Exam 1 Flashcards
Data science
An area of investigation which includes AI and it’s components as well as statistical analysis, data analysis
Machine learning
Element that allows a computer intelligence to learn
Artificial intelligence
Computer implementation of human intelligence
Artificial neural network
Computer version of a biological version of this
Supervised Learning
Type of ML that find a model based on a dataset where the values (targets) are known, to predict those values
Unsupervised learning
Type of ML to find a model based on a dataset to determine natural classifications without guidance from known classifications
Classifier
Find a classification
Clustering
Find natural classifications for a dataset without guidance
Regression
Predict a value based on a fit to trends in the dataset
Association
Identify patterns of association between variables or items
A learning rule in machine learning is
What ML algorithms uses to learn
A decision boundary is
A point, line, plane, or hyper plane separating different classes
Gradient descents
Updates the answer in the direction along the negative gradient
Steepest descents
Chooses the best learning rate in each step to minimize the number of iterations
Choice of optimization
Trade off in number of iterations and the speed of each iteration that produces a reliable result in the shortest time with the smallest resources (memory)
The gradient gives
A good direction but not a good distance to find the minimum
The problem of diminishing gradients can be handled by
Normalizing the gradient by dividing by its L2 norm
L1 Norm of a vector
Sun or average of absolute values of the vector elements
L2 Norm of a vector
Square root of sum or average of square of values of the vector elements
Linfinity Norm of a vector
The maximum number in the vector
L0 Norm of a vector
Number of non zero elements in the vector
Feature
Defining characteristics of a given dataset that allow for optimal learning
Observation
A sample of the system that may contain several measurements
Reason to do feature scaling
Tends to make the search for the minimum more direct
K fold cross validation
Divide the data randomly into datasets and choose all combinations to train
Overfitting
Fitting noise during training which over estimates how the model will perform on test data
Underfitting
Training does not capture the desired prediction
Accuracy
Number of correct predictions as a proportion of total observations
AUC
Area under the true positive vs False positive rate
True positive rate
TP/(TP + FN)
False positive rate
FP/(FP + TN)
Linear regression
Predicts value from data trends
Logistics regression
Generative discriminator classifying using relative probabilities, assumes independent predictors
SVM
Margin perceptron with regularizer including linear or nonlinear transform
Decision tree
Based on splitting observations on feature value thresholds, a weak learner
Random forest
Ensemble learning method using decision trees with randomly chosen datasets
K nearest neighbors
Subdivision of space into classes based on majority of members within a distance from a centroid, a lazy learner
K means
Unsupervised learning dividing observations into clusters
K means with DBSCAN
Clustering based on density of points in a classification
Two types of error in ML are
Bias and variance
An example of bias would be
An assumption about filtering or collection of data
An example of variance would be
Fitting nuances in the data that may be noise
A way to reduce variance would be
Filter or smooth the predictors
A way to reduce bias would be
Remove filters and use raw data
We can generally reduce overfitting by
Increasing the number of independent observations
