting point to create a fine-tuned contextualized version of a model, thus applying transfer learning.
Pretraining: The first step in training a foundation model, usually done as an unsupervised learning phase. Once foundation models are pretrained, they have a general capability. However, foundation models need to be improved through fine-tuning to gain greater accuracy.
Prioritized Experience Replay: A reinforcement learning technique that replays experiences with higher learning importance to improve training efficiency.
Privacy in AI: The practice of following considerations regarding safeguarding user data and maintaining privacy while utilizing AI technologies.
Probabilistic: In generative AI, probabilistic means that the models incorporate probability, which is used to estimate the likelihood of different outcomes and generate outputs that align with the learned probabilities.
Probabilistic Model: A probabilistic AI model makes decisions based on probabilities or likelihoods.
Procedural Generation: A type of algorithmic content creation, particularly prominent in video game development, that generates new gameplay elements like character designs, animations, and environments during gameplay.
ProGAN: A generative adversarial network that uses progressive growth to produce high-resolution images.
Progressive Growth of GANs: A training technique that increases GAN image resolution gradually, enhancing image quality and diversity.
Prompt: A phrase or individual keywords used as the initial input text or instructions given to a GenAI model to generate new content based on that starting point. It provides context and guides the model’s output. The prompt can be a few words or sentences that set the tone or specify the desired content.
Prompt chaining: An approach that uses multiple prompts to refine a request made by a model.
Prompt Engineering: Identifying inputs — prompts — that result in meaningful outputs. As of now, prompt engineering is essential for LLMs. LLMs are a fusion of layers of algorithms and, consequently, have limited controllability with few opportunities to control and override behavior. An example of prompt engineering is providing a collection of templates and wizards to direct a copywriting application.
Prompt Templates: Prompt templates are pre-defined recipes for generating prompts for a language models. AIPRM manages prompt templates of many types, augments them this user-defined variables and content from custom indexes (RAG).
Proximal Policy Optimization (PPO): A policy gradient method for reinforcement learning that optimizes policies gradually to ensure stable learning.
Q
Q-Function: A function that uses the Q-value to estimate future rewards for taking actions in given states, central to reinforcement learning.
Q-Learning: A model-free reinforcement learning method using the Q-function to find optimal action choices through trial and error.
Q-Value: A key value in reinforcement learning that quantifies the expected cumulative reward that will result from taking a specific action in a given state. It reflects the model’s learned knowledge about the potential outcomes and benefits of various actions. It helps the model make optimal decisions by guiding its choice of actions to maximize long-term rewards.
Quantum computing: Quantum computing is the process of using quantum-mechanical phenomena such as entanglement and superposition to perform calculations. Quantum machine learning uses these algorithms on quantum computers to expedite work because it performs much faster than a classic machine learning program and computer.
Quantum Machine Learning: Integration of quantum computing techniques into machine learning for enhanced computational power.
Question & Answer (Q&A): An AI technique that allows users to ask questions using common everyday language and receive the correct response back. With the advent of large language models (LLMs), question and answering has evolved to let users ask questions using common everyday language and use Retrieval Augmented Generation (RAG) approaches to generate a complete answer from the text fragments identified in the target document or corpus.
Question-Answering Systems: An AI system designed to comprehend and respond to questions posed in natural language.
R
RAG (Retrieval-Augmented Generation): RAG, or Retrieval-Augmented Generation, is an advanced AI framework that enhances large language models by pulling in current and precise information from external databases. This process ensures that the responses provided by these models are not only accurate but also up-to-date. Additionally, it offers users a clearer understanding of how these models generate their answers.
Rainbow DQN: The integration of multiple reinforcement learning techniques to improve deep Q-network performance.
Random Forest: A supervised machine learning algorithm that grows and combines multiple decision trees to create a “forest.” Used for both classification and regression problems in R and Python.
Random Search: A hyperparameter tuning strategy involving random parameter selection to find optimal configurations.
RBMT (Rules-based Machine Translation): Considered the “Classical Approach” of machine translation, it is based on linguistic information about source and target that allow words to have different meaning depending on the context.
Reasoning: AI reasoning is the process by which artificial intelligence systems solve problems, think critically, and create new knowledge by analyzing and processing available information, allowing them to make well-informed decisions across various tasks and domains.
Recall: Given a set of results from a processed document, recall is the percentage value that indicates how many correct results have been retrieved based on the expectations of the application. It can apply to any class of a predictive AI system such as search, categorization and entity recognition.
Example: You have an application that is supposed to find all the dog breeds in a document. If the application analyzes a document that mentions 10 dog breeds but only returns five values (all of which are correct), the system will have performed at 50% recall.
Rectified Linear Unit (ReLU): Activation function commonly used in neural networks, introducing non-linearity.
Recursive Prompting: Recursive prompting is a strategy for guiding AI models like OpenAI’s GPT-4 to produce higher-quality output. It involves providing the model with a series of prompts or questions that build upon previous responses, refining both the context and the AI’s understanding to achieve the desired result.
Recurrent Neural Networks (RNN): A neural network model commonly used in natural language processes and speech recognition allowing previous outputs to be used as inputs.
Regularization: Techniques preventing model overfitting by adding constraints to the learning process.
Reinforcement Learning: A machine learning method where agents learn to make decisions and improve their actions by interacting with an environment, receiving feedback through rewards or penalties, and incorporating guidance from human feedback to maximize rewards in real-world scenarios such as robotics and autonomous systems.
Relations: The identification of relationships is an advanced NLP function that presents information on how elements of a statement are related to each other. For example, “John is Mary’s father” will report that John and Mary are connected, and this datapoint will carry a link property that labels the connection as “family” or “parent-child.”
ReLU (Rectified Linear Unit): Activation function commonly used in neural networks, introducing non-linearity.
Responsible AI: Development and deployment of AI systems while considering ethical, societal, and transparency aspects.
Restricted Boltzmann Machines (RBMs): Unsupervised neural networks used for feature extraction and data representation.
Reward Shaping: Adapting reward functions in reinforcement learning to guide agents toward desired behaviors.
RL Simulation Environments: Simulated settings used to train and test reinforcement learning agents, allowing safe and efficient learning.
RLM (Representation learning)
RNN (Recurrent Neural Network): A neural network model commonly used in natural language process and speech recognition allowing previous outputs to be used as inputs.
RoBERTa (Robustly Optimized BERT-Pretraining Approach): An advanced pretraining technique for natural language processing, extending the capabilities of BERT.
Robustness in AI: The ability of AI models to perform consistently across various conditions and inputs.
S
SAC (Soft Actor-Critic): A reinforcement learning algorithm optimizing policies for maximum reward while also introducing entropy. It attempts to succeed at the task while acting as randomly as possible.
Saliency Maps: Visualizations that highlight influential regions in input data, helping to interpret model decisions.
SAO (Subject-Action-Object): Subject-Action-Object (SAO) is an NLP function that identifies the logical function of portions of sentences in terms of the elements that are acting as the subject of an action, the action itself, the object receiving the action (if one exists), and any adjuncts if present.
Scene Understanding: An AI’s capacity to comprehend and interpret visual scenes, crucial for computer vision applications like autonomous driving.
Self-Attention: An attention mechanism enabling neural networks to weigh the importance of different positions within a sequence to capture relationships and dependencies.
Self-Play in RL: A reinforcement learning technique where agents improve through self-generated challenges by competing against previous versions.
Self-Supervised Learning: A learning approach where AI systems generate their own training labels from available data.
Semantic Network: A form of knowledge representation, used in several natural language processing applications, where concepts are connected to each other by semantic relationship.
Semantic Search: The use of natural language technologies to improve user search capabilities by processing the relationship and underlying intent between words by identifying concepts and entities such as people and organizations are revealed along with their attributes and relationships.
Semantics: Semantics is the study of the meaning of words and sentences. It concerns the relation of linguistic forms to non-linguistic concepts and mental representations to explain how sentences are understood by the speakers of a language.
Semi-structured Data: Data that is structured in some way but does not obey the tabular structure of traditional databases or other conventional data tables most commonly organized in rows and columns. Attributes of the data are different even though they may be grouped together. A simple example is a form; a more advanced example is an object database where the data is represented in the form of objects that are related (e.g., automobile make relates to model relates to trim level).
Semi-Supervised Learning: A learning paradigm that combines labeled and unlabeled data to enhance model training.
Sentiment: Sentiment is the general disposition expressed in a text.
Sentiment Analysis: A model’s ability to assess sentiment expressed in text, often used in social media analysis. This can be applied to anything from a business document to a social media post. Sentiment is typically measured on a linear scale (negative, neutral or positive), but advanced implementations can categorize text in terms of emotions, moods, and feelings.
Sentient: The capability to possess consciousness, self-awareness, and subjective experiences. Achieving true sentience in AI systems is a topic of scientific exploration and philosophical debate.
Seq2Seq Models (Sequence-to-Sequence Models): Neural architectures that translate sequences from one domain to another, which are central to machine translation.
Sequence Generation: AI’s capability to create sequential data, prominently used in language generation tasks.
Sequence Modeling: A subfield of NLP that focuses on modeling sequential data such as text, speech, or time series data.
Example: A sequence model that can predict the next word in a sentence or generate coherent text.
Sequential Data: A type of information where the order and arrangement of elements hold significance, such as text or an event series.
SGD (Stochastic Gradient Descent): A variant of gradient descent that randomly samples a subset (or “mini-batch”) of the data in each iteration and uses only that subset to reduce error. This can lead to faster convergence and reduced computational requirements.
SHAP (SHapley Additive Explanations): Method for explaining machine learning model outputs, attributing predictions to input features.
Sigmoid Function: A mathematical function used in artificial neural networks to introduce non-linearity into the model. It maps input values to a range between 0 and 1 and smoothly transitions as inputs vary, resulting in an S-shaped (or sigmoid) curve. It’s particularly useful for binary classification tasks with outputs in the form of a probability score.
Similarity (and Correlation): Similarity is an NLP function that retrieves documents similar to a given document. It usually offers a score to indicate the closeness of each document to that used in a query. However, there are no standard ways to measure similarity. Thus, this measurement is often specific to an application versus generic or industry-wide use cases.
Simple Knowledge Organization System (SKOS): A common data model for knowledge organization systems such as thesauri, classification schemes, subject heading systems, and taxonomies.
SMOTE (Synthetic Minority Over-sampling Technique): A technique for generating synthetic samples to address a class imbalance in datasets.
Soft Actor-Critic (SAC): A reinforcement learning algorithm optimizing policies for maximum reward while also introducing entropy. It attempts to succeed at the task while acting as randomly as possible.
Softmax Function: A mathematical function that transforms raw scores into probability distributions, resulting in probabilities for each possible outcome.
Spatiotemporal Data Analysis: Examination of data with both spatial and temporal dimensions, crucial for understanding dynamic processes.
Spatiotemporal Sequence Forecasting: The process of collecting data across both space and time and using it to predict future developments over time, used in fields like climate modeling.
Specialized corpora: A focused collection of information or training data used to train an AI. Specialized corpora focuses on an industry — for example, banking, Insurance or health — or on a specific business or use case, such as legal documents.
Speech Analytics: The process of analyzing recordings or live calls with speech recognition software to find useful information and provide quality assurance. Speech analytics software identifies words and analyzes audio patterns to detect emotions and stress in a speaker’s voice.
Speech Recognition: Speech recognition or automatic speech recognition (ASR), computer speech recognition, or speech-to-text, enables a software program to process human speech into a written/text format.
Speech-to-Text Conversion: The conversion of spoken language into written text, essential for transcription and voice assistants.
Stacking: Stacking is a technique in AI that combines multiple algorithms to enhance overall performance. By blending the strengths of various AI models, stacking compensates for each model’s weaknesses and achieves a more accurate and robust output in diverse applications, such as image recognition and natural language processing.
Stable Diffusion: Stable diffusion is an artificial intelligence system that uses deep learning to generate images from text prompts.
Steerability: AI steerability refers to the ability to guide or control an AI system’s behavior and output according to human intentions or specific objectives. This involves designing AI models with mechanisms that understand and adhere to the preferences provided by users, while avoiding unintended or undesirable outcomes. Improving steerability requires ongoing research and refinement, including techniques like fine-tuning, rule-based systems, and implementing additional human feedback loops during AI development.
Stemming: The process of reducing words to their base form to simplify analysis, commonly applied in natural language processing.
Stochastic Gradient Descent (SGD): A variant of gradient descent that randomly samples a subset (or “mini-batch”) of the data in each iteration and uses only that subset to reduce error. This can lead to faster convergence and reduced computational requirements.
Stochastic Parrot: Stochastic parrots are AI systems that use statistics to convincingly generate human-like text, while lacking true semantic understanding behind the word patterns.
Stop Words: Frequently used words like articles that are disregarded in text analysis due to their minimal informative value.
Strong AI: Strong AI refers to machines possessing generalized intelligence and capabilities on par with human cognition.
Structured Data: Structured data is the data which conforms to a specific data model, has a well-defined structure, follows a consistent order and can be easily accessed and used by a person or a computer program. Structured data are usually stored in rigid schemas such as databases.
Style Transfer: A technique merging the artistic style of one image with the content from another, producing novel visuals.
StyleGAN: A generative adversarial network specializing in generating images with specific artistic styles.
Summarization: Summarization is the ability of generative models to analyze large texts and produce concise, condensed versions that accurately convey the core meaning and key points.
Supervised Learning: A type of machine learning in which a model is trained on labeled data to make predictions or classifications about new, unseen data.
Example: A supervised learning algorithm that can classify images of handwritten digits based on labeled training data.
Symbolic Methodology (Symbolic AI): A symbolic methodology is an approach to developing AI systems for NLP based on a deterministic, conditional approach. In other words, a symbolic approach designs a system using very specific, narrow instructions that guarantee the recognition of a linguistic pattern. Rule-based solutions tend to have a high degree of precision, though they may require more work than ML-based solutions to cover the entire scope of a problem, depending on the application.
Synthetic Data: Artificially-generated data created to train machine learning models, often used when real data is limited.
Synthetic Media: Any type of AI-created media content including images, videos, text, and audio, useful in creative and communications-related industries.
Synthetic Minority Over-sampling Technique (SMOTE): A technique for generating synthetic samples to address a class imbalance in datasets.
T
T-SNE (T-Distributed Stochastic Neighbor Embedding): A dimensionality reduction technique visualizing high-dimensional data in lower-dimensional space.
T5 (Text-to-Text Transfer Transformer): A transformer model trained to perform various text-to-text tasks, exemplifying versatile text processing.
Tagging: See Parts-of-Speech Tagging (aka POS Tagging).
Taxonomy: A taxonomy is a predetermined group of classes of a subset of knowledge (e.g., animals, drugs, etc.). It includes dependencies between elements in a “part of” or “type of” relationship, giving itself a multi-level, tree-like structure made of branches (the final node or element of every branch is known as a leaf). This creates order and hierarchy among knowledge subsets.
Companies use taxonomies to more concisely organize their documents which, in turn, enables internal or external users to more easily search for and locate the documents they need. They can be specific to a single company or become de-facto languages shared by companies across specific industries.
TD3 (Twin Delayed DDPG): An advanced AI technique that improves deep reinforcement learning by integrating several cutting-edge strategies including policy gradient, actor-critics, and enhanced deep Q-learning.
Temperature: A parameter that controls the degree of randomness or unpredictability of the LLM output. A higher value means greater deviation from the input; a lower value means the output is more deterministic.
Test Set: A test set is a collection of sample documents representative of the challenges and types of content an ML solution will face once in production. A test set is used to measure the accuracy of an ML system after it has gone through a round of training.
Text Analytics: Techniques used to process large volumes of unstructured text (or text that does not have a predefined, structured format) to derive insights, patterns, and understanding; the process can include determining and classifying the subjects of texts, summarizing texts, extracting key entities from texts, and identifying the tone or sentiment of texts.
Text Classification: The categorization of text into predefined classes or categories, crucial for content organization and sentiment analysis.
Text Generation: A process where an agent produces new text-based content, applicable in chatbots, creative writing, and more.
Text Summarization: A range of techniques that automatically produce short textual summaries from lengthy text documents, aiding information extraction. The principal purpose of this technology is to reduce employee time and effort required to acquire insight from content, either by signaling the value of reading the source(s), or by delivering value directly in the form of the summary.
Text-to-Image Generation: The creation of images based on textual descriptions.
Text-to-speech: Text-to-speech (TTS) is a technology that converts written text into spoken voice output. It allows users to hear written content being read aloud, typically using synthesized speech.
Text-to-Speech Conversion: The transformation of written text into spoken language, powering applications like voice assistants. Sometimes called “read aloud” technology.
Thesauri: Language or terminological resource “dictionary” describing relationships between lexical words and phrases in a formalized form of natural language(s), enabling the use of descriptions and relationships in text processing.
Tokenization: The process of breaking text into individual words or subwords (tokens) to input them into a language model.
Example: Tokenizing a sentence “I am ChatGPT” into the words: “I,” “am,” “Chat,” “G,” and “PT.”
Tokens: A unit of content corresponding to a subset of a word (often, the individual words used to compose a sentence). Tokens are processed internally by LLMs and can also be used as metrics for usage and billing. Breaking down text into these units allows AI models to process and analyze language at a granular level, enabling tasks like language generation.
Training Data: Training data refers to the set of examples or input data fed to an ML algorithm and used to train a generative AI model. This data consists of a collection of representative samples the model learns from to generate new content or make predictions.
Training Set: A training set is the pre-tagged sample data fed to an ML algorithm for it to learn about a problem, find patterns, and ultimately, produce a model that can recognize those same patterns in future analyses.
Transfer Learning: A technique in which a pre-trained model is used as a starting point for a new ML task. This strategy applies a pre-trained model to problems it’s unfamiliar with, allowing it to generalize knowledge it’s gained from a previous task.
Transformer: A type of neural network architecture that uses self-attention mechanisms to process sequential data, enabling AI models to interpret context and generate coherent output. It can be pre-trained on large datasets and then fine-tuned for specific tasks using transfer learning, reducing the need for extensive task-specific data.
Transformer Model: A type of neural network that uses self-attention to learn context and tracks relationships in sequential data.
TRPO (Trust Region Policy Optimization): A reinforcement learning algorithm that optimizes policies while respecting policy constraints.
Trustworthy AI: An AI system developed and used ethically, transparently, and with consideration for societal impact.
t-SNE (T-Distributed Stochastic Neighbor Embedding): A dimensionality reduction technique visualizing high-dimensional data in lower-dimensional space.
Tunable: An AI model that can be easily configured for specific requirements. For example, by industry such as healthcare, oil and gas, departmental accounting or human resources.
Tuning (Model Tuning or Fine-Tuning): The procedure of re-training a pre-trained language model using your own custom data. The weights of the original model are updated to account for the characteristics of the domain data and the task you are interested modeling. The customization generates the most accurate outcomes and best insights.
Turing test: Named after famed mathematician and computer scientist Alan Turing to evaluate a machine’s ability to exhibit intelligence equal to humans, especially in language and behavior. When facilitating the test, a human evaluator judges conversations between a human and machine. If the evaluator cannot distinguish between responses, then the machine passes the Turing test.
U
Uncertainty Estimation: The quantification of prediction uncertainty in neural network models. This is helpful because deep learning networks tend to make overly confident predictions, and wrong answers made with high confidence can cause serious problems in the real world. Also known as uncertainty quantification.
Underfitting: A problem that happens when a model is overly simple, causing it to fail to capture underlying patterns in the data.
Uniform Manifold Approximation and Projection (UMAP): A dimensionality reduction technique using Riemannian geometry and algebraic topology to enable visualization of complex data in lower dimensions.
Unsupervised Image-to-Image Translation: The transformation of images between domains without paired data, fostering creative image manipulation.
Unsupervised Learning: A type of machine learning where models are trained on unlabeled and uncategorized data, allowing them to discover hidden patterns and features without explicit supervision or target labels.
Unsupervised RL: A reinforcement learning approach without explicit external rewards, often relying on intrinsic motivation.
Unstructured Data: Unstructured data is any information that isn’t arranged in a pre-defined model or structure, making it tough to collect, process, and analyze. Lacking rigid constructs, unstructured data are often more representative of “real world” business information (examples – Web pages, images, videos, documents, audio).
V
VAE Disentanglement: When a variational autoencoder learns interpretable and independent features in data representation.
Value Iteration Networks: Neural networks employed for value iteration in reinforcement learning tasks.
Variational Autoencoder (VAEs): An autoencoder model that is trained to minimize reconstruction errors between the decoded data and the original data.
Vision Transformer (ViT): A transformer-like model that applies self-attention to image classification tasks.
Voice Processing: Voice processing in AI refers to the pipeline of speech-to-text conversion followed by text-to-speech synthesis.
Voice Synthesis: The AI-driven generation of spoken language, pivotal for applications like text-to-speech conversion.
Voxel-Based Model: A model that represents 3D space using small cubes called voxels, which are essentially three-dimensional pixels. This is vital for 3D scene understanding.
W
Wasserstein GAN: A GAN variant built for improved stability during training.
Weak AI (or narrow AI): AI that’s focused on a particular task and can’t learn beyond its skill set. These narrow systems excel at specific tasks within limited context, but lack generalized intelligence and adaptability outside their domain. Most of today’s AI is weak AI.
Weak-to-Strong Generalization: Weak-to-strong generalization is an AI training approach that uses less capable models to guide and constrain more powerful ones towards better generalization beyond their narrow training data.
Whisper: OpenAI’s Whisper is an AI system developed to perform automatic speech recognition (ASR), the task of transcribing spoken language into text.
Word Embedding: The numerical representation of a word in natural language processing, used to capture semantic relationships between words.
Word2Vec: A method for learning word embeddings from a large text corpus, facilitating semantic understanding.
World Model: AI models that simulate real-world dynamics to make predictions and decisions, vital in fields like robotics, weather forecasting, and computer vision.
X
Y
Z
Zero-Shot Learning: A technique that allows machine learning models to recognize and classify new, unfamiliar objects or concepts without receiving specific training or labeled examples for those classes, demonstrating the model’s ability to transfer knowledge autonomously.
Zero-to-One Problem: The zero-to-one problem refers to the difficulty of finding an initial solution when addressing complex challenges, which is often disproportionately challenging compared to subsequent progress.