Cheatt Sheett
Bias-Variancee Tradeoff

Whatt iss Bias?

Error between average model prediction and ground truth

The bias of the estimated function tells us the capacity of the underlying model to
predict the values
Whatt iss Variance?

Average variability in the model prediction for thee ggiven
iven dataset

The variance of the estimated function tells you how much the function can adjust
to the change in the dataset
High
h Bias Overly-simplified Model
Under-fitting
High error on both test and train data

h Variance
High Overly-complex Model
Over-fitting
Low error on train data and high on test
Starts modelling the noise in the input

<$YS $sR NBbS $sR

NBbSzs„$st{‚ <$YSzs„$st{‚
NBbS $sR <$YS $sR
NBbSzs„$st{‚

Minimum Error

$sR
zs„$st{‚
e„„B„
<$YSzs„$st{‚

=tI‚„TP$$t FZRSi$Y !%‚„TP$$t

_„‚P‚„„‚IS$PSR$9‚S _„‚P‚„„‚IS$PSR$9‚S
BPSIssR‚S$RSR8sjj BPSIssR‚S$RSjs„‚
Biass variancee Trade-off

Increasing bias reduces variance and vice-versa

Error = bias2 + variance +irreducible error

The best model is where the error is reduced.

Compromise between bias and variance
Source: https://www.cheatsheets.aqeel-anwar.com

Page 1 of 9
 Cheatt Sheett
Imbalanced
d Data
a in
n Classification
Blue: Label 1

Green: Label 0 Correct Predictions

Accuracy =
Total Predictions
Classifier that always predictss labell bluee yields prediction accuracy of 90%

y doesnt alwayss givee thee correctt insightt aboutt yourr trained

Accuracy d model
Accuracy:: %age correct prediction Correct prediction over total predictions One value for entire network
Precision:: Exactness of model From the detected cats, how many were Each class/label has a value
actually cats
Recall:: Completeness of model Correctly detected cats over total cats Each class/label has a value
F1
1 Score:: Combines Precision/Recall Harmonic mean of Precision and Recall Each class/label has a value

Performancee metricss associated

d with
h Classs 1
(Is your prediction correct?) (What did you predict)
Actual Labels True Negative
1 0
(Your prediction is correct) (You predicted 0)
TP FP
True False
Predicted Labels

Precision = Falsee +ve ratee =

1

Positive Positive TP + FP TN + FP

(Prec x Rec) TP + TN
F1
1 scoree = 2x Accuracy =
(Prec + Rec) TP + FN + FP + TN
False True
0

Negative Negative TN TP
Specificity
y= Recall,, Sensitivity =
TN +FP Truee +ve rate TP + FN

Possiblee solutions
1. Data
a Replication:: Replicate the available data until the Blue: Label 1
number of samples are comparable Green: Label 0
2. Syntheticc Data:: Images: Rotate, dilate, crop, add noise to Blue: Label 1
existing input images and create new data Green: Label 0
3. Modified
d Loss:: Modify the loss to reflect greater error when   
 


 
misclassifying smaller sample set
4. Changee thee algorithm:: Increase the model/algorithm complexity so that the two classes are perfectly
separable (Con: Overfitting)
Increase model
complexity

No straight line (y=ax) passing through origin can perfectly Straight line (y=ax+b) can perfectly separate data.
separate data. Bestt solution:: line y=0, predict all labels blue Green class will no longer be predicted as blue

Source: https://www.cheatsheets.aqeel-anwar.com

Page 2 of 9
 Cheatt Sheett
PCA
A Dimensionality
y Reduction
Whatt iss PCA?

Based on the dataset find a new set of orthogonal feature vectors in such a way that the
data spread is maximum in the direction of the feature vector (or dimension)

Rates the feature vector in the decreasing order of data spread (or variance)

The datapoints have maximum variance in the first feature vector, and minimum variance
in the last feature vector

The variance of the datapoints in the direction of feature vector can be termed as a
measure of information in that direction.
Steps
1. Standardize the datapoints
2. Find the covariance matrix from the given datapoints
3. Carry out eigen-value decomposition of the covariance matrix
4. Sort the eigenvalues and eigenvectors

Dimensionalityy Reductionn with

h PCA

Keep the first m out of n feature vectors rated by PCA. These m vectors will be the best m
vectors preserving the maximum information that could have been preserved with m
vectors on the given dataset
Steps:
1. Carry out steps 1-4 from above
2. Keep first m feature vectors from the sorted eigenvector matrix
3. Transform the data for the new basis (feature vectors)
4. The importance of the feature vector is proportional to the magnitude of the eigen value

Figure 1 Figure 2
Feature # 1 (F1)

Fe Feature # 1

Variance
Variance

1
# e

2
ur

e #
at

ur
at
Fe
ew

w
Ne
N

F1 F2 Feature # 2 (F2) Feature # 2 F1 F2

Figure 3 Figuree 1:: Datapoints with feature vectors as

x and y-axis
Figuree 2:: The cartesian coordinate system is
rotated to maximize the standard deviation
Variance
ew Feature # 1

along any one axis (new feature # 2)

1
#

2 Figuree 3:: Remove the feature vector with

re

#
u

ur
e minimum standard deviation of datapoints
at

at
Fe

Fe F1 F2 (new feature # 1) and project the data on

w
Ne
N

Feature # 2 new feature # 2

Source: https://www.cheatsheets.aqeel-anwar.com

Page 3 of 9
 Cheatt Sheett
Bayess Theorem
m and
d Classifier
Whatt iss Bayes Theorem?

Describes the probability of an event, based on prior knowledge of conditions that might be
related to the event.

P(A B)

How the probability of an event changes when
we have knowledge of another event Posterior
Probability
P(A) P(A B)
Usually a better
estimate than P(A)
Bayes Theorem
Example

Probability of fire P(F) = 1%

Probability of smoke P(S) = 10%
Likelihood P(A) Evidence

Prob of smoke given there is a fire P(S F) = 90%

What is the probability that there is a fire given P(B A) Prior P(B)
we see a smoke P(F ( S)?
) Probability

Maximum m Aposteriori Probability

y (MAP)) Estimation
The MAP estimate of the random variable y, given that we have observed iid (x1, x2, x3,  ), is
given by. We try to accommodate our prior knowledge when estimating.
ˆMAP y that maximizes the product of
prior and likelihood

Maximum m Likelihood
d Estimation
n (MLE)
The MAP estimate of the random variable y, given that we have observed iid (x1, x2, x3,  ), is
given by. We assume we dont have any
y prior
p knowledge
g of the quantity being estimated.
ˆ y that maximizes only the
MLE
likelihood
MLE is a special case of MAP where our prior is uniform (all values are equally likely)

Na
vee Bayes Classifierr (Instantiation

n off MAP
P ass classifier)
Suppose we have two classes, y=y1 and y=y2. Say we have more than one evidence/features (x1,
x2, x3,  ), using Bayes theorem

Bayes theorem assumes the features (x1, x2, x3,  ) are i.i.d. i.e

Source: https://www.cheatsheets.aqeel-anwar.com

Page 4 of 9
 Cheatt Sheett
Regression
n Analysis
What is Regression Analysis?
Fitting a function f(.) to datapoints yi=f(xi) under some error function. Based on the estimated
function and error, we have the following types of regression
1. Linear Regression:
Fits a line minimizing the sum of mean-squared error
for each datapoint.
2. Polynomial Regression:
Fits a polynomial of order k (k+1 unknowns) minimizing
the sum of mean-squared error for each datapoint.
3. Bayesian Regression:
For each datapoint, fits a gaussian distribution by
minimizing the mean-squared error. As the number ber of
data points xi increases, it conv
converges to point
estimates i.e.
4. Ridge Regression:
Can fit either a line, or polynomial minimizing the sum
of mean-squared error for each datapoint and the
weighted L2 norm of the function parameters beta.
5. LASSO Regression:
Can fit either a line, or polynomial minimizing the the
sum of mean-squared error for each datapoint and the
weighted L1 norm of the function parameters beta.
6. Logistic Regression (NOT regression, but classification):
Can fit either a line, or polynomial with sigmoid
activation minimizing the sum of mean-squared error for
each datapoint. The labels y are binary class labels.
Visual Representation:
Linear Regression Polynomial Regression Bayesian Linear Regression Logistic Regression
Label 1
y
y

Label 0

x x x x

Summary:
Whatt doess itt fit? Estimated
d function Error Function
Linear A line in n dimensions
Polynomial A polynomial of order k
Bayesian Linear Gaussian distribution for each point
Ridge Linear/polynomial
LASSO Linear/polynomial
Logistic Linear/polynomial with sigmoid

Source: https://www.cheatsheets.aqeel-anwar.com

Page 5 of 9
 Cheatt Sheett
Regularization
n in
n ML
Whatt iss Regularization
n in
n ML?

Regularization is an approach to address over-fitting in ML.

Overfitted model fails to generalize estimations on test data

When the underlying model to be learned is low bias/high
variance, or when we have small amount of data, the =tI‚„TP$$t FZRSi$Y !%‚„TP$$t
estimated model is prone to over-fitting. _„‚P‚„„‚IS$PSR$9‚S _„‚P‚„„‚IS$PSR$9‚S
BPSIssR‚S$RSR8sjj BPSIssR‚S$RSjs„‚

Regularization reduces the variance of the model
Types of Regularization: Figure 1. Overfitting
1. Modify the loss function:

L2 Regularization: Prevents the weights from getting too large (defined by L2 norm). Larger
p
the weights, more complex the model is,, more chances of overfitting.
g

L1 Regularization: Prevents the weights from getting too large (defined by L1 norm). Larger
the weights, more complex the model is, more chances of overfitting. L1 regularization
introduces sparsity in the weights. It forces more weights to be zero, than reducing the the
g
average magnitude of all weights

Entropy: Used for the models that output probability. Forces the probability distribution
niform distribution.
towards uniform

2. Modify data sampling:

Data augmentation: Create more data from available data by randomly cropping, dilating,
rotating, adding small amount of noise etc.

K-fold Cross-validation: Divide the data into k groups. Train on (k-1) groups and test on 1
group. Try all k possible combinations.

3. Change training approach:

Injecting noise: Add random noise to the weights when they are being learned. It pushes the
model to be relatively insensitive to small variations in the weights, hence regularization

Dropout: Generally used for neural networks. Connections between consecutive layers are
randomly dropped based on a dropout-ratio and the remaining network is trained in the
current iteration. In the next iteration, another set of random connections are dropped.
5-fold cross-validation Original Network Dropout-ratio = 30%
Test Train
Train Test Train

Train Test Train

Train Test Train

Train Test Connections = 16 Active = 11 (70%) Active = 11 (70%)

Figure 2. K-fold CV Figure 3. Drop-out

Source: https://www.cheatsheets.aqeel-anwar.com

Page 6 of 9
 Cheatt Sheett
Famouss CNNs
AlexNet  2012
Why: AlexNet was born out of the need to improve the results of
the ImageNet challenge.
What: The network consists of 5 Convolutional (CONV) layers and 3
Fully Connected (FC) layers. The activation used is the Rectified
Linear Unit (ReLU).
How: Data augmentation is carried out to reduce over-fitting, Uses
Local response localization.

VGGNet  2014
Why: VGGNet was born out of the need to reduce the # of
parameters in the CONV layers and improve on training time
What: There are multiple variants of VGGNet (VGG16, VGG19, etc.)
How: The important point to note here is that all the conv kernels are
of size 3x3 and maxpool kernels are of size 2x2 with a stride of two.

ResNet  2015
Why: Neural Networks are notorious for not being able to find a
simpler mapping when it exists. ResNet solves that.
What: There are multiple versions of ResNetXX architectures where

XX denotes the number of layers. The most used ones are ResNet50
and ResNet101. Since the vanishing gradient problem was taken care of
(more about it in the How part), CNN started to get deeper and deeper
How: ResNet architecture makes use of shortcut connections do solve
the vanishing gradient problem. The basic building block of ResNet is
a Residual block that is repeated throughout the network.
Filter
Concatenation

Weight layer
3x3 5x5
f(x) x 1x1
Conv Conv
1x1 Conv

Weight layer Conv 1x1 1x1 3x3

Conv Conv Maxpool

+ Previous
f(x)+x Layer

Figure 1 ResNet Block Figure 2 Inception Block

Inception  2014
Why: Lager kernels are preferred for more global features, on the other
hand, smaller kernels provide good results in detecting area-specific
features. For effective recognition of such a variable-sized feature, we
need kernels of different sizes. That is what Inception does.
What:: The Inception network architecture consists of several inception
modules of the following structure. Each inception module consists of
four operations in parallel, 1x1 conv layer, 3x3 conv layer, 5x5 conv
layer, max pooling
How:: Inception increases the network space from which the best
network is to be chosen via training. Each inception module can
capture salient features at different levels.

Source: https://www.cheatsheets.aqeel-anwar.com

Page 7 of 9
 Cheatt Sheett
Convolutionall Neurall Network
Convolutional Neural Network:
The data gets into the CNN through the input layer and passes
through various hidden layers before getting to the output layer.
The output of the network is compared to the actual labels in
terms of loss or error. The partial derivatives of this loss w.r.t the
trainable weights are calculated, and the weights are updated
through one of the various methods using backpropagation.

CNN Template:
Most of the commonly used hidden layers (not all) follow a
pattern
1. Layer function: Basic transforming function such as
convolutional or fully connected layer.
a. Fully Connected: Linear functions between the input and the
output.
a. Convolutional Layers: These layers are applied to 2D (3D) input feature maps. The trainable weights are a 2D (3D)
kernel/filter that moves across the input feature map, generating dot products with the overlapping region of the input
feature map.
b.Transposed Convolutional (DeConvolutional) Layer: Usually used to increase the size of the output feature map
(Upsampling) The idea behind the transposed convolutional layer is to undo (not exactly) the convolutional layer
Fully Connected Layer Convolutional Layer
w11*x
x1 1+ b1
+ b1 y1
w21*x2
x2
1
3 +b
1 *x
x3 w3

Input Node Output Node Input Map Kernel Output Map

2. Pooling: Non-trainable layer to change the size of the feature map

a. Max/Average Pooling: Decrease the spatial size of the input layer based on
selecting the maximum/average value in receptive field defined by the kernel
b. UnPooling: A non-trainable layer used to increase the spatial size of the input
layer based on placing the input pixel at a certain index in the receptive field
of the output defined by the kernel.
3. Normalization: Usually used just before the activation functions to limit the
unbounded activation from increasing the output layer values too high
a. Local Response Normalization LRN: A non-trainable layer that square-normalizes the pixel values in a feature map
within a local neighborhood.
b. Batch Normalization: A trainable approach to normalizing the data by learning scale and shift variable during training.
3. Activation: Introduce non-linearity so CNN can 5. Loss function: Quantifies how far off the CNN prediction
efficiently map non-linear complex mapping. is from the actual labels.
a. Non-parametric/Static functions: Linear, ReLU a. Regression Loss Functions: MAE, MSE, Huber loss
b. Parametric functions: ELU, tanh, sigmoid, Leaky ReLU b. Classification Loss Functions: Cross entropy, Hinge loss
MSE Loss MAE Loss Huber Loss
c. Bounded functions: tanh, sigmoid 4.0 2.0 22.00  
mse = (x − x̂)2 mae = |x − x̂| 1
2 (x − x̂)
2
: |x − x̂| < γ
3.5 1.75 1.75 γ|x − x̂| − 12 γ 2 : else
γ =1.9
=
3.0 1.5 15
1.5
2.5 1.25 1.25
2.0 1.0 1.0
1.5 0.75 0.75
1.0 0.5 0.5
0.5 0.25 0.25
0.0 0.0 0.0
-2.0 -1.0 0.0 1.0 2.0 -2.0 -1.0 0.0 1.0 2.0 -2.0 -1.0 0.0 1.0 2.0

Hinge Loss Cross Entropy Loss

1.0
3.0  
max(0, 1 − x̂) : x = 1 −ylog(p) − (1 − y)log(1 − p)
2.5
max(0, 1 + x̂) : x = −11 8.0 0.8

2.0 6.0 0.6

1.5
4.0 0.4
1.0
2.0
0.5 0.2

0.0 0.0 0.0

-2.0 -1.0 0.0 1.0 2.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

Source: https://www.cheatsheets.aqeel-anwar.com

Page 8 of 9
 Cheatt Sheett
Ensemblee Learning
g in
n ML
What is Ensemble Learning? Wisdom of the crowd
Combine multiple weak models/learners into one predictive model to reduce bias, variance and/or improve accuracy.

Types of Ensemble Learning: N number of weak learners

1.Bagging: Trains N different weak models (usually of same types
homogenous) with N non-overlapping subset of the
input dataset in parallel. In the test phase, each model is evaluated. The label with the greatest number of predictions is
selected as the prediction. Bagging methods reduces variance of the prediction

2.Boosting: Trains N different weak models (usually of same types
homogenous) with the complete dataset in a
sequential order. The datapoints wrongly classified with previous weak model is provided more weights to that they can
be classified by the next weak leaner properly. In the test phase, each model is evaluated and based on the test error of
each weak model, the prediction is weighted for voting. Boosting methods decreases the bias of the prediction.

3.Stacking: Trains N different weak models (usually of different types
heterogenous) with one of the two subsets of the
dataset in parallel. Once the weak learners are trained, they are used to trained a meta learner to combine their
predictions and carry out final prediction using the other subset. In test phase, each model predicts its label, these set of
labels are fed to the meta learner which generates the final prediction.

The block diagrams, and comparison table for each of these three methods can be seen below.
Ensemble Method
Boosting Ensemble Method
Bagging
Input Dataset Step #1 Input Dataset
Step #1 Create N subsets
Assign equal weights Complete dataset from original Subset #1 Subset #2 Subset #3 Subset #4
to all the datapoints dataset, one for each
in the dataset weak model

Uniform weights
Step #2
Train each weak
Weak Model Weak Model Weak Model Weak Model
Step #2a Step #2b model with an
Train a weak model Train Weak
Based on the final error on the independent #1 #2 #3 #4
with equal weights to trained weak model, calculate a subset, in
Model #1 parallel
all the datapoints scalar alpha.

Use alpha to increase the weights of
wrongly classified points, and
decrease the weights of correctly
alpha1 Adjusted weights classified points
Step #3
In the test phase, predict from
each weak model and vote their Voting
Step #3b predictions to get final prediction
Step #3a Train Weak
Based on the final error on the
Train a weak model Model #2 trained weak model, calculate a
with adjusted weights scalar alpha.
on all the datapoints
Use alpha to increase the weights of
in the dataset wrongly classified points, and Final Prediction
decrease the weights of correctly
alpha2 Adjusted weights classified points

Train Weak Ensemble Method
Stacking

Model #3
Step #1
Create 2 subsets from Input Dataset
original dataset, one
for training weak Subset #1
Weak Learners Subset #3#2
Meta Learner
Subset
alpha3 Adjusted weights models and one for
meta-model

Train Weak
Step #(n+1)a Model #4 Step #2
Train a weak model Train each weak
with adjusted weights model with the
Train Weak Train Weak Train Weak Train Weak
on all the datapoints weak learner Model #1 Model #2 Model #3 Model #4
in the dataset dataset
alpha3

x x x x Input Dataset
Subset #1
Weak Learners Subset #2
Meta Learner
Step #n+2
In the test phase, predict from each
weak model and vote their predictions
weighted by the corresponding alpha to
get final prediction Step #3
Voting Train a meta-
learner for which Trained Weak Trained Weak Trained Weak Trained Weak
the input is the
outputs of the Model Model Model Model
weak models for #1 #2 #3 #4
the Meta Learner
dataset
Final Prediction

Parameter Bagging Boosting Stacking

Meta Model
Focuses on Reducing variance Reducing bias Improving accuracy
Nature of weak
Homogenous Homogenous Heterogenous Step #4
learners is In the test phase, feed the input to the
weak models, collect the output and feed
Weak learners are Learned voting it to the meta model. The output of the
Final Prediction
Simple voting Weighted voting meta model is the final prediction
aggregated by (meta-learner)

Source: https://www.cheatsheets.aqeel-anwar.com

Page 9 of 9