CSE 432 Midterm

machine learning is the study of _________ that improve their __________ at some ________ with ___________

algorithms; performance; task; experience

well-defined learning task: < ____ >

P, T, E

machine learning is good at recognizing ________, recognizing _________, and _________

patterns; anomalies; prediction

deep learning is a type of __________ ________ _________

artificial neural network

more than 2 hidden layers makes it a _____ _______ _________ (____)

deep neural network (DNN)

the most popular machine learning algorithm

deep learning

the ______ in Python is important because it indicates a block of code

indentation

comments in python use ___; block comments use three ___ or ___

#; ‘; “

variables in python are _____ _________ and must start with a ______ or the ________ character, no _________

case sensitive; letter; underscore; numbers

boolean in Python is declared as _______

bool

Python for loop syntax through myList

for x in myList:

Python for loop syntax for range of 10

for x in range(10):

used in Python to store data values in key:value pairs; they are _______ and do not allow ________

dictionaries; ordered; duplicates

Python function definition

def myFunction(input):

what do you add to you parameter for an arbitrary number of arguments?

*

declare arr [1 2 3 4] as an numpy array

arr = np.array([1, 2, 3, 4])

declare 2×2 matrix (mat) as numpy array

mat = np.array([1,1],[2,2])

check the dimension of numpy array (arr) TWO WAYS

arr.ndim; arr.shape

comprehensive library for creating static, animated, and interactive visualizations in Python

matplotlib

a vector is a ____ _____

1D array

matrix transpose is an operator that _____ the matrix over its _______, in turn switching the _____ and ______

flips; diagonal; rows; columns

v = [a,b]
f(v) = a² + b²
what is f’(v) with respect to v?

f’(v) = [2a, 2b]

for derivatives with a matrix or vector, we normally multiply the ________ and the ____ ________

transpose; one vector

python code to find magnitude of vector x

y = x**2
s = np.sum(y)
d = np.sqrt(s)

python add/subtract vectors x and y

x + y; x - y

numpy dot product for x and y

np.dot(x,y)

matplotlib plot function for x, y

plt.plot(x, y, label='My Plot’, linewidth=2.0)

KNN is ___ __________ _________

non parameter learning

non-parameter learning y = _____
parameter learning y = ____

f(X, X_train); f(X,W)

non-parameter learning needs the _____ ________ ________ and is very slow in ______ with almost no ______ ________

entire training dataset; inferring; training process

similar to using a dictionary to find definitions or synonyms

non-parameter learning

parameter learning requires the ____, is very _____ in _______, but takes more ______ in _______

weight; fast; inferring; time; training

similar to having the word in your brain to recognize it at once

parameter learning

gives you the ground truth

loss function

common loss function

Loss(y, y^) = sum(y-y^)²

with different combinations of theta0 and theta1, we obtain different ______ ______, it is a ____ surface

loss values; 3D

loss value shows how close your _________ __________ ___________ is to the ________ ________

machine learning algorithm; ground truth

for a loss value, the _______ the ________

lower; better

machine learning aims to find the best ________ that ____ ________ could obtain the _______ value

parameters; loss function; lowest

how do we get the smallest loss value

gradient descent

each step of gradient descent uses all of the training examples - this is known as …

batch gradient descent

your step size in gradient descent is known as the _________ _____

learning rate

output is decrete in _________

classification

output is continuous in _________

regression

machine learning is a ____-______ approach

data driven

data: any __________ fact, value, text, sound, or picture not being _______ and __________

unprocessed; interpreted; analyzed

a set of data collected for machine learning based task

dataset

a set of data used to discover predictive relationships

training dataset

a set of data used to asses the strength and utility of a predictive relationship

test dataset

the attributes to each data sample

features

KNN stands for:

k nearest neighbors

for KNN:

1. calculate the ____ _________ for every _____ _______

2. select the ____ data points with the _________ _________

3. ________ based on the k point (new data point should belong to same category as the _______ )

L-2 distance; data point; K; smallest distance; voting; majority

can you still use KNN if there is more than one feature for distance calculations?

yes

when setting up KNN, you can choose two parameters:

1. the best ______ of ___ for ________

2. the best _______ for ________

value; k; voting; distance; measuring

the parameters you set of KNN are known as _____________ and are not ________ by the machine learning _______ itself

hyperparameters; adapted; algorithm

a set of examples used to tune the hyperparameters

validation dataset

never use _____ data to _____ _______

test; train model

cross validation: when dataset is ______, ______ data, try each fold as _______ and _______

small; split; validation; average

cross validation is _________ in deep learning

uncommon

learning from labeled examples

supervised learning

draw from inferences from datasets consisting of input data without labeled responses

unsupervised learning

supervised learning has pairs with an ______ object and a desired ______ value

input; ouput

unsupervised learning finds ______ ________ or _________ in data

hidden patterns; grouping

K-Means Algorithm:

1. initialize ____ _______ _______

2. assign _____ ______ to ________ clusters

3. update _______ _________ by calculating _________

4. repeat ___ and ___ until _________

5. select optimal number of ________

K center centroids; data points; nearest; center centroids; average; 2; 3; convergence; clusters

non-parameter learning requires computation of all of the _______ ________, taking more ______ and ________

training dataset; time; memory

non-parameter/parameter, supervised/unsupervised
KNN:
K-Means:
Linear Regression:

non-parameter, supervised; non-parameter, unsupervised; parameter, supervised

KNN and K-Means are __________ tasks whereas linear regression is a _________ task

classification; regression

linear regression steps

propose model; gradient descent; get parameters and test

image recognition is _______; stock price prediction is ________

classification; regression

softmax classifier: build upon ________ _________; _____ score of class k to __________ of being in this class; __________ of being in different classes sum up to ____

linear classification; map; probability; probabilities; 1

loss over the dataset is the _________ ______ for all _________

average loss; examples

three loss functions

MAE; MSE; Cross Entropy

MAE: ______ ________ __________
Equation:

mean absolution error; abs(y^ - y)

MSE: ______ _________ _________
Equation:

mean square error; (y^ - y)²

Cross Entropy is the _________ _____ likelihood of the __________ ________ as the loss

negative log; correct class

cross entropy for the following:
true label: [1 0 0 0 0]
softmax: [0.1 0.5 0.1 0.1 0.2]

-(1*log(0.1))+(0*log(0.5))+(0*log(0.1))+(0*log(0.1))+(0*log(0.2))

Regularization:
- it is likely different ___ has the same _____
- regularization helps to _______ ________ and avoid _________

W; loss; express preference; overfitting

L(W) including regularization

L(W) = data loss + regularization

overfitting: model tries to fit not only the __________ relation between _____ and ______ but also the _______ ________; ________ ______________ helps select simple models

regular; inputs; outputs; sampling errors; weight regularization

numerical gradient: __________, ______, easy to _____
analytic gradient: ______, _______, _______ prone
—> in practice we use _______ but check with _________

approximate; slow; write; exact; fast; error; analytic; numerical

with backpropogation, given f(x, y, z), you’ll end up getting which derivatives

df/dx; df/dy; df/dz

in backpropogation, multiply the _________ by the ______ ___________

upstream; local gradient

tool used for forward and back propogation

computational graph

the local gradient is the _________

derivative

the input to the local gradient is found from __________-____________

forward-propogation

current gradient =

local gradient * upstream gradient

what do we assume to begin back propogation if forward not clear

2

the input layer for a neural network

the first layer

the output layer for a neural network

the last layerl

layers in between input and output layers of neural networks

hidden layers

neurons between ________ layers are typically connected, neurons _______ the ______ layer are not connected

adjacent; within; same

the input layer of a neural network is _________ meaning the ________ is the input

transparent; output

two parts in the neurons of hidden layer

accumulation of product; activation function

connection among neurons has a _______ and it is the parameter that should be ________

weight; learned

which layer has a “special” activation function

output

squashes numbers [0, 1] and is popular, used in RNN

sigmoid

squashes numbers [-1,1] and zero-centered, used in RNN

tanh

squashes number [0, infinity] and does not saturate, used in CNN and FCN

RELU

squashes numbers [-infinity, infinity], does not saturate

leaky RELU

for choosing activation functions, typically choose _____, then _____ _______ if that doesn’t work; sometimes _____, not normally _______

RELU; leaky RELU; tanh; sigmoid