General

Question 1

What are “labels”?

Answer 1

The thing we’re predicting. e.g. the “y” variable in a simple linear regression.

Question 2

What is a “feature”?

Answer 2

An input variable, e.g. the “x” in a simple linear regression. There may be multiple input variables, indicated as {x₁, x₂, … x_N}.

Question 3

What is an “example”?

Answer 3

An instance of data, either “labeled” or “unlabeled”. Labeled examples are used to train the model, which can then predict the labels on unlabeled examples.

Question 4

What is “training” and “inference”?

Answer 4

Training is using labeled data to create, or “learn”, the model. Inference is using the model to make predictions on unlabeled data.

Question 5

What is “regression” and “classification”?

Answer 5

A regression model predicts continuous values, e.g. prices. A classification model predicts discrete values, e.g. spam or not spam, or the animal in an image.

Question 6

What is “linear regression”?

Answer 6

Linear regression is a method for finding the straight line or hyperplane that best fits a set of points

Question 7

What is “L₂ Loss”?

Answer 7

Also known as “squared error”, this is simply the sum of the “(prediction - actual)²” for the examples.

Question 8

How would you write the equation for a simple linear regression model?

Answer 8

y’ = b + w₁ x₁, where y’ is the predicted label, b is the bias (y intercept), w₁ is weight of feature 1, and x₁ is a feature (a known input). With more features, “b + w₁ x₁ + w₂ x₂…”

Question 9

What is the process of trying many example models to find one that minimizes loss called?

Answer 9

Empirical risk minimization.

Question 10

What is MSE?

Answer 10

Mean square error. The sum of the squared losses for example example, divided by the number of examples. It is commonly used, but not the only function, or best for all cases.

Question 11

What does it mean that a model has “converged”?

Answer 11

In computing the model, the loss has reach a point with no or little improvement in each iteration.

Question 12

What is the shape of the “loss” vs “weight” plot for regression problems?

Answer 12

Convex (i.e. bowl). They only have one minimum (i.e. one place with a slope is exactly 0).

Question 13

What is the symbol “Δ”?

Answer 13

The uppercase form of the Greek letter delta, it is used to represent “change in”. e.g. in the “slop intercept” equation “y = mx + b”, then “Δy = m Δx”.

Question 14

What is “∂y / ∂x”

Answer 14

The partial derivative of f with respect to x (the derivative of f considered as a function of x alone).

Question 15

What is “∇f”

Answer 15

The gradient of a function, denoted as shown, is the vector of partial derivatives with respect to all of the independent variables.

Question 16

What is a “gradient”?

Answer 16

The gradient is a multi-variable generalization of the derivative. Like the derivative, the gradient represents the slope of the tangent of the graph of the function. More precisely, the gradient points in the direction of the greatest rate of increase of the function, and its magnitude is the slope of the graph in that direction

Question 17

What is the gradient decent algorithm?

Answer 17

The gradient descent algorithm takes a step in the direction of the negative gradient in order to reduce loss as quickly as possible. To determine the next point along the loss function curve, the gradient descent algorithm adds some fraction of the gradient’s magnitude to the starting point.

Question 18

What is the “learning rate”?

Answer 18

Gradient descent algorithms multiply the gradient by a scalar known as the learning rate (also sometimes called step size) to determine the next point.

Question 19

What are hyperparameters?

Answer 19

Hyperparameters are the knobs that programmers tweak in machine learning algorithms. Most machine learning programmers spend a fair amount of time tuning the learning rate. eg. If you pick a learning rate that is too small, learning will take too long.

Question 20

In gradient descent, what is a “batch”?

Answer 20

The total number of examples you use to calculate the gradient in a single iteration.

Question 21

What is SGD and “mini-batch SGD”?

Answer 21

Stochastic Gradient Decent uses one random example for each iteration of training. This is fast but can be noisy. “mini-batch” chooses a number of random samples (typically 10 - 1000), which has less noise and is still much more efficient that “full batch”.

Question 22

What is root mean squared error? What is nice about it?

Answer 22

It is just the square root of MSE (mean squared error). A nice property of RMSE is that it can be interpreted on the same scale as the original targets

Question 23

What two sets are the data typically divided into?

Answer 23

The training set, used to build the model, and the test set, used to validate the accuracy of the model. The subsets should be consistent throughout training, and never train on the test data.

Question 24

What other subset might be added to the training and test sets, and why?

Answer 24

A validation set can be used to avoid overfitting. Tweak the hyperparameters until the validation set does well, then verify against the test set that it holds.

Question 25

What is “one-hot encoding”?

Answer 25

Take a highly variable column, like a “street name” string, and represent as a sparse vector, where all values are 0 except the index representing that street.

Question 26

What is “feature engineering”?

Answer 26

Turning raw data into a feature vector. Machine learning models typically expect examples to be represented as real-numbered vectors. The vector is constructed by deriving features for each field, then concatenating them all together.

Question 27

What is “scaling”, and when/why is it useful?

Answer 27

This puts feature values within a similar range (e.g. -1 to 1, or 0 to 1). This helps gradient decent converge quicker, avoids NaNs due to out of range operations, and avoids some features having an outsided impact.

Question 28

What is the “Z score”?

Answer 28

The number of standard deviations from the mean a value is. It is a useful way of scaling. (e.g. “z-score = (value - mean) / std dev.” ).

Question 29

What is a “feature cross”?

Answer 29

A feature cross is a “synthetic feature” formed by multiplying (crossing) two or more features

Question 30

What is the value of a “feature cross”?

Answer 30

It allows the encoding of non-linearity in a linear system.

Question 31

What are feature crosses commonly used for?

Answer 31

Combining one-hot encodings as a type of conjunction, e.g. country=use & lang=es, or lat=32 & long=45. Creating a feature crosses on 5 different lat & longs, results in a 25-element one-hot encoding for the lat & long.