Computer Vision License Validation

Business problem: Id verification system(valid or invalid) say driving license as id. How do we go about solving this business problem using Deep learning

Input - License Id Images

Approach

• Feature Definition
• Defining Elements
• Historical data
• Labeling / Annotation

Vision

• Problem #1 - Extract Face images
• Problem #2 - OCR, License Id, Dates, LicenseNumber, Authority
• Problem #3 - Detection for Signature Extracting
• Data Validation - Blurriness - Image Sharpening / Laplacian / Sobel / Canny edge to sharpen images. Non-readable - Far / Validation - Near View

Backend Validation

• API call
• Face Match
• Similarity Score
• Output - Valid License

Keep Exploring!!!

Applied Machine Learning Notes

Supervised Learning

• Classification (Discrete Labels)
• Regression (Output is continuous, Example - Age, Stock prices)
• Past data + Past Outputs used

Unsupervised Learning

• Dimensionality reduction (Data in higher dimensions, Remove dimension without losing lot of information)
• Reducing dimensionality makes it easy for computation (Continuous values)
• Clustering (Discrete labels)
• No Past outputs, Only current data

Reinforcement Learning

• All Game Playing is unsupervised
• Learning Policy
• Negative / Positive reward for each step

Type of Models

• Inductive (Learn model, Learn from a function) vs Transductive (Lazy learning ex- Opinion from like minded people)
• Online (Learn from every new incoming tweet) vs Offline (Look past 1 Yeat tweet)
• Generative (Apply Gaussian on Data, Use ML and compute Mean / Variance) vs Discriminative (Two sides of Line)
• Parametric vs Non-Parametric Models

Happy Learning!!!

DBMS Session Three

Relational Model Classification

Key - Subset of attribute
Super Key -Sufficient to identify tuple uniquely
Considerations for primary key

• not null values
• Few attributes
• Key often used in data access clauses

Relational Operators

• Selection (With Filters Applied)
• Projection (Select with explicitly specified columns)
• Cartesian product
• Union
• Difference
• Intersection
• Join

SQL Refreshers

• SELECT, WHERE, Aggregate (Group by, Having), JOINs, String operations
• SET Operations, Handling NULL values, Subqueries
• DELETION, No EXISTS, Conditional Updates
• Views, Materialized views
• Authentication & Authorization (Roles & Permissions)

Happy Learning!!!

Data Models

Captured Notes from Session #2 - Data Models

Hierarchical Data Models

• Tree like structures
• Used in Windows Registry
• Frequent Use (IMS)
• DL / 1 Programming language for IMS
• Difficult to reorganize

Graph / Network Model

• Organize collection of records in form of directed graph
• 3 way relationships can't be maintained

ER Model

• Defined in terms of Entity, Relationships
• Never Caught on Physical Model

Object Oriented Database model

• Difficult mapping programming objects to database objects

Relational Model

• Better physical data independence
• Better logical independence
• Won because of linear algebra

Happy Learning!!!

t-test and z-test

Problems to workout (Good Compiled List)

References
Link1
Link2

Chi Square Test for Independence

• Uses a cross classification table to examine the nature of the relationship between these variables
• Tables are sometimes referred to as contingency tables
• Determine variables are dependent on each other or not

Approach

• H0: chi square test for independence is conducted by assuming that there is no relationship between the two variables
• Ha: alternative hypothesis is that there is some relationship between the variables

The general formula for the degrees of freedom is the number of rows minus one, times the number of columns minus 1.

In terms of independence and dependence these hypotheses could be stated

• H0 : X and Y are independent
• H1 : X and Y are dependent

Expected Frequency = ((row total)*(column total))/Total Population

I liked the example provided in link  

Problem - Test for a Relationship between Sex and Class

X (Sex)

Y (Social Class) Male(M) Female(F) Total

Upper Middle (A) 33 29 62

Middle (B) 153 181             334

Working (C) 103 81 184

Lower (D) 16 14 30

Total 305 305            610

Table 10.12: Social Class Cross Classified by Sex of Respondents

Expected Frequency = ((row total)*(column total))/Total Population

1-pchisq(4.8748,df=3)

 0.1811978

Significance is greater than or equal to 0.05, you don't reject the null hypothesis

Results match with the problem although approach is different. The sum total sum is 610 (Total Sum)

Happy Learning!!!

Stats - Chi-Square Goodness of Fit Test

Purpose -  Test association of variables in two-way tables

The chi-square test is defined for the hypothesis:
H0: The data follow a specified distribution
Ha: The data do not follow the specified distribution
This means that if the significance value is less than 0.05, you reject the null hypothesis; if significance is greater than or equal to 0.05, you don't reject the null hypothesis

Formula is

I liked the example mentioned in notes

Problem - Testing an octadedral die to see if it is biased

Score 1 2 3 4 5 6 7 8

Frequency 7 10 11 9 12 10 14 7 (Observed)

Degree of Freedom = Number of entries - 1. Here is is 8-1 = 7

Test the hypothesis H0 - The Die is Fair

H1: Die is not fair

Significance level alpha = 0.005

Expected frequency is uniform distribution of Ei = Sum of all observed scores / 8(Number of items)

= 80/8 = 10

The expected values will be

Score 1 2 3 4 5 6 7 8

Frequency 10 10 10 10 10 10 10 10 (Expected)

To compute the score we need to find values of (Oi-Ei ), ((Oi-Ei )*(Oi-Ei ))/ Ei

For each element between  both the arrays

Compute chisquare value (R Command)

1-pchisq(4,df=7)

0.7797774

This is above significance level > 0.05. So we cannot reject null hypothesis

Answer - The Die is Fair

Happy Learning!!!

K Means Clustering

I'm slowly moving in Stats with a lot of learning. This post is from my class notes

K-means clustering

• Finding groups of object similar to one another
• Partitioning cluster approach
• Mean moves every time (Within first few iterations it will converge)
• Classify a given data set through a certain number of clusters
• This does not fit well for Sparse / Dense clusters

Great 5 Minute Video



Step 1 - "Figure out centric of region"
Step 2 - "Select K Data points randomly"
Step 3 - "Assign each data point to nearest centre"
Step 4 - "Recalculate the new centroids"
Step 5 - "Repeat Step 3,4"

More Reads - K-Means Clustering

DTW  - Dynamic Time Warping Algorithm. DTW - allowing similar shapes to match even if they are out of phase in the time axis

Ref - Link

Happy Learning!!!

5 minute quick learning - Naïve Bayes

Good 5 Minute Learning!!!

Conditional Probability - Easy Walkthrough

I take iterations to understand / try out a concept. Going back and learning after sometime it's interesting. This post is a quick explanation on Conditional Probability.

P(A) - Probability of Event A to occur
P(A/B) - Probability of A given that B has already occurred. - This we refer as conditional probability

Problem - Roll a fair die. Let A be event of odd outcomes. B be event where outcome <=3. What is probability of A and Probability of A given B has already occurred

A = Odd Outcomes = {1,3,5} = 3
B = Outcome <=3 = {1,2,3} = 3
Sample space = {1,2,3,4,5,6} = 6

Probability P(A) = |A| / |S| = 3/6 = 1/2

Probability P(A/B) = Probability of A given that B has already occurred

From B outcomes {1,2,3}, Possible A values are = {1,3}

p(A/B) = Events{1,3} / Events of B{1,3,5}
p(A/B)  = 2/3

Bayes Theorem

P(CD) = P(C/D)P(D)

P(CD) = P(D/C)P(C)

Equating both the formulas

P(D/C)P(C) = P(C/D)P(D)

P(D/C) = (P(C/D)P(D)) / P(C)

P(C/D) = ((P(D/C)P(C)) / P(D)

Happy 2 Minute Quick Learning!!!

R Working on Normal Distributions & Binomial Distributions

Cumulative density function - CDF is summation of all the probabilities within a range. The CDF is the integral of the PDF (Probability density function).

Read Jack Lion Heart's answer to What is the difference between a probability density function and a cumulative distribution function? on Quora

• “d” - returns the height of the probability density function
• “p” - returns the cumulative density function
• “q” - returns the inverse cumulative density function (quantiles)
• “r” - returns randomly generated numbers

Reference - Link

R Examples

pnorm 

pnorm(700,500,100)

Mean - 500

Variance - 100

This score 700 is better than 97% of other scores

v <- c(-1,1,2)

dnorm(v)

plot(v,dnorm(v))

plot(v,pnorm(v))

dnorm(0,mean(v),sd(v))

Link Ref

Normal distribution is defined by the following probability density function, where μ is the population mean and σ2 is the variance.

Chi-squared Distribution - If X1,X2,…,Xm are m independent random variables having the standard normal distribution, then the following quantity follows a Chi-Squared distribution with m degrees of freedom

Binormial Distribution - binomial distribution is a discrete probability distribution. It describes the outcome of n independent trials in an experiment. 

Two Extra Parameters - number of trials and the probability of success for a single trial

Distribution function

x <- seq(0,50,by=1)

y <- dbinom(x,50,0.2)

plot(x,y)

50 - Number of Trials

0.2 - Probability of success for each trial

Cumulative Probability Density Function

x <- seq(0,50,by=1)

y <- pbinom(x,50,0.5)

plot(x,y)

50 - Number of Trials

0.5 - Probability of success for each trial

Random Probability Density Function

x <- seq(0,50,by=1)

y <- rbinom(x,50,0.5)

plot(x,y)

Happy Learning, More Learning Needed...It's vast....Lot more efforts needed :)

Maths - Basics

Good Read
Link1

Happy Learning!!!

Class Notes - Information Retrieval

Information Retrieval Notes

• Document Corpus (Collection of links / indexed web structures)

Examples of Information Retrieval

• When user is humming songs, based on it if we identify song then its classified as IR problem
• Multimedia IR (musics, video, analysing music videos)
• Photo Search (Visual IR)

Information Retrieval Application Areas

• Text Information Retrieval
• Web Search
• Social Media Search
• Micro Blogs
• Twitter Blogs

Boolean Information Retrieval

• Simplest model
• Restricted Queries
• queries are boolean expressions

Inverted Index

• For each item we have a list
• Like index of a book (Topic, Pagenumber), Close to glossary
• Document, Tokenize the text
• Inside document order tokens, heuristics to combine multiple tokens for index construction
• Document Frequency - How many times term appears

Challenges

• Ordering (Right to Left)
• Proximity Search Leveraging the context)
• Encoding 
• Normalization and keyword detection based on locale
• Accents, patterns
• Stemming (Chopping end of words to obtain root word)
• Porter algorithm for stemming
• Stopwords, normalization, tokenization, lower casing, stemming, non-latin alphabets, compounds, numbers
• Skip pointers(Find elements common between both lists, increment until they match)

Happy Learning!!!

Central Limit Theorem

Normal Distribution



Standard Normal Distribution - Mean = 0, Variance = 1

Distribution approaches to normal distribution for larger set of variables. 

"As n increases, the distribution of sample mean approaches normal distribution"

Central Limit Theorem - Almost all measurable "random" variables in real world follow some kind of normal distribution.

Good Link

"Sampling distribution of the sampling means approaches a normal distribution as the sample size gets larger"

"Average of your sample means will be the population mean"




Happy Learning!!!

Linear Algebra Playlist

Linear Algebra and Calculus basics Playlist bookmarked based on reference from my colleague

Pauls Online Calculus Notes

MIT Linear Algebra Playlist

Linear Algebra


Calculus

More Reads
Link1
Link2
LA in 4 Pages
LA Machine Learning
Link3
Link4
Link5

Happy Learning!!!

Linear Regression Basics

To Understand Linear Regression basics of Slope, Correlation was useful.

Slope Revision

Finding the slope of a line from its graph: Slope of a line


Slope = Change in Y / Change in X
Slope is constant for a line

Simple Linear Regression

• One Explanatory Variable Simple Regression
• More than one Explanatory Variable multiple Regression
• X - Independant Variable(Explanatory), Y - Dependant Variable (Response)
• Good fitting line (Reasonable for predicting relationship) - Measure of Strength of Relationship (Co-relation)
• Correlation - A & B are observed at Same time
• Methods of Least Squares to estimate B0 and B1
• Residual = Observed - Predicted value (Above Line +ve, Below Line -ve)

Reference Videos

Squared error of regression line: Introduction to the idea that one can find a line that minimizes the squared distances to the points


Happy Learning!!!

Covariance and Correlation - Random Variable - Probability

This video was useful to understand the covariance and correlation relationship

Happy Learning!!!

Working with R - InterQuartile Range

Concept - IQR - InterQuartile Range

IQR = Q3 - Q1 = 3rd Quartile - 1st Quartile

• Median - Arrange data from lowest to highest
• On Even dataset - Average of two most middle numbers
• On Odd dataset - Single Number that is halfway into the set

Dataset - 5,6,12,13,15,18,22,50

Q2 = (13+15)/2 = 14 - Median of Data Value

Q1 = (6+12)/2 = 9 - Median Before Q2

Q3 = (18+22)/2 = 20 - Median After Q2

IQR = Q3-Q1 = 20-9 = 11

BoxPlot is used to identify outliers

For Above Dataset

• Minimum Value - 5
• Q1 - 9
• Q2 - 14
• Q3 - 20
• Maximum Value - 50

This is the mathematical concept. This is used for finding outliers.

Outlier - Much larger or smaller than other values in data set. IQR obtained by subtracting third vs first quartile. 

Finding Outliers

1. Any value < Q1-1.5(IQR) or > Q3+1.5(IQR) is an outlier

2. Any Value < (9-1.5(11)) = -7.5

    Any Value > 20+1.5(11) = 20+16.5 = 36.5

This Video was useful to understand the concept before trying out in R

Computing using R

IQR between 25th percentile and the 75th

dataset <-c( 5,6,12,13,15,18,22,50 )
quantile(x=dataset, probs= c(.25,.75))
IQR(x=dataset)
boxplot(dataset)

Sample Output

Outlier highlighted in circles

Happy Learning!!!

Class 3 - Statistics Notes

This was mostly on probability distribution functions. Couple of one liners from session

Conditional Probability Distribution - Value of one random variable not impacting another then random variables are independent

Variance - How many values < mean and > mean

Covariance - If X and Y are two independent variables Covariance is zero
Correlation Values between -1 and 1
Correlation

Probability "Chebyshev Inequality"
Chebyshev Inequality - For computing mean for specific region (Integrated over smaller region). Probability in tail for any random variable.

Related Reads

• Link1
• Jointly distributed random variables

Happy Learning!!!

Video Analytics Class 2

My Notes

• Linear Filter - Linear combination of neighbours
• Box filter - All values constant [1's]
• Corr-relation - Masked and Moved across Image
• Gradient  - due to surface normal discontinuity, depth discontinuity, illumination discontinuity
• LOG - Laplace of Gaussian. LOG capable of finding edges
• Salt and Pepper Image - Image has random black and white 

Basics

• Represent Image as a Matrix
• Represent Image as a function
• Point, local operations, histogram equalization, moving average model
• Cross Correlation g = H X F
• Gaussian filter (Removes High frequency, blurring, smoothens image)
• Symmetric Matrix (When you shift rows into columns it would appear the same ( aij = aji, for all indices i and j) example link 

Convolution Basics

Programmatic Walkthru - link

From link 

From link

From Link
FFT

Convolution Applications

• Smoother image
• Gaussian (Point spread function)
• Different Kinds of filter (Box, Gaussian filter)

Cross Correlation - Assess how similar are two different functions. Compares position by position. 

Correlation Walkthru

From Link 

Mathematics concepts to learn

• Vector Product
• Eigen Value decomposition
• First Derivative, Second Derivative

Vertical and horizontal edge detection filters - Sobel, Roberts, Prewitt (Veritical, Horizontal, Diagonal edge detection filters).

Good Read Link
MIT Course Slides link

Related Reads







Happy Learning!!!