Ultimate Artificial 


Intelligence Glossary
2023 Edition
Introduction

For over 4 years now, we have

published our “Ultimate Artificial

Intelligence Glossary.” Each year

we expand upon the terms to


help continue to advance the

knowledge of AI.

2018 2019 2020

As more and more companies

explore AI to improve business

processes, our glossary helps

navigate commonly used terms to

help accelerate AI readiness.

Whether you’re just learning about

the nuances of Al or want a

2021 2022

refresher, this guide is for you.

 Activation Function Anchor Box

A In artificial neural networks, the The archetypal location, size, and shape

activation function of a node defines the for finding bounding boxes in an object

output of that node given an input or set detection problem. For example, square

A/B Testing of inputs. anchor boxes are typically used in face

detection models.
A statistical way of comparing two (or

more) techniques, typically an incumbent Active Learning

against a new rival. It aims to determine Annotation
A machine learning term that refers to
which technique performs better, and
various methods for actively improving the A markup placed on an image (bounding
whether the difference is statistically
performance of trained models. boxes for object detection, polygons or a
significant.
segmentation map for segmentation) to

teach the model the ground truth.

Adversarial Example

Specialised inputs created with the

Annotation Format
purpose of confusing a neural network,

resulting in the misclassification of a given Particular way of encoding an annotation.

input. These notorious inputs are There are many ways to describe a

indistinguishable to the human eye, but bounding box's size and position (JSON,

cause the network to fail to identify the XML, TXT, etc) and to delineate which

contents of the image. annotation goes with which image.

Adversarial Machine Learning Annotation Group

A research field that lies at the intersection Describes what types of objects you are

of machine learning (ML) and computer identifying. For example, "chess pieces" or

security. It enables the safe adoption of ML "vehicles."

techniques in adversarial settings like

spam filtering, malware detection, etc.

Application Programming

Interface (API)
AI Algorithms
A set of commands, functions, protocols,
Accuracy Extended subset of machine learning that
and objects that programmers can use to
Refers to the percentage of correct tells the computer how to learn to operate
create software or interact with an external
predictions the classifier made. on its own through a set of rules or
system.
instructions.
Architecture What is Autonomous AI?
A specific neural network layout (layers,
The most advanced form of AI is autonomous artificial intelligence, in which
neurons, blocks, etc). These often come in
processes are automated to generate the intelligence that allows machines, bots
multiple sizes whose design is similar
and systems to act on their own, independent of human intervention. It is often
except for the number of parameters.
used in autonomous vehicles.

Artificial Intelligence This field of AI is still very new, and researchers are continually refining their

algorithms and their approaches to the problem, but it entails multiple layers.

A computational system that simulates

parts of human intelligence but focuses

on one narrow task. Sensing Planning

Building a model of the constantly Finding the best path forward re quires
Artificial Neural Network shifting world requires a collection of studying the model and also importing

sensors that are usually cameras and information from other sources like
A learning model created to act like a
often controlled lighting from lasers or mapping software, weather forecasts,
human brain that solves tasks that are too
other sources. The sensors usually also traffic sensors and more.
difficult for traditional computer systems


to solve.
include position information from GPS
or some other independent mechanism.
Control
Audio Speech Recognition (ASR) After a path is chosen, any device
Fusion
must ensure that the motors and
A technology that processes human
The details from the various sensors steering work to move along the path
speech into readable text.
must be organized into a single, without being diverted by bumps or

coherent view of what’s happening small obstacles.

AutoML around the autonomous vehicle. The

sensor fusion algorithms must sort In general, information flows from the
Automates each step of the ML workflow
through the details and construct a top layer of the sensors down to the
so that it’s easier for users with minimal
reliable model that can be used in later control layer as decisions are made.
effort and machine learning expertise.
stages for planning. There are feedback loops, though,

that bring information from the lower

Automation Bias layers back up to the top to improve
Perception
sensing, planning and perception.
When a human decision maker favors
After the model is constructed, the
recommendations made by an automated
system must begin to identify
decision-making system over information
important areas like any roads or
made without automation, even when the
paths or moving objects.
automated decision-making system


makes errors.
 Batch Size Boosting
B The number of training examples utilized A machine learning technique that

in one iteration. iteratively combines a set of simple and

not very accurate classifiers (referred to as

Backward Chaining "weak" classifiers) into a classifier with

Bayes's Theorem
high accuracy (a "strong" classifier) by
A method where the model starts with the
A famous theorem used by statisticians upweighting the examples that the model
desired output and works in reverse to find
to describe the probability of an event is currently misclassifying.
data that might support it.
based on prior knowledge of conditions

that might be related to an occurrence.

Bootstrapping
Base Workflow
Bootstrapping is any test or metric that
One of Clarifai's prebuilt models that can Bias
uses random sampling with replacement
be built upon to create a custom model. It
When an Al algorithm produces results that and falls under the broader class of
pre-indexes inputs for search and provides
are systemically prejudiced due to erroneous resampling methods. Bootstrapping
a default embedding space.
assumptions in the machine learning process. assigns measures of accuracy (bias,

variance, confidence intervals, prediction

Baseline error, etc.) to sample estimates.

Big Data
A model used as a reference point for
Big data refers to data that is so large, fast
comparing how well another model
or complex that it's difficult or impossible
Bounding Box
(typically, a more complex one) is
to process using traditional methods. In an image, the (x, y) coordinates of a
performing. Baseline models help
rectangle around an area of interest.
developers quantify the minimal expected

performance that a new model must Binary Classification

achieve to be useful. 01
kid.

The task of classifying elements of a set

into two groups on the basis of a

Batch classification rule i.e. a model that evaluates

email messages and outputs either "spam"

The set of examples used in one

or "not spam" is a binary classifier.
iteration (that is, one gradient update)


of model training.
Black Box AI

Batch Inference An Al system whose inputs and operations Brute Force Search
are not visible to the user. A black box, in a A search that isn't limited by clustering/
Asynchronous process that is executing
general sense, is an impenetrable system. approximations; it searches across all
predictions based on existing models and
inputs. Often more time-consuming and
observations, and then stores the output.
expensive, but more thorough.
C Checkpoint
Data that captures the state of the
variables of a model at a particular time.
Cluster
A group of observations that show
similarities to each other and are organized
Checkpoints enable exporting model by similarities.
Calibration Layer weights, performing training across
A post-prediction adjustment, typically to multiple sessions and continuing training
account for prediction bias. The adjusted past errors. Clustering
predictions and probabilities should match A method of unsupervised learning and
the distribution of an observed set
common statistical data analysis
of labels.
Class technique. In this method, observations
One of a set of enumerated target values that show similarities to each other are
for a label. For example, in a binary organized into groups (clusters).
Chatbot classification model that detects spam, the
Simulates human conversation, using two classes are spam and not spam. In a
response workflows or artificial intelligence multi-class classification model that
to interact with people based on verbal identifies dog breeds, the classes would
and written cues. Chatbots have become be poodle, beagle, pug, etc.
increasingly sophisticated in recent years
and in the future may be indistinguishable Class Balance
from humans.
The relative distribution between the
number of examples of each class used to
train a model. A model performs better if
there are a relatively even number of
examples for each class. Cognitive Computing
A computerized model that mimics the way
Classification the human brain thinks. It involves self­
Process of grouping and categorizing learning through the use of data mining,
objects and ideas recognized, natural language processing, and pattern
differentiated, and understood in data. recognition.

Classifier
An algorithm that implements classification.
It refers to the mathematical function
implemented by a classification algorithm
that maps input data to a category.
Computer Vision Container Curse of Dimensionality
Field of AI that trains computers to A virtualized environment that packages its The curse of dimensionality refers to
interpret and understand the visual world. dependencies together into a portable various phenomena that arise when
Using digital images from cameras and environment. Docker is one common way analyzing and organizing data in high-
videos and deep learning models, to create containers. dimensional spaces that do not occur in
machines can accurately identify and low-dimensional settings such as the
classify objects — and then react to what three-dimensional physical space of
they “see.” Convolutional Filter everyday experience.
A convolution is a type of block that
helps a model learn information about
Concept relationships between nearby pixels. Custom Dataset
Describes an input, similar to a "tag" or A set of images and annotations pertaining
"keyword." There are two types: those that to a domain specific problem. In contrast
you specify to train a model, and those Convolutional Neural Network to a research benchmark dataset like coco
that a model assigns as a prediction. Convolutional neural networks are deep or Pascal voe.
artificial neural networks that are used
primarily to classify images (e.g. name
Confidence what they see), cluster them by similarity Custom Training
A model is inherently statistical. Along with (photo search), and perform object The process of teaching a model to make
its prediction, it also outputs a confidence recognition within scenes. certain predictions.
value that quantifies how "sure" it is that
its prediction is correct.
CoreML
A proprietary format used to encode
Confidence Threshold weights for Apple devices that takes
We often discard predictions that fall advantage of the hardware accelerated
below a certain bar. This bar is the neural engine present on iPhone and
confidence threshold. iPad devices.

Confusion Matrix CreateML

A confusion matrix is a table that is often A no-code training tool created by Apple
used to describe the performance of a that will train machine learning models and
classification model on a set of test data export to CoreML. It supports classification
for which the true values are known. The and object detection along with several
confusion matrix itself is relatively simple types of non computer-vision models

to understand, but the related terminology (such as sound, activity, and text
can be… confusing. classification).
 Dataset Deep Neural Network
D A collection of data and a ground truth of An artificial neural network (A NN) with
outputs that you use to train machine multiple layers between the input and

learning models by example. output layers. It uses sophisticated

Data mathematical modeling to process data

In the data science and AI world, any in complex ways.

De-duplication
collection of information in a digital form.

It’s important to distinguish between The removal of identical data, or data

Deploy
structured and unstructured data; that is so similar that for all intents and

structured data is highly specific and is purposes it can be considered duplicate Taking the results of a trained model and

stored in a predefined format such as an data. Using visual search, a similarity using them to do inference on real world

spreadsheet table, whereas unstructured threshold can be set to decide what data. This could mean hosting a model on

data is a conglomeration of many varied should be removed. a server or installing it to an edge device.

types of data that are stored in their native

formats, such as images, video, audio, and

Detection Mode
text. Data is also a plural, with the singular

being “datum”.
j
Also known as ob ect detection. A model that

identifies the presence, location and type of

j
ob ects within images or video frames.
Data Annotation

The process of labeling datasets to be

Diversity, Equity & Inclusion (DEI)
used as inputs for machine learning
Term used to describe policies and
models.
programs that promote the representation

and participation of different groups of

Data Curation individuals, including people of different

The process of collecting, organizing, ages, races and ethnicities, abilities and

cleaning, labeling, and maintaining data for disabilities, genders, religions, cultures,

use in training and testing models. and sexual orientations.

Deep Learning

Data Mining
The general term for machine learning Domain Adaptation
using layered (or deep) algorithms to
The process by which patterns are A technique to improve the performance of a
learn patterns in data. It is most often
discovered within large sets of data with model where there is little data in the target
used for supervised learning problems.
the goal of extracting useful information domain by using knowledge learned by

from it. another model in a related domain. An example

could be training a model to recognize taxis

using a model that recognizes cars.

E Embedding Space
The d-dimensional vector space that
features from a higher-dimensional vector
F
space are mapped to. Ideally, the
Edge AI embedding space contains a structure that F Score
Data is processed on the same device that yields meaningful mathematical results. A weighted average of the true positive
produces it, or at most on a nearby rate of recall and precision.
computer with no reliance on distant cloud Emotional AI
servers or other remote computing nodes. Facial Recognition
AI can work faster, and respond more Technologies that use affective computing
accurately to time-sensitive events. and artificial intelligence techniques to An application capability of identifying or
sense, learn about and interact with human
 verifying a person from an image or a
emotional life. video frame by comparing selected facial
features from the image and a face
Ensemble Models database.
Machine learning approach to combine
multiple other models in the prediction False Positives
process. While the individual models may An error where a model falsely predicts the
not perform very well, when combined presence of the desired outcome in an
they can be very powerful indeed. input, when in reality it is not present
(Actual No, Predicted Yes).
Extensible Markup Language
Edge Computing (XML) False Negatives
A distributed computing framework that A language and file format for storing, An error where a model falsely predicts an
brings enterprise applications closer to transmitting, and reconstructing arbitrary input as not having a desired outcome,
data sources such as loT devices or local data. It defines a set of rules for encoding when one is actually present. (Actual Yes,
edge servers. documents in a format that is both 
 Predicted No).
human-readable andmachine-readable.
Embeddings FastAI
A categorical feature represented as a A library built on top of PyTorch for rapid
continuous-valued feature. Typically,
 prototyping and experimentation. There is
an embedding is a translation of a 
 a companion course that teaches the
high­-dimensional vector into a low- fundamentals of machine learning.
dimensional space.
Feature Extraction
When image features at various levels of
complexity are extracted from the image
G Ground Truth
The "answer key" for your dataset. This is
how you judge how well your model is
data. Typical examples of such features doing and calculate the loss function we
are: Lines, edges, and ridges. Localized Generalization use for gradient descent. It's also what we
interest points such as corners, blobs, or Refers to a model's ability to make correct use to calculate our metrics. Having a good
points.More complex features may be predictions on new, previously unseen data ground truth is extremely important. Your
related to texture, shape, or motion.
as opposed to the data used to train the model will learn to predict based on the
The process by which data that is too large model. ground truth you give it to replicate.
to be processed is transformed into a
reduced representation set of features
such as texture, shape, lines, and edges. Generative Adversarial

Networks (GANs) H
A class of artificial intelligence algorithms
used in unsupervised machine learning, Hashing
implemented by a system of two neural
networks contesting with each other in a In machine learning, a mechanism for
zero-sum game framework. This technique bucketing categorical data, particularly
can generate photographs that look at when the number of categories is large,
least superficially authentic to human but the number of categories actually
observers, having many realistic appearing in the dataset is

characteristics (though in tests people can comparatively small.
tell real from generated in many cases).
Folksonomy Hidden Layer
User-generated system of classifying and
Generative AI A synthetic layer in a neural network
Models that can be trained using existing between the input layer (that is, the
organizing online content into different features) and the output layer (the
categories by the use of metadata such as content like text, audio files, or images to
create new original content. prediction). Hidden layers typically contain
electronic tags. an activation function (such as ReLU) for
training. A deep neural network contains
Framework Grid Search more than one hidden layer.
Deep learning frameworks implement Grid search is a tuning technique that
neural network concepts. Some are attempts to compute the optimal values of
designed for training and inference— hyperparameters for training models by
TensorFlow, PyTorch, FastAI, etc. And performing an exhaustive search through a
others are designed particularly for speedy subset of hyperparameters.
inference—OpenVino, TensorRT, etc.
Holdout Data Hyperparameter Image Segmentation
Examples intentionally not used during The levers by which you can tune your The process of dividing a digital image

training. The validation dataset and test model during training. These include
 into multiple segments with the goal of

dataset are examples of holdout data. It things like learning rate and batch size. simplifying the representation of an

helps evaluate your model's ability to You can experiment with changing image into something that is easier to

generalize to data other than the data on hyperparameters to see which ones analyze. Segmentation divides whole

which it was trained. perform best with a given model for
 images into pixel groupings, which can

your dataset. then be labeled and classified.

Hosted Model
Information Retrieval
I
A set of trained weights located in the

cloud that you can receive predictions The area of Computer Science studying

from via an API. the process of searching for information

in a document, searching for documents

Inference themselves, and also searching for
Human Workforce (“Labelers”)
Making predictions using the weights you metadata that describes data and for
Workers who can help to complete work on save after training your model. databases of texts, images or sounds.
an as-needed basis, which for purposes

usually means labeling data (images).

ImageNet Input

A large visual database designed for
 Any information or data sent to a computer

use in visual object recognition
 for processing is considered input.

software research.

Input Layer
Image Recognition The first layer (the one that receives the

The ability of software to identify
 input data) in a neural network.

objects, places, people, writing and


actions in images.
Intelligent Character


Recognition (ICR)
Implicit Bias
Related technology to OCR designed to
Automatically making an association or
recognize handwritten characters.
assumption based on one's mental models

and memories. Implicit bias can affect how

data is collected and classified, and how

machine learning systems are designed

and developed.
J L M
Jetson Label Machine Intelligence

An edge computing device created by Assigning a class or category to a specific An umbrella term that encompasses

NVIDIA that includes an onboard GPU. object in your dataset. machine learning, deep learning and

classical learning algorithms.

JSON Labeling
Machine Learning
A freeform data serialization format Also known as data labeling; the

originally created as part of JavaScript process of annotating datasets to train A general term for algorithms that can

but now used much more broadly. Many machine learning models. learn patterns from existing data and use

annotation formats use JSON to encode these patterns to make predictions or

their bounding boxes. decisions with new data.

Labeling Criteria

A labeling requirements guide which

Jupyter Notebook Masked Language Model
includes instructions for the labeling

A common data science tool that process itself as well as written definitions A language model that predicts the

enables you to execute Python code and a multitude of visual examples for probability of candidate tokens to fill in

visually. Each " ecute by hitting each concept. blanks in a sequence.

"Ctrl+Enter". The results of the

execution are displayed below the cell.

Labeler Metadata

AI-automated tool using end-to-end Information about an analog or digital

K
workflows to label images and video at object, a component of an object, or a

scale to create high-quality training data. coherent collection of objects. Metadata

describing digital content is often

structured (e.g., with tagging or markup).

Knowledge Graph

Collection of nodes and edges where the

nodes represent concepts, entities, Misclassification Rate

relationships, and events, and the edges Rate used to gauge how often a model’s
represent the connections between them. predictions are wrong.
MLOps Types of Models
Also known as Machine Learning
It’s important to understand the different types of models for different data types.

Operations. Best practices for

organizations to operationalize machine
learning. Often involves collaboration
between data scientists and devops Audio Transcription Model Domain Model
professionals to manage production ML. Takes audio containing speech and Focuses on understanding a single
converts it into text. The text files domain, such as travel, weddings,
Modality allow audio to be searched for key food, not-safe-for-work (NSFW), etc.
terms, or AI models to transmit text
A high-level data category. For example,
instead of audio over networks, which
numbers, text, images, video and audio are Embedding Model
is much smaller and faster.
five different modalities. Computers and models can’t
understand images and text like
Model Classification Model humans do. Embedding models take
Reads an input such as text, image, unstructured input like images, audio,
The representation of what a machine
audio, or video data and generates an text, and video and transform them
learning system has learned from the
output that classifies it into a into a series of numbers called
training data.
category. For example, a language vectors which can then be input into
classification model might read a the prediction models.
ModelOps sentence and determine whether it's
in French, Spanish, or Italian.
As defined by Gartner, ModelOps is focused Masked Language Model
primarily on the governance and life cycle
A language model that predicts the
management of a wide range of Detection Model probability of which words make the
operationalized artificial intelligence (AI) and
Detection comprises two tasks; listing most sense to fill in blanks in a
decision models, including machine learning,
“what” things appear in an image, and sequence. A simple example could be
knowledge graphs, rules, optimization,
“where” they appear. Results are “Good _____, how are you?” where
linguistic and agent-based models.
returned as bounding boxes along probable candidate words could be
with the names of the detected items. “morning”, “day”, or “evening”.
Model Size
The number of parameters (or neurons)
 More model types on the next page...
a model has. This can also be measured

in terms of the size of the weights file

on disk.
Types of Models (Continued) Model Training

The process of determining the best

model.
Multimodal Model Named Entity Recognition

A model whose inputs and/or outputs Model

include more than one modality. For
A method that is used for recognizing
example, consider a model that takes
entities such as people, dates,
both an image and a text caption (two
organizations, and locations that are
modalities) as features, and outputs a
present in a text document.
score indicating how appropriate the

text caption is for the image.

Pre-trained Model

A model or the component of a model,

that has been preliminary trained,

Monte Carlo Simulation

generally using another data set. (for Used to model the probability of different

example, finding lines, corners, and outcomes in a process that cannot easily

patterns of colors). Pre-training on a be predicted due to the intervention of

large dataset like the huge Common random variables. It’s a technique used to

Objects in Context (COCO), which has understand the impact of risk and

330,000 images with 1.5 million uncertainty. It was developed while

objects to detect, can reduce the working on nuclear weapons in the 1940s,

number of custom images you need to and was given the code name “Monte

obtain satisfactory results. Carlo” in reference to the Monte Carlo

Casino in Monaco, where one of the

inventor’s uncles would borrow money

Segmentation Model
from relatives to gamble.
Predictive Model Instead of bounding boxes returned

A model that uses observations for each concept, this model indicates

via a heat map and trace (think of a

Multi-class Classification
measured in a sample to gauge the

probability that a different sample or coloring book) of regions for each Classification problems that distinguish

remainder of the population will
 concept. between more than two classes. For

exhibit the same behavior or have
 example, there are approximately 53 species

the same outcome. of maple trees, so a model that categorized

maple tree species would be multi-class.

N
Neuron
A unit in an Artificial Neural Network

processing multiple input values to

O
generate a single output value.
Natural Language Processing Object Detection

(NLP) Neural Network A computer technology related to

computer vision and image processing that

A branch of Al that helps computers Series of algorithms that endeavors to
deals with detecting instances of semantic
understand, interpret, and manipulate

human language. This field of study

recognize underlying relationships in a set
objects of a certain class (such as humans,
focuses on helping machines understand
of data through a process that mimics the
buildings, or cars ) in digital images and
way the human brain operates.
videos. This technique also involves
human language in order to improve
localizing the object in question, which
human-computer interfaces.
Noise differentiates it from classification, which

only tells the type of object.

Signals with no causal relation to the target
Natural Language Understanding
function.
Determining a user's intentions based on

what the user typed or said. For example, a Object Recognition

Normalization
search engine uses natural language Also known object classification. A
understanding to determine what the user The process of converting an actual range computer vision technique for identifying
is searching for based on what the user of values into a standard range of values, objects in images or videos.
typed or said. typically -1 to +1 or O to 1.

Neuro-Symbolic Object Tracking

The process of following a specific object

The combining of neural and symbolic AI
of interest, or multiple objects, in a given
architectures to address complementary
scene. It traditionally has applications in
strengths and weaknesses of each,
video and real-world interactions where
providing a robust AI capable of reasoning,
observations are made following an initial
learning, and cognitive modeling.
object detection.

Neural Architecture Search

Automatically trying many variations of

model layouts and hyperparameters to find

the optimal configuration.

On-premise Software Optimization

Software that is installed and runs on The selection of the best element (with P
computers located on the premises of the regard to some criterion) from some set of

organization using that software versus at available alternatives.

a remote facility such as a server farm or Parameter

on the cloud. Any characteristic that can be used to help
Output
define or classify a system. In Al, they are
Predictions made after the input uploaded used to clarify exactly what an algorithm
One Shot Classification
to or fed into a model are processed by
 should be seeking to identify as important
A model that only requires that you have the model. data when performing its target function.
one training example of each class you

want to predict on. The model is still

Outsourced Labeling Pattern Recognition
trained on several instances, but they only

have to be in a similar domain as your Paying people to annotate, or label, your A branch of machine learning that focuses
training example. data. Its effectiveness can depend on the on the recognition of patterns and
domain expertise of annotators. Providing regularities in data, although it is in some
a comprehensive labeling criteria is crucial cases considered to be nearly synonymous
Open Neural Network
for training annotators before beginning
 with machine learning.
Exchange (ONNX) a project.

ONNX is an open format to represent

Pipeline
machine learning models.
The process of going from raw images to

prediction. Usually this encompasses

OpenAI collecting images, annotation, data

Mission is to ensure that artificial general inspection and quality assurance,

intelligence benefits all humanity. transformation, preprocessing and

augmentation, training, evaluation,

deployment, inference (and then repeating

Optical Character
 the cycle to improve the predictions).

Recognition (OCR)
Overfitting
A computer system that takes images of Precision
typed, handwritten, or printed text and A machine learning problem where an
Indicator of a machine learning model's
converts them into machine-readable text. algorithm is unable to discern information
performance – the quality of a positive
that is relevant to its assigned task from
prediction made by the model. Refers to
information which is irrelevant within
the number of true positives divided by the
training data. Overfitting inhibits the
total number of positive predictions.
algorithm's predictive performance when

dealing with new data.

Polygon
A (usually non-rectangular) region defining
an object with more detail than a
Precision (Recognition)
A rate that measures how often a model is
correct when it predicts “yes.”
R
rectangular bounding box. Polygon
annotations can be used to train Recall (Sensitivity)
segmentation models or to enhance
Prediction
The fraction of relevant instances that
performance of object-detection models An attempt by a model to replicate the have been retrieved over the total amount
by enabling a more accurate bounding box ground truth. A prediction usually contains of relevant instances.
to be maintained after augmentation. a confidence value for each class.
Receiver Operating
Prevalence
Characteristic (ROC) Curve
The rate of how often the “yes” condition
actually occurs in a sample. This is a commonly used graph that
summarizes the performance of a
classifiers over all possible thresholds. It is
Production generated by plotting the True Positive
The deployment environment where
 Rate (y-axis) against the False Positive
the model will run in the wild on
 Rate (x-axis) as you vary the threshold for
real-world images (as opposed to the assigning observations to a given class.
testing environment where the model

is developed). Recurrent Neural Network

A type of artificial network with loops in
Pruning them, allowing recorded information, like
Positive Predictive Value (PPV) data and outcomes, to persist by being
The use of a search algorithm to cut off
Very similar to precision, except that it undesirable solutions to a problem in an Al passed from one step of the network to
takes prevalence into account. In the case system. It reduces the number of decisions the next. They can be thought of as
where the classes are perfectly balanced that can be made by the Al system. multiple copies of the same network with
(meaning the prevalence is 50%), the each passing information to its successor.
positive predictive value is equivalent

to precision. PyTorch
Regression
A popular open source deep learning
framework developed by Facebook.
 A statistical measure used to determine
Prediction the strength of the relationships between
It focuses on accelerating the path

An attempt by a model to replicate the from research prototyping to
 dependent and independent variables.
ground truth. A prediction usually contains production deployment.
a confidence value for each class.
Reinforcement Learning Specificity

A type of machine learning in which S The rate of how often a model predicts

machines are "taught" to achieve their “no,” when it’s actually “no.”

target function through a process of

experimentation and reward receiving Search Query

Standard Classification
positive reinforcement when its processes A query that a user feeds into a search
produce the desired result and negative engine to satisfy his or her information The process by which an input is assigned

reinforcement when they do not. This is needs. If the query itself is a piece of visual to one of a fixed set of categories. In

differentiated from supervised learning, content then that is what is known as a machine learning, this is often achieved by

which would require an annotation for every "visual search query." learning a function that maps an input to a

individual action the algorithmwould take. score for each potential category.

Selective Filtering
Strong AI
When a model ignores "noise" to focus on

valuable information. A theoretical form of AI that replicates

human functions, such as reasoning,

planning, and problem-solving.

Siamese Networks

A different way of classifying image where

Structured Data
ReLU instead of training one model to learn to

classify image inputs it trains two neural Data that resides in a fixed field within a
In the context of artificial neural networks,
networks that learn simultaneously to find file or record. Structured data is typically
the ReLU (rectified linear unit) activation
similarity between images. stored in a relational database. It can
function is an activation function which
consist of numbers and text, and sourcing
outputs the same as its input if the input is
can happen automatically or manually, as
positive, and zero if the input is negative. A
Signal long as it's within an RDBMS structure.
related function is the leaky rectified linear
Inputs, information, data.
unit (leaky rectified linear unit) which

assigns a small positive slope for x < 0. Supervised Learning

Software Development Kit (SDK)
A machine learning approach that's

Responsible AI A set of software development tools that defined by its use of labeled datasets.

allows for the creation of applications on a These datasets are designed to train or
Umbrella term for aspects of making
specific platform. "supervise" algorithms into classifying data
appropriate business and ethical choices
or predicting outcomes accurately. Using
when adopting AI, including business and
labeled inputs and outputs, the model can
societal value, risk, trust, transparency,
measure its accuracy and learn over time.
fairness, bias mitigation, explainability,

accountability, safety, privacy, and

regulatory compliance.
Symbiotic Intelligence Train

A combination of human and artificial T j

The process iteratively ad usts your

intelligence. Instead of relying on memory, model’s parameters to converge o n


or having to open a book, or visit a the weights that optimally mimic the


website, an enhanced human could have Target Function training data.

access to all of the information that is The end goal of an algorithm.

stored on the internet, and an advanced AI
Training Datset
could feed the relevant data points to the

human brain, enabling the human to be

Taxonomy An initial dataset used to train machine

fully in control. In essence, a taxonomy is a model’s learning algorithms. Models create and
worldview, or the framework for how your '
refine their rules using this data. It s a set

model sees its training data. In practice, it’s of data samples used to fit the parameters
Synthetic Intelligence
a list of visually-distinct model concepts of a machine learning model to training it

Synthetic intelligence (SI) is an alternative and the definitions of those concepts. by example.

term for artificial intelligence emphasizing

that the intelligence of machines need not

Temporal Data Transfer Learning
be an imitation or in any way artificial; it

can be a genuine form of intelligence. An Data recorded at different points in time. Transferring information from one machine

analogy can be made with simulated learning task to another. It might involve

diamonds (such as cubic zirconia) versus transferring knowledge from the solution
TensorFlow of a simpler task to a more complex one, or
synthetic diamonds (real diamonds made

of carbon created by humans). An open-source software library also used involve transferring knowledge from a task

for machine learning applications such as where there is more data to one where

neural networks. there is less data.

Synthetic Data

Images that are created rather than

Test Dataset Transformer
collected.

The sample of data used to provide an A neural network that transforms a

unbiased evaluation of a final model fit on sequence of elements (like words in a

the training dataset. sentence) into another sequence to solve

sequence-to-sequence tasks.

Torch
True Positives
A scientific computing framework with
Actual positives that are correctly
wide support for machine learning
identified as actual “Yes” or predicted
algorithms, written in C and lua.
“Yes.”
True Negatives Unsupervised Learning Variance
Actual negatives that are correctly Uses machine learning algorithms to The error due to sensitivity to fluctuations
identified as an actual “No” or predicted analyze and cluster unlabeled datasets. in the training set computed as the
“No.” These algorithms discover hidden patterns expectation of the squared deviation of a
or data groupings without the need for random variable from its mean.
human intervention. Its ability to discover
Turing Test similarities and differences in information
A test developed by Alan Turing in 1950, make it the ideal solution for exploratory Verify/Verification
used to identify true artificial intelligence.
 data analysis, cross-selling strategies, The process of verifying that labeled data
It tested a machine’s ability to exhibit customer segmentation, and image has been labeled correctly in adherence to
intelligent behavior equivalent to, or recognition. the ground truth.
indistinguishable from, that of a human.

V
Video Frame Interpolation
Is to synthesize several frames in the
middle of two adjacent frames of video.

Video Frame Interpolation can be applied.

I’m not a robot
Validate
The model is given new, previously unseen Visual Dictionary
data, and then metrics are collected on A document that defines every model
I’m not a robot how well it performs predictions on them. concept with a written definition and also

This is analogous to a human learning math a wide array of visual examples. This helps
problems using one set of questions, then establish ground truth by providing
tested to see if they learned properly with confirmation that each involved party
a different set of questions. understands the model’s worldview,


U Validation Data Set

The sample of data used to provide an
or taxonomy.

Visual Recognition
unbiased evaluation of a model fit on the
Unstructured Data training dataset while tuning model
The ability of software to identify objects,
places, people, writing, and actions in
Unstructured data is information that hyperparameters. The evaluation becomes images and videos.
either does not have a pre-defined data more biased as skill on the validation
model or is not organized in a pre-defined dataset is incorporated into the model
manner. Unstructured data may include configuration.
documents, images, video and audio.
Visual Match Width

Instead of doing a search which W The number of neurons in a particular layer

returns the items in the database in of a neural network.
sorted order, a visual match couple be

considered returning a yes/no answer Weak AI

of whether the query is close enough
Workflows
Also known as narrow Al, weak Al refers to
to any item in the database to be a non-sentient computer system that Workflows enable users to make
considered a “match.” operates within a predetermined range of predictions on a graph that combines one
skills and usually focuses on a singular or more pre-trained, custom models and
task or small set of tasks. Most Al in use fixed function model operators using a
Visual Search
today is weak Al. single API call.
The ability of software to find visually

similar content based on an image or

video query. Weight

A coefficient for a feature in a linear model, Y

or an edge in a deep network. The goal of

training a linear model is to determine the

ideal weight for each feature. If a weight is YAML

0, then its corresponding feature does not A markup language originally invented by
contribute to the model. Yahoo that is now commonly used as a

format for configuration files.

About Clarifai

Clarifai is the leading deep learning AI platform for computer vision, natural language processing, and automatic speech

recognition. We help enterprises and public sector organizations transform unstructured images, video, text, and audio data into

structured data, significantly faster and more accurately than humans would be able to do on their own. Founded in 2013 by Matt

Zeiler, Ph.D., Clarifai has been a market leader in computer vision AI since winning the top five places in image classification at

the 2013 ImageNet Challenge. Clarifai, headquartered in Wilmington, DE, is continuing to grow with more than 90 employees in

North America and Europe. For more information, please visit: www.clarifai.com.
www.clarifai.com