Types of Optimization Algos.
https://medium.com/towards-data-science/types-of-optimization-algorithms-used-in-neural-networks-and-ways-to-optimize-gradient-95ae5d39529f
0 notes
Confusing statistical terms alpha and beta
http://www.theanalysisfactor.com/confusing-statistical-terms-1-alpha-and-beta/
0 notes
K means Clustering - 10601
Clustering:
Fully Unsupervisd Algo
Automatically partition unlabeled data points into groups of data points
Useful for
Organizin data
Understanding Hidden Structure
Appliations:
Cluster NewsArticles, Web Pages or search results by topic.
Cluster Protein sequences by func or genes according to expression profile.
Cluster user of Social network based on intrest(Commuity detection) - Ad target
Optimization Background:
Coordinate Descent:
Goal - Minimize a function J(theta)
eg. theta = argmin(J(theta))
Idea: Pick one dimension, and minimize along that dimension.
Algorithm:
1. Choosing Initial Point heta
2. Repeat until stopping criteria is reached
a. Theta1 = argmin J(theta1,theta2,….thetan)
b. Theta2 = argmin J(theta1,theta2,….thetan)
c. Theta1 = argmin J(theta1,theta2,….thetan)
.
.
d. Thetan = argmin J(theta1,theta2,….thetan)
Note: Steps abc..,d are exact line search alo ng same axis.
In some cases this algo can get stuck and start oscillating along two points. In this scenario we can use something called as Block Coordinate Descent.
Block Coordinate Descent:
Here: An eg. With two block alpha and beta where theta = [alpha, beta]
Goal: alpha,beta = argmin J(alpha,beta) where alpha in Ra and beta in Rb
Idea: Minimize over an entire group of variables at a time.
Algo:
1.) Choose alpha and beta
2.) Repeat until stopping criteria is reached
a. Alpha – argmin (J(alpha,beta))
b. Beta – argmin (J(alpha,beta))
Clustering:
Left figure - Distance between two points in same cluster is less than different cluster
Right Figure - Distance between two points in same cluster is more than different cluster
Now we define a Object function to minimize.
What is clustering? Goal is to partition unlabeled instances into group of similar points.
Input: Unlabeld data D = {X1,X2,X3….Xn}, X(i) belongs Rm
*We do not know the labels!
Output:
View#1: Labeled Instances {(X(1),Z(1)), (X(2),Z(2)), (X(3),Z(3))….. (X(n),Z(n))}
Where Z(1) E {1,2,3,….,K} k is # of clusters
View#2: Clusterings: C1,C2,…..Ck where k is number of clusters Ci = {X(i): Z(i)=j} pints in jth partition
Questions:
1.) How many clusters are there?
2.) How do we define similarity between points? Eg. Eucledian distance
Object based clusteting:
Eg K-Means Objective -
Input: D = {Xi}i=1N
Cluster Centers: {c1,c2,…,ck} = c
Decision Rule: Assign point X(i) to its nearest cluster center cj
Objective:
C= argmin sumi=1N minj in {1,..,k}||X(i)-cj||2
�- � �S�<|6[�
0 notes
AI Makes Stunning Photos From Your Drawings | Two Minute Papers
https://www.youtube.com/watch?v=u7kQ5lNfUfg&list=PLujxSBD-JXgnqDD1n-V30pKtp6Q886x7e&index=141
0 notes
Nice Vizualizations for Imp DS theoretical concepts
http://students.brown.edu/seeing-theory/
0 notes
Interesting gallery of Python Notebooks
https://github.com/jupyter/jupyter/wiki/A-gallery-of-interesting-Jupyter-Notebooks
0 notes
Exploratory Data Analysis
Graphical Representation used for EDA.(List is incomplete)
histogram
swarmplots
empirical cummulaive distribution functions
0 notes
determining if two different people are tweeting from Pres Trump’s twitter Account
http://www.datasciencecentral.com/profiles/blogs/data-science-reveals-trump-tweets-are-written-by-two-people
0 notes
What is Bayesian ML?
http://www.kdnuggets.com/2016/07/bayesian-machine-learning-explained.html
0 notes
A complete guide to data exploration
https://www.analyticsvidhya.com/blog/2016/01/guide-data-exploration/
0 notes
Steel hacks 2017, Musicon - Dynamic Song Playing based on User Activity
Recently, I along with my friends from CMU participated in Steel Hacks 2017. We build a mobile application which identifies user activity using Android Accelerator and plays song from Spotify based on activity and user preferences.
I worked on the User Activity Recognition part where I used Android’s Google Accelerator api to receive the co-ordinates of Phone Movement. Based on this we calculated the speed with which the user is moving and categorized it into different modes like: Still, Slow Walking Brisk Walking, Jogging and Sprint.
Now Based on these different modes we pulled the Users Spotify details and played the song based on the activity and his preferences.
For more details please go to the below dev post link.
https://devpost.com/software/musicon-2zy5iq
1 note
·
View note
Steel hacks 2017, Musicon - Dynamic Song Playing based on User Activity
Recently, Me along with my friends from CMU participated in Steel Hacks 2017. We build a mobile application which identifies user activity using Android Accelerator and plays song from Spotify based on activity and user preferences.
I worked on the User Activity Recognition part where I used Android’s Google Accelerator api to receive the co-ordinates of Phone Movement. Based on this we calculated the speed with which the user is moving and categorized it into different modes like: Still, Slow Walking Brisk Walking, Jogging and Sprint.
Now Based on these different modes we pulled the Users Spotify details and played the song based on the activity and his preferences.
For more details please go to the below dev post link.
https://devpost.com/software/musicon-2zy5iq
1 note
·
View note
https://youtu.be/fqxxv5Cyd4E?t=1037
Regression Analysis
Linear Reg Most powerful method in stat modelling.
Applications of Regressions
Extract and exploit the relation between dependent and indep variable.
To talk about causal inference. What factors drive the dependent variable. So one can test hypothesis about the causal relation.
Simple Approximation
Data set
Independent Variable:
Xi = (Xi1,Xi2,….Xip)T, i = 1,2,…n
Dependent Varibale:
Yi , i = 1,2,…,n
General Linear Model – For each case i, the conditional distribution is given by [yi|xi]
Here, yi = i + εi
Where yi = β1xi,1, β2xi,2,….. βpxi,p,
β1, β2,….. βp are regression parameters
εi = residual error variable (varies over all cases)
Why to change the stock prices onto Logarithmic scale.
Its better to show stock prices on percentage scale than abosulte value. Log gives a good idea about percentage change.
1 note
·
View note
Deep Learning for Multivariate Financial Time Series
https://www.diva-portal.org/smash/get/diva2:820891/FULLTEXT01.pdf?utm_content=bufferc5a37&utm_medium=social&utm_source=twitter.com&utm_campaign=buffer
0 notes
Time Complexity for Different Python functions(Big Oh notation)
Source: https://wiki.python.org/moin/TimeComplexity
0 notes
Intro to clustering
https://www.analyticsvidhya.com/blog/2016/11/an-introduction-to-clustering-and-different-methods-of-clustering/
0 notes
Basic Prob every data scientist should know
https://www.analyticsvidhya.com/blog/2017/02/basic-probability-data-science-with-examples/
0 notes