Machine Learning Terms And Code Understanding

Front 1

Labels

Back 1

What the computer code is trying to predict (eg if an email is spam or not, or if a set of particles is a jet or not)

Front 2

Features

Back 2

Input variables to predict label, eg. Words in an email

Front 3

Example

Back 3

Piece of data,

Labeled has a label and feature

Unlabled has feature only

Front 4

Model

Back 4

Thing that does the predicting, ie maps feature, x, to a predicted label y'

Front 5

Regression model

Classification model

Back 5

Regression: predicts continuous variables

Classification: predicts discrete values

Front 6

Emperical risk minimization

Back 6

Training a model to reduce loss

Front 7

Gradient decent algorithm

Back 7

Pick a random w1 value. Calculate gradient of loss vs w graph at that point. Pick a new point that moves as a fraction of the gradient vector.

Front 8

Learning rate and optimum learning rate

Back 8

Learning rate: contant to multiply the gradient to get a new point

Optimum

1D = is 1/f"(x)

2+D = inverse of Hessian (matrix of 2nd Partial Derivatives)

! For general convex functions, story is more complex

Front 9

All module imports for creating (basic) machine learning model with tensor flow

Back 9

Front 10

Stochastic gradient decent (SGD)

Mini batch gradient decent (mini-batch SGD)

Back 10

SGD = Batch size of 1

mbSGD = typically 10 to 1000 examples chosen at random

Front 11

Building a model with tensorflow, all steps

Back 11

Front 12

Step 1: define features and configure feature columns

Back 12

Front 13

Step 2: define the target

Back 13

Front 14

Step 3: configure linear regressor

Back 14

Front 15

Step 4

Back 15

Front 16

Input function

Back 16

Front 17

Step 5 (actual training function with tensor flow)

Back 17

Front 18

Step 6 evaluation and prediction

Back 18

Front 19

One-hot encoding

Multi-hot encoding

Back 19

One hot

Make a feature of strings into a vector where each dimension represents a possible string (oov- out of vocabulary- is every string not accounted for, eg "hdkehijrla" for a road name)

Each vector is a 1 or 0 (bool) for if the feature is equal to that.

Multi-hot

One-hot but many dimensions have a 1.

Front 20

Sparse representation

Back 20

Only consider and store non zero values in one-hot or multi-hot vectors (so if feature is street name and there are a million streets, the vector doesn't have to have a million dimensions)

Front 21

Z score

Back 21

Scaling using stdevs from the mean

Scaledvalue = (val - mean) / stdev

Usually between -3 and 3

Front 22

Code to create buckets from lattitude using pandas

Back 22