Machine Learning Glossary (Google)

What's inside

A/B testing

A statistical way of comparing two (or more) techniques, typically an incumbent against a new rival. A/B testing aims to determine not only which technique performs better but also to understand whether the difference is statistically significant. A/B testing usually considers only two techniques using one measurement, but it can be applied to any finite number of techniques and measures.

action

In reinforcement learning, the mechanism by which the **agent** transitions between **states** of the **environment**. The agent chooses the action by using a **policy**.

activation function

A function (for example, **ReLU** or **sigmoid**) that takes in the weighted sum of all of the inputs from the previous layer and then generates and passes an output value (typically nonlinear) to the next layer.

active learning

A **training** approach in which the algorithm chooses some of the data it learns from. Active learning is particularly valuable when **labeled examples** are scarce or expensive to obtain. Instead of blindly seeking a diverse range of labeled examples, an active learning algorithm selectively seeks the particular range of examples it needs for learning.

AdaGrad

A sophisticated gradient descent algorithm that rescales the gradients of each parameter, effectively giving each parameter an independent **learning rate**. For a full explanation, see this paper.

agent

In reinforcement learning, the entity that uses a **policy** to maximize expected **return** gained from transitioning between **states** of the **environment**.

agglomerative clustering

See **hierarchical clustering**.

AR

Abbreviation for **augmented reality**.

area under the PR curve

See **PR AUC (Area under the PR Curve)**.

area under the ROC curve

See **AUC (Area under the ROC curve)**.

artificial general intelligence

A non-human mechanism that demonstrates a broad range of problem solving, creativity, and adaptability. For example, a program demonstrating artificial general intelligence could translate text, compose symphonies, and excel at games that have not yet been invented.

artificial intelligence

A non-human program or model that can solve sophisticated tasks. For example, a program or model that translates text or a program or model that identifies diseases from radiologic images both exhibit artificial intelligence.

Formally, **machine learning** is a sub-field of artificial intelligence. However, in recent years, some organizations have begun using the terms artificial intelligence and machine learning interchangeably.

attribute

Synonym for **feature**. In fairness, attributes often refer to characteristics pertaining to individuals.

AUC (Area under the ROC Curve)

An evaluation metric that considers all possible **classification thresholds**.

The Area Under the **ROC curve** is the probability that a classifier will be more confident that a randomly chosen positive example is actually positive than that a randomly chosen negative example is positive.

augmented reality

A technology that superimposes a computer-generated image on a user's view of the real world, thus providing a composite view.

automation bias

When a human decision maker favors recommendations made by an automated decision-making system over information made without automation, even when the automated decision-making system makes errors.

average precision

A metric for summarizing the performance of a ranked sequence of results. Average precision is calculated by taking the average of the **precision** values for each relevant result (each result in the ranked list where the recall increases relative to the previous result).

See also **Area under the PR Curve**.

backpropagation

The primary algorithm for performing **gradient descent** on **neural networks**. First, the output values of each node are calculated (and cached) in a forward pass. Then, the **partial derivative** of the error with respect to each parameter is calculated in a backward pass through the graph.

bag of words

A representation of the words in a phrase or passage, irrespective of order. For example, bag of words represents the following three phrases identically:

the dog jumps

jumps the dog

dog jumps the

Each word is mapped to an index in a **sparse vector**, where the vector has an index for every word in the vocabulary. For example, the phrase the dog jumps is mapped into a feature vector with non-zero values at the three indices corresponding to the words the, dog, and jumps. The non-zero value can be any of the following:

A 1 to indicate the presence of a word.

A count of the number of times a word appears in the bag. For example, if the phrase were the maroon dog is a dog with maroon fur, then both maroon and dog would be represented as 2, while the other words would be represented as 1.

Some other value, such as the logarithm of the count of the number of times a word appears in the bag.

baseline

A **model** used as a reference point for comparing how well another model (typically, a more complex one) is performing. For example, a **logistic regression model** might serve as a good baseline for a **deep model**.

For a particular problem, the baseline helps model developers quantify the minimal expected performance that a new model must achieve for the new model to be useful.

batch

The set of examples used in one **iteration** (that is, one **gradient** update) of **model training**.

See also **batch size**.

batch normalization

**Normalizing** the input or output of the **activation functions** in a **hidden layer**. Batch normalization can provide the following benefits:

Make **neural networks** more stable by protecting against **outlier** weights.

Enable higher **learning rates**.

Reduce **overfitting**.

batch size

The number of examples in a **batch**. For example, the batch size of **SGD** is 1, while the batch size of a **mini-batch** is usually between 10 and 1000. Batch size is usually fixed during **training** and **inference**; however, **TensorFlow** does permit dynamic batch sizes.

Bayesian neural network

A probabilistic **neural network** that accounts for uncertainty in **weights** and outputs. A standard neural network regression model typically **predicts** a scalar value; for example, a model predicts a house price of 853,000. By contrast, a Bayesian neural network predicts a distribution of values; for example, a model predicts a house price of 853,000 with a standard deviation of 67,200. A Bayesian neural network relies on Bayes' Theorem to calculate uncertainties in weights and predictions. A Bayesian neural network can be useful when it is important to quantify uncertainty, such as in models related to pharmaceuticals. Bayesian neural networks can also help prevent **overfitting**.

Bellman equation

In reinforcement learning, the following identity satisfied by the optimal **Q-function**:

\[Q(s, a) = r(s, a) + \gamma \mathbb{E}_{s'|s,a} \max_{a'} Q(s', a'))\]

**Reinforcement learning** algorithms apply this identity to create **Q-learning** via the following update rule:

\[Q(s,a) \gets Q(s,a) + \alpha \left[r(s,a) + \gamma \displaystyle\max_{\substack{a_1}} Q(s’,a’) - Q(s,a) \right] \]

Beyond reinforcement learning, the Bellman equation has applications to dynamic programming. See the Wikipedia entry for Bellman Equation.

bias (ethics/fairness)

1. Stereotyping, prejudice or favoritism towards some things, people, or groups over others. These biases can affect collection and interpretation of data, the design of a system, and how users interact with a system. Forms of this type of bias include:

**automation bias**

**confirmation bias**

**experimenter’s bias**

**group attribution bias**

**implicit bias**

**in-group bias**

**out-group homogeneity bias**

2. Systematic error introduced by a sampling or reporting procedure. Forms of this type of bias include:

**coverage bias**

**non-response bias**

**participation bias**

**reporting bias**

**sampling bias**

**selection bias**

Not to be confused with the bias term in machine learning models or **prediction bias**.

bigram

An **N-gram** in which N=2.

binary classification

A type of **classification** task that outputs one of two mutually exclusive **classes**. For example, a machine learning model that evaluates email messages and outputs either "spam" or "not spam" is a **binary classifier**.

binning

See **bucketing**.

boosting

A machine learning technique that iteratively combines a set of simple and not very accurate classifiers (referred to as "weak" classifiers) into a classifier with high accuracy (a "strong" classifier) by **upweighting** the examples that the model is currently misclassfying.

bounding box

In an image, the (x, y) coordinates of a rectangle around an area of interest, such as the dog in the image below.

broadcasting

Expanding the shape of an operand in a matrix math operation to **dimensions** compatible for that operation. For instance, linear algebra requires that the two operands in a matrix addition operation must have the same dimensions. Consequently, you can't add a matrix of shape (m, n) to a vector of length n. Broadcasting enables this operation by virtually expanding the vector of length n to a matrix of shape (m,n) by replicating the same values down each column.

For example, given the following definitions, linear algebra prohibits A+B because A and B have different dimensions:

A = [[7, 10, 4], [13, 5, 9]] B = [2]

However, broadcasting enables the operation A+B by virtually expanding B to:

[[2, 2, 2], [2, 2, 2]]

Thus, A+B is now a valid operation:

[[7, 10, 4], + [[2, 2, 2], = [[ 9, 12, 6], [13, 5, 9]] [2, 2, 2]] [15, 7, 11]]

See the following description of broadcasting in NumPy for more details.

bucketing

Converting a (usually **continuous**) feature into multiple binary features called buckets or bins, typically based on value range. For example, instead of representing temperature as a single continuous floating-point feature, you could chop ranges of temperatures into discrete bins. Given temperature data sensitive to a tenth of a degree, all temperatures between 0.0 and 15.0 degrees could be put into one bin, 15.1 to 30.0 degrees could be a second bin, and 30.1 to 50.0 degrees could be a third bin.

calibration layer

A post-prediction adjustment, typically to account for **prediction bias**. The adjusted predictions and probabilities should match the distribution of an observed set of labels.

candidate generation

The initial set of recommendations chosen by a recommendation system. For example, consider a bookstore that offers 100,000 titles. The candidate generation phase creates a much smaller list of suitable books for a particular user, say 500. But even 500 books is way too many to recommend to a user. Subsequent, more expensive, phases of a recommendation system (such as **scoring** and **re-ranking**) whittle down those 500 to a much smaller, more useful set of recommendations.

candidate sampling

A training-time optimization in which a probability is calculated for all the positive labels, using, for example, **softmax**, but only for a random sample of negative labels. For example, if we have an example labeled beagle and dog candidate sampling computes the predicted probabilities and corresponding loss terms for the beagle and dog class outputs in addition to a random subset of the remaining classes (cat, lollipop, fence). The idea is that the **negative classes** can learn from less frequent negative reinforcement as long as **positive classes** always get proper positive reinforcement, and this is indeed observed empirically. The motivation for candidate sampling is a computational efficiency win from not computing predictions for all negatives.

categorical data

**Features** having a discrete set of possible values. For example, consider a categorical feature named house style, which has a discrete set of three possible values: Tudor, ranch, colonial. By representing house style as categorical data, the model can learn the separate impacts of Tudor, ranch, and colonial on house price.

Sometimes, values in the discrete set are mutually exclusive, and only one value can be applied to a given example. For example, a car maker categorical feature would probably permit only a single value (Toyota) per example. Other times, more than one value may be applicable. A single car could be painted more than one different color, so a car color categorical feature would likely permit a single example to have multiple values (for example, red and white).

Categorical features are sometimes called **discrete features**.

Contrast with **numerical data**.

centroid

The center of a cluster as determined by a **k-means** or **k-median** algorithm. For instance, if k is 3, then the k-means or k-median algorithm finds 3 centroids.

centroid-based clustering

A category of **clustering** algorithms that organizes data into nonhierarchical clusters. **k-means** is the most widely used centroid-based clustering algorithm.

Contrast with **hierarchical clustering** algorithms.

checkpoint

Data that captures the state of the variables of a model at a particular time. Checkpoints enable exporting model **weights**, as well as performing training across multiple sessions. Checkpoints also enable training to continue past errors (for example, job preemption). Note that the **graph** itself is not included in a checkpoint.

class

One of a set of enumerated target values for a label. For example, in a **binary classification** model that detects spam, the two classes are spam and not spam. In a **multi-class classification** model that identifies dog breeds, the classes would be poodle, beagle, pug, and so on.

classification model

A type of machine learning model for distinguishing among two or more discrete classes. For example, a natural language processing classification model could determine whether an input sentence was in French, Spanish, or Italian. Compare with **regression model**.

classification threshold

A scalar-value criterion that is applied to a model's predicted score in order to separate the **positive class** from the **negative class**. Used when mapping **logistic regression** results to **binary classification**. For example, consider a logistic regression model that determines the probability of a given email message being spam. If the classification threshold is 0.9, then logistic regression values above 0.9 are classified as spam and those below 0.9 are classified as not spam.

class-imbalanced dataset

A **binary classification** problem in which the **labels** for the two classes have significantly different frequencies. For example, a disease dataset in which 0.0001 of examples have positive labels and 0.9999 have negative labels is a class-imbalanced problem, but a football game predictor in which 0.51 of examples label one team winning and 0.49 label the other team winning is not a class-imbalanced problem.

clipping

A technique for handling **outliers**. Specifically, reducing feature values that are greater than a set maximum value down to that maximum value. Also, increasing feature values that are less than a specific minimum value up to that minimum value.

For example, suppose that only a few feature values fall outside the range 40–60. In this case, you could do the following:

Clip all values over 60 to be exactly 60.

Clip all values under 40 to be exactly 40.

In addition to bringing input values within a designated range, clipping can also used to force gradient values within a designated range during training.

Cloud TPU

A specialized hardware accelerator designed to speed up machine learning workloads on Google Cloud Platform.

clustering

Grouping related **examples**, particularly during **unsupervised learning**. Once all the examples are grouped, a human can optionally supply meaning to each cluster.

Many clustering algorithms exist. For example, the **k-means** algorithm clusters examples based on their proximity to a **centroid**, as in the following diagram:

A human researcher could then review the clusters and, for example, label cluster 1 as "dwarf trees" and cluster 2 as "full-size trees."

As another example, consider a clustering algorithm based on an example's distance from a center point, illustrated as follows:

co-adaptation

When **neurons** predict patterns in training data by relying almost exclusively on outputs of specific other neurons instead of relying on the network's behavior as a whole. When the patterns that cause co-adaption are not present in validation data, then co-adaptation causes overfitting. **Dropout regularization** reduces co-adaptation because dropout ensures neurons cannot rely solely on specific other neurons.

collaborative filtering

Making **predictions** about the interests of one user based on the interests of many other users. Collaborative filtering is often used in **recommendation systems**.

confirmation bias

The tendency to search for, interpret, favor, and recall information in a way that confirms one's preexisting beliefs or hypotheses. Machine learning developers may inadvertently collect or label data in ways that influence an outcome supporting their existing beliefs. Confirmation bias is a form of **implicit bias**.

Experimenter's bias is a form of confirmation bias in which an experimenter continues training models until a preexisting hypothesis is confirmed.

confusion matrix

An NxN table that summarizes how successful a **classification model's** predictions were; that is, the correlation between the label and the model's classification. One axis of a confusion matrix is the **label** that the model predicted, and the other axis is the actual label. N represents the number of **classes**. In a **binary classification** problem, N=2. For example, here is a sample confusion matrix for a binary classification problem:

Tumor (predicted) Non-Tumor (predicted)

Tumor (actual) 18 1

Non-Tumor (actual) 6 452

The preceding confusion matrix shows that of the 19 samples that actually had tumors, the model correctly classified 18 as having tumors (18 **true positives**), and incorrectly classified 1 as not having a tumor (1 **false negative**). Similarly, of 458 samples that actually did not have tumors, 452 were correctly classified (452 **true negatives**) and 6 were incorrectly classified (6 **false positives**).

The confusion matrix for a **multi-class classification** problem can help you determine mistake patterns. For example, a confusion matrix could reveal that a model trained to recognize handwritten digits tends to mistakenly predict 9 instead of 4, or 1 instead of 7.

Confusion matrices contain sufficient information to calculate a variety of performance metrics, including **precision** and **recall**.

continuous feature

A floating-point feature with an infinite range of possible values. Contrast with **discrete feature**.

convenience sampling

Using a dataset not gathered scientifically in order to run quick experiments. Later on, it's essential to switch to a scientifically gathered dataset.

convergence

Informally, often refers to a state reached during **training** in which training **loss** and **validation** loss change very little or not at all with each iteration after a certain number of iterations. In other words, a model reaches convergence when additional training on the current data will not improve the model. In **deep learning**, loss values sometimes stay constant or nearly so for many iterations before finally descending, temporarily producing a false sense of convergence.

See also **early stopping**.

See also Boyd and Vandenberghe, Convex Optimization.

convex function

A function in which the region above the graph of the function is a **convex set**. The prototypical convex function is shaped something like the letter U. For example, the following are all convex functions:

By contrast, the following function is not convex. Notice how the region above the graph is not a convex set:

A strictly convex function has exactly one local minimum point, which is also the global minimum point. The classic U-shaped functions are strictly convex functions. However, some convex functions (for example, straight lines) are not U-shaped.

A lot of the common **loss functions**, including the following, are convex functions:

**L2 loss**

**Log Loss**

**L1 regularization**

**L2 regularization**

Many variations of **gradient descent** are guaranteed to find a point close to the minimum of a strictly convex function. Similarly, many variations of **stochastic gradient descent** have a high probability (though, not a guarantee) of finding a point close to the minimum of a strictly convex function.

The sum of two convex functions (for example, L2 loss + L1 regularization) is a convex function.

**Deep models** are never convex functions. Remarkably, algorithms designed for **convex optimization** tend to find reasonably good solutions on deep networks anyway, even though those solutions are not guaranteed to be a global minimum.

convex optimization

The process of using mathematical techniques such as **gradient descent** to find the minimum of a **convex function**. A great deal of research in machine learning has focused on formulating various problems as convex optimization problems and in solving those problems more efficiently.

For complete details, see Boyd and Vandenberghe, Convex Optimization.

convex set

A subset of Euclidean space such that a line drawn between any two points in the subset remains completely within the subset. For instance, the following two shapes are convex sets:

By contrast, the following two shapes are not convex sets:

convolution

In mathematics, casually speaking, a mixture of two functions. In machine learning, a convolution mixes the convolutional filter and the input matrix in order to train **weights**.

The term "convolution" in machine learning is often a shorthand way of referring to either **convolutional operation** or **convolutional layer**.

Without convolutions, a machine learning algorithm would have to learn a separate weight for every cell in a large **tensor**. For example, a machine learning algorithm training on 2K x 2K images would be forced to find 4M separate weights. Thanks to convolutions, a machine learning algorithm only has to find weights for every cell in the **convolutional filter**, dramatically reducing the memory needed to train the model. When the convolutional filter is applied, it is simply replicated across cells such that each is multiplied by the filter.

convolutional filter

One of the two actors in a **convolutional operation**. (The other actor is a slice of an input matrix.) A convolutional filter is a matrix having the same **rank** as the input matrix, but a smaller shape. For example, given a 28x28 input matrix, the filter could be any 2D matrix smaller than 28x28.

In photographic manipulation, all the cells in a convolutional filter are typically set to a constant pattern of ones and zeroes. In machine learning, convolutional filters are typically seeded with random numbers and then the network **trains** the ideal values.

convolutional layer

A layer of a **deep neural network** in which a **convolutional filter** passes along an input matrix. For example, consider the following 3x3 **convolutional filter**:

The following animation shows a convolutional layer consisting of 9 convolutional operations involving the 5x5 input matrix. Notice that each convolutional operation works on a different 3x3 slice of the input matrix. The resulting 3x3 matrix (on the right) consists of the results of the 9 convolutional operations:

convolutional neural network

A **neural network** in which at least one layer is a **convolutional layer**. A typical convolutional neural network consists of some combination of the following layers:

**convolutional layers**

**pooling layers**

**dense layers**

Convolutional neural networks have had great success in certain kinds of problems, such as image recognition.

convolutional operation

The following two-step mathematical operation:

Element-wise multiplication of the **convolutional filter** and a slice of an input matrix. (The slice of the input matrix has the same rank and size as the convolutional filter.)

Summation of all the values in the resulting product matrix.

For example, consider the following 5x5 input matrix:

Now imagine the following 2x2 convolutional filter:

Each convolutional operation involves a single 2x2 slice of the input matrix. For instance, suppose we use the 2x2 slice at the top-left of the input matrix. So, the convolution operation on this slice looks as follows:

A **convolutional layer** consists of a series of convolutional operations, each acting on a different slice of the input matrix.

cost

Synonym for **loss**.

counterfactual fairness

A **fairness metric** that checks whether a classifier produces the same result for one individual as it does for another individual who is identical to the first, except with respect to one or more **sensitive attributes**. Evaluating a classifier for counterfactual fairness is one method for surfacing potential sources of bias in a model.

See "When Worlds Collide: Integrating Different Counterfactual Assumptions in Fairness" for a more detailed discussion of counterfactual fairness.

coverage bias

See **selection bias**.

crash blossom

A sentence or phrase with an ambiguous meaning. Crash blossoms present a significant problem in **natural language understanding**. For example, the headline Red Tape Holds Up Skyscraper is a crash blossom because an NLU model could interpret the headline literally or figuratively.

critic

Synonym for **Deep Q-Network**.

cross-entropy

A generalization of **Log Loss** to **multi-class classification problems**. Cross-entropy quantifies the difference between two probability distributions. See also **perplexity**.

cross-validation

A mechanism for estimating how well a model will generalize to new data by testing the model against one or more non-overlapping data subsets withheld from the **training set**.

custom Estimator

An **Estimator** that you write yourself by following these directions.

Contrast with **premade Estimators**.

data analysis

Obtaining an understanding of data by considering samples, measurement, and visualization. Data analysis can be particularly useful when a dataset is first received, before one builds the first model. It is also crucial in understanding experiments and debugging problems with the system.

data augmentation

Artificially boosting the range and number of **training** examples by transforming existing examples to create additional examples. For example, suppose images are one of your features, but your dataset doesn't contain enough image examples for the model to learn useful associations. Ideally, you'd add enough **labeled** images to your dataset to enable your model to train properly. If that's not possible, data augmentation can rotate, stretch, and reflect each image to produce many variants of the original picture, possibly yielding enough labeled data to enable excellent training.

DataFrame

A popular datatype for representing datasets in **pandas**. A DataFrame is analogous to a table. Each column of the DataFrame has a name (a header), and each row is identified by a number.

data set or dataset

A collection of **examples**.

Dataset API (tf.data)

A high-level **TensorFlow** API for reading data and transforming it into a form that a machine learning algorithm requires. A tf.data.Dataset object represents a sequence of elements, in which each element contains one or more **Tensors**. A tf.data.Iterator object provides access to the elements of a Dataset.

For details about the Dataset API, see Importing Data in the TensorFlow Programmer's Guide.

decision boundary

The separator between classes learned by a model in a **binary class** or **multi-class classification problems**. For example, in the following image representing a binary classification problem, the decision boundary is the frontier between the orange class and the blue class:

decision threshold

Synonym for **classification threshold**.

decision tree

A model represented as a sequence of branching statements. For example, the following over-simplified decision tree branches a few times to predict the price of a house (in thousands of USD). According to this decision tree, a house larger than 160 square meters, having more than three bedrooms, and built less than 10 years ago would have a predicted price of 510 thousand USD.

Machine learning can generate deep decision trees.

deep model

A type of **neural network** containing multiple **hidden layers**.

Contrast with **wide model**.

deep neural network

Synonym for **deep model**.

Deep Q-Network (DQN)

In **Q-learning**, a deep **neural network** that predicts **Q-functions**.

Critic is a synonym for Deep Q-Network.

demographic parity

A **fairness metric** that is satisfied if the results of a model's classification are not dependent on a given **sensitive attribute**.

For example, if both Lilliputians and Brobdingnagians apply to Glubbdubdrib University, demographic parity is achieved if the percentage of Lilliputians admitted is the same as the percentage of Brobdingnagians admitted, irrespective of whether one group is on average more qualified than the other.

Contrast with **equalized odds** and **equality of opportunity**, which permit classification results in aggregate to depend on sensitive attributes, but do not permit classification results for certain specified ground-truth labels to depend on sensitive attributes. See "Attacking discrimination with smarter machine learning" for a visualization exploring the tradeoffs when optimizing for demographic parity.

dense feature

A **feature** in which most values are non-zero, typically a **Tensor** of floating-point values. Contrast with **sparse feature**.

dense layer

Synonym for **fully connected layer**.

depth

The number of **layers** (including any **embedding** layers) in a **neural network** that learn weights. For example, a neural network with 5 **hidden layers** and 1 output layer has a depth of 6.

depthwise separable convolutional neural network (sepCNN)

A **convolutional neural network** architecture based on Inception, but where Inception modules are replaced with depthwise separable convolutions. Also known as Xception.

A depthwise separable convolution (also abbreviated as separable convolution) factors a standard 3-D convolution into two separate convolution operations that are more computationally efficient: first, a depthwise convolution, with a depth of 1 (n ✕ n ✕ 1), and then second, a pointwise convolution, with length and width of 1 (1 ✕ 1 ✕ n).

To learn more, see Xception: Deep Learning with Depthwise Separable Convolutions.

device

A category of hardware that can run a TensorFlow session, including CPUs, GPUs, and **TPUs**.

dimension reduction

Decreasing the number of dimensions used to represent a particular feature in a feature vector, typically by converting to an **embedding**.

dimensions

Overloaded term having any of the following definitions:

The number of levels of coordinates in a **Tensor**. For example:

A scalar has zero dimensions; for example, ["Hello"].

A vector has one dimension; for example, [3, 5, 7, 11].

A matrix has two dimensions; for example, [[2, 4, 18], [5, 7, 14]].

You can uniquely specify a particular cell in a one-dimensional vector with one coordinate; you need two coordinates to uniquely specify a particular cell in a two-dimensional matrix.

The number of entries in a **feature vector**.

The number of elements in an **embedding** layer.

discrete feature

A **feature** with a finite set of possible values. For example, a feature whose values may only be animal, vegetable, or mineral is a discrete (or categorical) feature. Contrast with **continuous feature**.

discriminative model

A **model** that predicts labels from a set of one or more features. More formally, discriminative models define the conditional probability of an output given the features and weights; that is:

p(output | features, weights)

For example, a model that predicts whether an email is spam from features and weights is a discriminative model.

The vast majority of supervised learning models, including classification and regression models, are discriminative models.

Contrast with **generative model**.

discriminator

A system that determines whether examples are real or fake.

The subsystem within a **generative adversarial network** that determines whether the examples created by the **generator** are real or fake.

disparate impact

Making decisions about people that impact different population subgroups disproportionately. This usually refers to situations where an algorithmic decision-making process harms or benefits some subgroups more than others.

For example, suppose an algorithm that determines a Lilliputian's eligibility for a miniature-home loan is more likely to classify them as “ineligible” if their mailing address contains a certain postal code. If Big-Endian Lilliputians are more likely to have mailing addresses with this postal code than Little-Endian Lilliputians, then this algorithm may result in disparate impact.

Contrast with **disparate treatment**, which focuses on disparities that result when subgroup characteristics are explicit inputs to an algorithmic decision-making process.

disparate treatment

Factoring subjects' **sensitive attributes** into an algorithmic decision-making process such that different subgroups of people are treated differently.

For example, consider an algorithm that determines Lilliputians’ eligibility for a miniature-home loan based on the data they provide in their loan application. If the algorithm uses a Lilliputian’s affiliation as Big-Endian or Little-Endian as an input, it is enacting disparate treatment along that dimension.

Contrast with **disparate impact**, which focuses on disparities in the societal impacts of algorithmic decisions on subgroups, irrespective of whether those subgroups are inputs to the model.

Warning: Because sensitive attributes are almost always correlated with other features the data may have, explicitly removing sensitive attribute information does not guarantee that subgroups will be treated equally. For example, removing sensitive demographic attributes from a training data set that still includes postal code as a feature may address disparate treatment of subgroups, but there still might be disparate impact upon these groups because postal code might serve as a **proxy** for other demographic information.

divisive clustering

See **hierarchical clustering**.

downsampling

Overloaded term that can mean either of the following:

Reducing the amount of information in a feature in order to train a model more efficiently. For example, before training an image recognition model, downsampling high-resolution images to a lower-resolution format.

Training on a disproportionately low percentage of over-represented class examples in order to improve model training on under-represented classes. For example, in a **class-imbalanced dataset**, models tend to learn a lot about the **majority class** and not enough about the **minority class**. Downsampling helps balance the amount of training on the majority and minority classes.

DQN

Abbreviation for **Deep Q-Network**.

dropout regularization

A form of **regularization** useful in training **neural networks**. Dropout regularization works by removing a random selection of a fixed number of the units in a network layer for a single gradient step. The more units dropped out, the stronger the regularization. This is analogous to training the network to emulate an exponentially large ensemble of smaller networks. For full details, see Dropout: A Simple Way to Prevent Neural Networks from Overfitting.

dynamic model

A **model** that is trained online in a continuously updating fashion. That is, data is continuously entering the model.

eager execution

A TensorFlow programming environment in which **operations** run immediately. By contrast, operations called in **graph execution** don't run until they are explicitly evaluated. Eager execution is an imperative interface, much like the code in most programming languages. Eager execution programs are generally far easier to debug than graph execution programs.

early stopping

A method for **regularization** that involves ending model training before training loss finishes decreasing. In early stopping, you end model training when the loss on a **validation dataset** starts to increase, that is, when **generalization** performance worsens.

embeddings

A categorical feature represented as a continuous-valued feature. Typically, an embedding is a translation of a high-dimensional vector into a low-dimensional space. For example, you can represent the words in an English sentence in either of the following two ways:

As a million-element (high-dimensional) **sparse vector** in which all elements are integers. Each cell in the vector represents a separate English word; the value in a cell represents the number of times that word appears in a sentence. Since a single English sentence is unlikely to contain more than 50 words, nearly every cell in the vector will contain a 0. The few cells that aren't 0 will contain a low integer (usually 1) representing the number of times that word appeared in the sentence.

As a several-hundred-element (low-dimensional) **dense vector** in which each element holds a floating-point value between 0 and 1. This is an embedding.

In TensorFlow, embeddings are trained by **backpropagating** **loss** just like any other parameter in a **neural network**.

embedding space

The d-dimensional vector space that features from a higher-dimensional vector space are mapped to. Ideally, the embedding space contains a structure that yields meaningful mathematical results; for example, in an ideal embedding space, addition and subtraction of embeddings can solve word analogy tasks.

The dot product of two embeddings is a measure of their similarity.

empirical risk minimization (ERM)

Choosing the function that minimizes loss on the training set. Contrast with **structural risk minimization**.

ensemble

A merger of the predictions of multiple **models**. You can create an ensemble via one or more of the following:

different initializations

different **hyperparameters**

different overall structure

Deep and wide models are a kind of ensemble.

environment

In reinforcement learning, the world that contains the **agent** and allows the agent to observe that world's **state**. For example, the represented world can be a game like chess, or a physical world like a maze. When the agent applies an **action** to the environment, then the environment transitions between states.

episode

In reinforcement learning, each of the repeated attempts by the **agent** to learn an **environment**.

epoch

A full training pass over the entire dataset such that each example has been seen once. Thus, an epoch represents N/**batch size** training **iterations**, where N is the total number of examples.

epsilon greedy policy

In reinforcement learning, a **policy** that either follows a **random policy** with epsilon probability or a **greedy policy** otherwise. For example, if epsilon is 0.9, then the policy follows a random policy 90% of the time and a greedy policy 10% of the time.

Over successive episodes, the algorithm reduces epsilon’s value in order to shift from following a random policy to following a greedy policy. By shifting the policy, the agent first randomly explores the environment and then greedily exploits the results of random exploration.

equality

A **fairness metric** that checks whether, for a preferred **label** (one that confers an advantage or benefit to a person) and a given **attribute**, a classifier predicts that preferred label equally well for all values of that attribute. In other words, equality of opportunity measures whether the people who should qualify for an opportunity are equally likely to do so regardless of their group membership.

For example, suppose Glubbdubdrib University admits both Lilliputians and Brobdingnagians to a rigorous mathematics program. Lilliputians’ secondary schools offer a robust curriculum of math classes, and the vast majority of students are qualified for the university program. Brobdingnagians’ secondary schools don’t offer math classes at all, and as a result, far fewer of their students are qualified. Equality of opportunity is satisfied for the preferred label of "admitted" with respect to nationality (Lilliputian or Brobdingnagian) if qualified students are equally likely to be admitted irrespective of whether they're a Lilliputian or a Brobdingnagian.

For example, let's say 100 Lilliputians and 100 Brobdingnagians apply to Glubbdubdrib University, and admissions decisions are made as follows:

Table 1. Lilliputian applicants (90% are qualified)

Qualified Unqualified

Admitted 45 3

Rejected 45 7

Total 90 10

Percentage of qualified students admitted: 45/90 = 50%

Percentage of unqualified students rejected: 7/10 = 70%

Total percentage of Lilliputian students admitted: (45+3)/100 = 48%

Table 2. Brobdingnagian applicants (10% are qualified):

Qualified Unqualified

Admitted 5 9

Rejected 5 81

Total 10 90

Percentage of qualified students admitted: 5/10 = 50%

Percentage of unqualified students rejected: 81/90 = 90%

Total percentage of Brobdingnagian students admitted: (5+9)/100 = 14%

The preceding examples satisfy equality of opportunity for acceptance of qualified students because qualified Lilliputians and Brobdingnagians both have a 50% chance of being admitted.

Note: While equality of opportunity is satisfied, the following two fairness metrics are not satisfied:

**demographic parity**: Lilliputians and Brobdingnagians are admitted to the university at different rates; 48% of Lilliputians students are admitted, but only 14% of Brobdingnagian students are admitted.

**equalized odds**: While qualified Lilliputian and Brobdingnagian students both have the same chance of being admitted, the additional constraint that unqualified Lilliputians and Brobdingnagians both have the same chance of being rejected is not satisfied. Unqualified Lilliputians have a 70% rejection rate, whereas unqualified Brobdingnagians have a 90% rejection rate.

See "Equality of Opportunity in Supervised Learning" for a more detailed discussion of equality of opportunity. Also see "Attacking discrimination with smarter machine learning" for a visualization exploring the tradeoffs when optimizing for equality of opportunity.

equalized odds

A **fairness metric** that checks if, for any particular label and attribute, a classifier predicts that label equally well for all values of that attribute.

For example, suppose Glubbdubdrib University admits both Lilliputians and Brobdingnagians to a rigorous mathematics program. Lilliputians' secondary schools offer a robust curriculum of math classes, and the vast majority of students are qualified for the university program. Brobdingnagians' secondary schools don’t offer math classes at all, and as a result, far fewer of their students are qualified. Equalized odds is satisfied provided that no matter whether an applicant is a Lilliputian or a Brobdingnagian, if they are qualified, they are equally as likely to get admitted to the program, and if they are not qualified, they are equally as likely to get rejected.

Let’s say 100 Lilliputians and 100 Brobdingnagians apply to Glubbdubdrib University, and admissions decisions are made as follows:

Table 3. Lilliputian applicants (90% are qualified)

Qualified Unqualified

Admitted 45 2

Rejected 45 8

Total 90 10

Percentage of qualified students admitted: 45/90 = 50%

Percentage of unqualified students rejected: 8/10 = 80%

Total percentage of Lilliputian students admitted: (45+2)/100 = 47%

Table 4. Brobdingnagian applicants (10% are qualified):

Qualified Unqualified

Admitted 5 18

Rejected 5 72

Total 10 90

Percentage of qualified students admitted: 5/10 = 50%

Percentage of unqualified students rejected: 72/90 = 80%

Total percentage of Brobdingnagian students admitted: (5+18)/100 = 23%

Equalized odds is satisfied because qualified Lilliputian and Brobdingnagian students both have a 50% chance of being admitted, and unqualified Lilliputian and Brobdingnagian have an 80% chance of being rejected.

Note: While equalized odds is satisfied here, **demographic parity** is not satisfied. Lilliputian and Brobdingnagian students are admitted to Glubbdubdrib University at different rates; 47% of Lilliputian students are admitted, and 23% of Brobdingnagian students are admitted. Equalized odds is formally defined in "Equality of Opportunity in Supervised Learning" as follows: "predictor Ŷ satisfies equalized odds with respect to protected attribute A and outcome Y if Ŷ and A are independent, conditional on Y."

Note: Contrast equalized odds with the more relaxed **equality of opportunity** metric.

Estimator

An instance of the tf.Estimator class, which encapsulates logic that builds a TensorFlow graph and runs a TensorFlow session. You may create your own **custom Estimators** (as described here) or instantiate **premade Estimators** created by others.

example

One row of a dataset. An example contains one or more **features** and possibly a **label**. See also **labeled example** and **unlabeled example**.

experience replay

In reinforcement learning, a **DQN** technique used to reduce temporal correlations in training data. The **agent** stores state transitions in a **replay buffer**, and then samples transitions from the replay buffer to create training data.

experimenter's bias

See **confirmation bias**.

exploding gradient problem

The tendency for **gradients** in a **deep neural networks** (especially **recurrent neural networks**) to become surprisingly steep (high). Steep gradients result in very large updates to the weights of each node in a deep neural network.

Models suffering from the exploding gradient problem become difficult or impossible to train. **Gradient clipping** can mitigate this problem.

Compare to **vanishing gradient problem**.

fairness constraint

Applying a constraint to an algorithm to ensure one or more definitions of fairness are satisfied. Examples of fairness constraints include:

**Post-processing** your model's output.

Altering the **loss function** to incorporate a penalty for violating a **fairness metric**.

Directly adding a mathematical constraint to an optimization problem.

fairness metric

A mathematical definition of “fairness” that is measurable. Some commonly used fairness metrics include:

**equalized odds**

**predictive parity**

**counterfactual fairness**

**demographic parity**

Many fairness metrics are mutually exclusive; see **incompatibility of fairness metrics**.

false negative (FN)

An example in which the model mistakenly predicted the **negative class**. For example, the model inferred that a particular email message was not spam (the negative class), but that email message actually was spam.

false positive (FP)

An example in which the model mistakenly predicted the **positive class**. For example, the model inferred that a particular email message was spam (the positive class), but that email message was actually not spam.

feature

An input variable used in making **predictions**.

Feature column (tf.feature_column)

A function that specifies how a model should interpret a particular feature. A list that collects the output returned by calls to such functions is a required parameter to all **Estimators** constructors.

The tf.feature_column functions enable models to easily experiment with different representations of input features. For details, see the Feature Columns chapter in the TensorFlow Programmers Guide.

"Feature column" is Google-specific terminology. A feature column is referred to as a "namespace" in the VW system (at Yahoo/Microsoft), or a field.

feature cross

A **synthetic feature** formed by crossing (taking a Cartesian product of) individual binary features obtained from **categorical data** or from **continuous features** via **bucketing**. Feature crosses help represent nonlinear relationships.

feature engineering

The process of determining which **features** might be useful in training a model, and then converting raw data from log files and other sources into said features. In TensorFlow, feature engineering often means converting raw log file entries to **tf.Example** protocol buffers. See also tf.Transform.

Feature engineering is sometimes called feature extraction.

feature extraction

Overloaded term having either of the following definitions:

Retrieving intermediate feature representations calculated by an **unsupervised** or pretrained model (for example, **hidden layer** values in a **neural network**) for use in another model as input.

Synonym for **feature engineering**.

feature set

The group of **features** your machine learning model trains on. For example, postal code, property size, and property condition might comprise a simple feature set for a model that predicts housing prices.

feature spec

Describes the information required to extract **features** data from the **tf.Example** protocol buffer. Because the tf.Example protocol buffer is just a container for data, you must specify the following:

the data to extract (that is, the keys for the features)

the data type (for example, float or int)

The length (fixed or variable)

The **Estimator API** provides facilities for producing a feature spec from a list of **FeatureColumns**.

feature vector

The list of feature values representing an **example** passed into a model.

federated learning

A distributed machine learning approach that **trains** machine learning **models** using decentralized **examples** residing on devices such as smartphones. In federated learning, a subset of devices downloads the current model from a central coordinating server. The devices use the examples stored on the devices to make improvements to the model. The devices then upload the model improvements (but not the training examples) to the coordinating server, where they are aggregated with other updates to yield an improved global model. After the aggregation, the model updates computed by devices are no longer needed, and can be discarded.

Since the training examples are never uploaded, federated learning follows the privacy principles of focused data collection and data minimization.

For more information about federated learning, see this tutorial.

feedback loop

In machine learning, a situation in which a model's predictions influence the training data for the same model or another model. For example, a model that recommends movies will influence the movies that people see, which will then influence subsequent movie recommendation models.

feedforward neural network (FFN)

A neural network without cyclic or recursive connections. For example, traditional **deep neural networks** are feedforward neural networks. Contrast with **recurrent neural networks**, which are cyclic.

few-shot learning

A machine learning approach, often used for object classification, designed to learn effective classifiers from only a small number of training examples.

See also **one-shot learning**.

fine tuning

Perform a secondary optimization to adjust the parameters of an already trained **model** to fit a new problem. Fine tuning often refers to refitting the weights of a trained **unsupervised** model to a **supervised** model.

forget gate

The portion of a **Long Short-Term Memory** cell that regulates the flow of information through the cell. Forget gates maintain context by deciding which information to discard from the cell state.

full softmax

See **softmax**. Contrast with **candidate sampling**.

fully connected layer

A **hidden layer** in which each **node** is connected to every node in the subsequent hidden layer.

A fully connected layer is also known as a **dense layer**.

GAN

Abbreviation for **generative adversarial network**.

generalization

Refers to your model's ability to make correct predictions on new, previously unseen data as opposed to the data used to train the model.

generalization curve

A **loss curve** showing both the **training set** and the **validation set**. A generalization curve can help you detect possible **overfitting**. For example, the following generalization curve suggests overfitting because loss for the validation set ultimately becomes significantly higher than for the training set.

generalized linear model

A generalization of **least squares regression** models, which are based on Gaussian noise, to other types of models based on other types of noise, such as Poisson noise or categorical noise. Examples of generalized linear models include:

**logistic regression**

multi-class regression

least squares regression

The parameters of a generalized linear model can be found through **convex optimization**.

Generalized linear models exhibit the following properties:

The average prediction of the optimal least squares regression model is equal to the average label on the training data.

The average probability predicted by the optimal logistic regression model is equal to the average label on the training data.

The power of a generalized linear model is limited by its features. Unlike a deep model, a generalized linear model cannot "learn new features."

generative adversarial network (GAN)

A system to create new data in which a **generator** creates data and a **discriminator** determines whether that created data is valid or invalid.

generative model

Practically speaking, a model that does either of the following:

Creates (generates) new examples from the training dataset. For example, a generative model could create poetry after training on a dataset of poems. The **generator** part of a **generative adversarial network** falls into this category.

Determines the probability that a new example comes from the training set, or was created from the same mechanism that created the training set. For example, after training on a dataset consisting of English sentences, a generative model could determine the probability that new input is a valid English sentence.

A generative model can theoretically discern the distribution of examples or particular features in a dataset. That is:

p(examples)

Unsupervised learning models are generative.

Contrast with **discriminative models**.

generator

The subsystem within a **generative adversarial network** that creates new **examples**.

Contrast with **discriminative model**.

gradient

The vector of **partial derivatives** with respect to all of the independent variables. In machine learning, the gradient is the vector of partial derivatives of the model function. The gradient points in the direction of steepest ascent.

gradient clipping

A commonly used mechanism to mitigate the **exploding gradient problem** by artificially limiting (clipping) the maximum value of gradients when using **gradient descent** to train a model.

gradient descent

A technique to minimize **loss** by computing the gradients of loss with respect to the model's parameters, conditioned on training data. Informally, gradient descent iteratively adjusts parameters, gradually finding the best combination of **weights** and bias to minimize loss.

graph

In TensorFlow, a computation specification. Nodes in the graph represent operations. Edges are directed and represent passing the result of an operation (a **Tensor**) as an operand to another operation. Use **TensorBoard** to visualize a graph.

graph execution

A TensorFlow programming environment in which the program first constructs a **graph** and then executes all or part of that graph. Graph execution is the default execution mode in TensorFlow 1.x.

Contrast with **eager execution**.

greedy policy

In reinforcement learning, a **policy** that always chooses the action with the highest expected **return**.

ground truth

The correct answer. Reality. Since reality is often subjective, expert **raters** typically are the proxy for ground truth.

group attribution bias

Assuming that what is true for an individual is also true for everyone in that group. The effects of group attribution bias can be exacerbated if a **convenience sampling** is used for data collection. In a non-representative sample, attributions may be made that do not reflect reality.

See also **out-group homogeneity bias** and **in-group bias**.

hashing

In machine learning, a mechanism for bucketing **categorical data**, particularly when the number of categories is large, but the number of categories actually appearing in the dataset is comparatively small.

For example, Earth is home to about 60,000 tree species. You could represent each of the 60,000 tree species in 60,000 separate categorical buckets. Alternatively, if only 200 of those tree species actually appear in a dataset, you could use hashing to divide tree species into perhaps 500 buckets.

A single bucket could contain multiple tree species. For example, hashing could place baobab and red maple—two genetically dissimilar species—into the same bucket. Regardless, hashing is still a good way to map large categorical sets into the desired number of buckets. Hashing turns a categorical feature having a large number of possible values into a much smaller number of values by grouping values in a deterministic way.

For more information on hashing, see the Feature Columns chapter in the TensorFlow Programmers Guide.

heuristic

A quick solution to a problem, which may or may not be the best solution. For example, "With a heuristic, we achieved 86% accuracy. When we switched to a deep neural network, accuracy went up to 98%."

hidden layer

A synthetic layer in a **neural network** between the **input layer** (that is, the features) and the **output layer** (the prediction). Hidden layers typically contain an **activation function** (such as **ReLU**) for training. A **deep neural network** contains more than one hidden layer.

hierarchical clustering

A category of **clustering** algorithms that create a tree of clusters. Hierarchical clustering is well-suited to hierarchical data, such as botanical taxonomies. There are two types of hierarchical clustering algorithms:

Agglomerative clustering first assigns every example to its own cluster, and iteratively merges the closest clusters to create a hierarchical tree.

Divisive clustering first groups all examples into one cluster and then iteratively divides the cluster into a hierarchical tree.

Contrast with **centroid-based clustering**.

holdout data

**Examples** intentionally not used ("held out") during training. The **validation dataset** and **test dataset** are examples of holdout data. Holdout data helps evaluate your model's ability to generalize to data other than the data it was trained on. The loss on the holdout set provides a better estimate of the loss on an unseen dataset than does the loss on the training set.

hyperparameter

The "knobs" that you tweak during successive runs of training a model. For example, **learning rate** is a hyperparameter.

Contrast with **parameter**.

hyperplane

A boundary that separates a space into two subspaces. For example, a line is a hyperplane in two dimensions and a plane is a hyperplane in three dimensions. More typically in machine learning, a hyperplane is the boundary separating a high-dimensional space. **Kernel Support Vector Machines** use hyperplanes to separate positive classes from negative classes, often in a very high-dimensional space.

i.i.d.

Abbreviation for **independently and identically distributed**.

image recognition

A process that classifies object(s), pattern(s), or concept(s) in an image. Image recognition is also known as image classification.

For more information, see ML Practicum: Image Classification.

imbalanced dataset

Synonym for **class-imbalanced dataset**.

implicit bias

Automatically making an association or assumption based on one’s mental models and memories. Implicit bias can affect the following:

How data is collected and classified.

How machine learning systems are designed and developed.

For example, when building a classifier to identify wedding photos, an engineer may use the presence of a white dress in a photo as a feature. However, white dresses have been customary only during certain eras and in certain cultures.

See also **confirmation bias**.

incompatibility of fairness metrics

The idea that some notions of fairness are mutually incompatible and cannot be satisfied simultaneously. As a result, there is no single universal **metric** for quantifying fairness that can be applied to all ML problems.

While this may seem discouraging, incompatibility of fairness metrics doesn’t imply that fairness efforts are fruitless. Instead, it suggests that fairness must be defined contextually for a given ML problem, with the goal of preventing harms specific to its use cases.

See "On the (im)possibility of fairness" for a more detailed discussion of this topic.

independently and identically distributed (i.i.d)

Data drawn from a distribution that doesn't change, and where each value drawn doesn't depend on values that have been drawn previously. An i.i.d. is the ideal gas of machine learning—a useful mathematical construct but almost never exactly found in the real world. For example, the distribution of visitors to a web page may be i.i.d. over a brief window of time; that is, the distribution doesn't change during that brief window and one person's visit is generally independent of another's visit. However, if you expand that window of time, seasonal differences in the web page's visitors may appear.

individual fairness

A fairness metric that checks whether similar individuals are classified similarly. For example, Brobdingnagian Academy might want to satisfy individual fairness by ensuring that two students with identical grades and standardized test scores are equally likely to gain admission.

Note that individual fairness relies entirely on how you define "similarity" (in this case, grades and test scores), and you can run the risk of introducing new fairness problems if your similarity metric misses important information (such as the rigor of a student’s curriculum).

See "Fairness Through Awareness" for a more detailed discussion of individual fairness.

inference

In machine learning, often refers to the process of making predictions by applying the trained model to **unlabeled examples**. In statistics, inference refers to the process of fitting the parameters of a distribution conditioned on some observed data. (See the Wikipedia article on statistical inference.)

in-group bias

Showing partiality to one's own group or own characteristics. If testers or raters consist of the machine learning developer's friends, family, or colleagues, then in-group bias may invalidate product testing or the dataset.

In-group bias is a form of **group attribution bias**. See also **out-group homogeneity bias**.

input function

In TensorFlow, a function that returns input data to the training, evaluation, or prediction method of an **Estimator**. For example, the training input function returns a **batch** of features and labels from the **training set**.

input layer

The first layer (the one that receives the input data) in a **neural network**.

instance

Synonym for **example**.

interpretability

The degree to which a model's predictions can be readily explained. Deep models are often non-interpretable; that is, a deep model's different layers can be hard to decipher. By contrast, linear regression models and **wide models** are typically far more interpretable.

inter-rater agreement

A measurement of how often human raters agree when doing a task. If raters disagree, the task instructions may need to be improved. Also sometimes called inter-annotator agreement or inter-rater reliability. See also Cohen's kappa, which is one of the most popular inter-rater agreement measurements.

intersection over union (IoU)

The intersection of two sets divided by their union. In machine-learning image-detection tasks, IoU is used to measure the accuracy of the model’s predicted **bounding box** with respect to the **ground-truth** bounding box. In this case, the IoU for the two boxes is the ratio between the overlapping area and the total area, and its value ranges from 0 (no overlap of predicted bounding box and ground-truth bounding box) to 1 (predicted bounding box and ground-truth bounding box have the exact same coordinates).

For example, in the image below:

The predicted bounding box (the coordinates delimiting where the model predicts the night table in the painting is located) is outlined in purple.

The ground-truth bounding box (the coordinates delimiting where the night table in the painting is actually located) is outlined in green.

Here, the intersection of the bounding boxes for prediction and ground truth (below left) is 1, and the union of the bounding boxes for prediction and ground truth (below right) is 7, so the IoU is \(\frac{1}{7}\).

IoU

Abbreviation for **intersection over union**.

item matrix

In **recommendation systems**, a matrix of **embeddings** generated by **matrix factorization** that holds latent signals about each **item**. Each row of the item matrix holds the value of a single latent feature for all items. For example, consider a movie recommendation system. Each column in the item matrix represents a single movie. The latent signals might represent genres, or might be harder-to-interpret signals that involve complex interactions among genre, stars, movie age, or other factors.

The item matrix has the same number of columns as the target matrix that is being factorized. For example, given a movie recommendation system that evaluates 10,000 movie titles, the item matrix will have 10,000 columns.

items

In a **recommendation system**, the entities that a system recommends. For example, videos are the items that a video store recommends, while books are the items that a bookstore recommends.

iteration

A single update of a model's weights during training. An iteration consists of computing the gradients of the parameters with respect to the loss on a single **batch** of data.

Keras

A popular Python machine learning API. Keras runs on several deep learning frameworks, including TensorFlow, where it is made available as tf.keras.

keypoints

The coordinates of particular features in an image. For example, for an **image recognition** model that distinguishes flower species, keypoints might be the center of each petal, the stem, the stamen, and so on.

Kernel Support Vector Machines (KSVMs)

A classification algorithm that seeks to maximize the margin between **positive** and **negative classes** by mapping input data vectors to a higher dimensional space. For example, consider a classification problem in which the input dataset has a hundred features. To maximize the margin between positive and negative classes, a KSVM could internally map those features into a million-dimension space. KSVMs uses a loss function called **hinge loss**.

k-means

A popular **clustering** algorithm that groups examples in unsupervised learning. The k-means algorithm basically does the following:

Iteratively determines the best k center points (known as **centroids**).

Assigns each example to the closest centroid. Those examples nearest the same centroid belong to the same group.

The k-means algorithm picks centroid locations to minimize the cumulative square of the distances from each example to its closest centroid.

For example, consider the following plot of dog height to dog width:

If k=3, the k-means algorithm will determine three centroids. Each example is assigned to its closest centroid, yielding three groups:

Imagine that a manufacturer wants to determine the ideal sizes for small, medium, and large sweaters for dogs. The three centroids identify the mean height and mean width of each dog in that cluster. So, the manufacturer should probably base sweater sizes on those three centroids. Note that the centroid of a cluster is typically not an example in the cluster.

The preceding illustrations shows k-means for examples with only two features (height and width). Note that k-means can group examples across many features.

L1 loss

**Loss** function based on the absolute value of the difference between the values that a model is predicting and the actual values of the **labels**. L1 loss is less sensitive to outliers than **L2 loss**.

L1 regularization

A type of **regularization** that penalizes weights in proportion to the sum of the absolute values of the weights. In models relying on **sparse features**, L1 regularization helps drive the weights of irrelevant or barely relevant features to exactly 0, which removes those features from the model. Contrast with **L2 regularization**.

L2 loss

See **squared loss**.

L2 regularization

A type of **regularization** that penalizes weights in proportion to the sum of the squares of the weights. L2 regularization helps drive outlier weights (those with high positive or low negative values) closer to 0 but not quite to 0. (Contrast with **L1 regularization**.) L2 regularization always improves generalization in linear models.

label

In supervised learning, the "answer" or "result" portion of an **example**. Each example in a labeled dataset consists of one or more features and a label. For instance, in a housing dataset, the features might include the number of bedrooms, the number of bathrooms, and the age of the house, while the label might be the house's price. In a spam detection dataset, the features might include the subject line, the sender, and the email message itself, while the label would probably be either "spam" or "not spam."

labeled example

An example that contains **features** and a **label**. In supervised training, models learn from labeled examples.

lambda

Synonym for **regularization rate**.

(This is an overloaded term. Here we're focusing on the term's definition within **regularization**.)

landmarks

Synonym for **keypoints**.

layer

A set of **neurons** in a **neural network** that process a set of input features, or the output of those neurons.

Also, an abstraction in TensorFlow. Layers are Python functions that take **Tensors** and configuration options as input and produce other tensors as output. Once the necessary Tensors have been composed, the user can convert the result into an **Estimator** via a **model function**.

Layers API (tf.layers)

A TensorFlow API for constructing a **deep** neural network as a composition of layers. The Layers API enables you to build different types of **layers**, such as:

tf.layers.Dense for a **fully-connected layer**.

tf.layers.Conv2D for a convolutional layer.

When writing a **custom Estimator**, you compose Layers objects to define the characteristics of all the **hidden layers**.

The Layers API follows the **Keras** layers API conventions. That is, aside from a different prefix, all functions in the Layers API have the same names and signatures as their counterparts in the Keras layers API.

learning rate

A scalar used to train a model via gradient descent. During each iteration, the **gradient descent** algorithm multiplies the learning rate by the gradient. The resulting product is called the gradient step.

Learning rate is a key **hyperparameter**.

least squares regression

A linear regression model trained by minimizing **L2 Loss**.

linear regression

Using the raw output (\(y'\)) of a **linear model** as the actual prediction in a **regression model**. The goal of a regression problem is to make a real-valued prediction. For example, if the raw output (\(y'\)) of a linear model is 8.37, then the prediction is 8.37.

Contrast linear regression with **logistic regression**. Also, contrast regression with **classification**.

logistic regression

A **classification model** that uses a **sigmoid function** to convert a **linear model's** raw prediction (\(y'\)) into a value between 0 and 1. You can interpret the value between 0 and 1 in either of the following two ways:

As a probability that the example belongs to the **positive class** in a binary classification problem.

As a value to be compared against a **classification threshold**. If the value is equal to or above the classification threshold, the system classifies the example as the positive class. Conversely, if the value is below the given threshold, the system classifies the example as the **negative class**. For example, suppose the classification threshold is 0.82:

Imagine an example that produces a raw prediction (\(y'\)) of 2.6. The sigmoid of 2.6 is 0.93. Since 0.93 is greater than 0.82, the system classifies this example as the positive class.

Imagine a different example that produces a raw prediction of 1.3. The sigmoid of 1.3 is 0.79. Since 0.79 is less than 0.82, the system classifies that example as the negative class.

Although logistic regression is often used in **binary classification** problems, logistic regression can also be used in **multi-class classification** problems (where it becomes called multi-class logistic regression or multinomial regression).

logits

The vector of raw (non-normalized) predictions that a classification model generates, which is ordinarily then passed to a normalization function. If the model is solving a **multi-class classification** problem, logits typically become an input to the softmax function. The softmax function then generates a vector of (normalized) probabilities with one value for each possible class.

In addition, logits sometimes refer to the element-wise inverse of the **sigmoid function**. For more information, see tf.nn.sigmoid_cross_entropy_with_logits.

Log Loss

The **loss** function used in binary **logistic regression**.

Long Short-Term Memory (LSTM)

A type of cell in a **recurrent neural network** used to process sequences of data in applications such as handwriting recognition, machine translation, and image captioning. LSTMs address the **vanishing gradient problem** that occurs when training RNNs due to long data sequences by maintaining history in an internal memory state based on new input and context from previous cells in the RNN.

loss

A measure of how far a model's **predictions** are from its **label**. Or, to phrase it more pessimistically, a measure of how bad the model is. To determine this value, a model must define a loss function. For example, linear regression models typically use **mean squared error** for a loss function, while logistic regression models use **Log Loss**.

loss curve

A graph of **loss** as a function of training **iterations**. For example:

The loss curve can help you determine when your model is **converging**, **overfitting**, or **underfitting**.

+197 more cards