Deep learning review

DEEP LEARNING : DESIGNING
INTELLIGENT EXPERTS
Manas Gaur
Delhi Technological University

OUTLAY OF THE
PRESENTATION
 What is Computational Social Science?
 Need for Computational Social Science
 ELECTION PREDICTION : HOT FIELD for Computational Social Science
 BLENDING-Machine Learning and Computational Social Science
 What is S.L.E.P.S
 Architecture of SLEPS
 Functional Requirement of SLEPS
 Class Diagram of SLEPS
 Sequence Diagram of SLEPS
 Component Diagram of SLEPS
 Conclusion and Future Work
 References

WHAT IS DEEP LEARNING
 Deep Learning is a new area of Machine Learning research.
 The objective of Deep Learning is moving Machine Learning closer to one of
its original goals: Artificial Intelligence.
 Deep Learning is about learning multiple levels of representation and
abstraction that help to make sense of data such as images, sound, and
text.
 Different Algorithms in Deep Learning :
 Restricted Boltzmann Machine
 Deep Belief Networks
 Logistic Regression in Deep Learning
 Neural Network

SUPERVISED LEARNING
Testing:
What is ?is?

SEMI SUPERVISED
LEARNING
Testing:
What is ?is?

SELF TAUGHT LEARNING
Testing:
What is ?is?

Samples from
Full posterior
inference
Samples from
Feed-forward
Inference
(control)
Input images
Hierarchical Probabilistic Inference

SELF LEARNING
ELECTORAL
RESULTS
PREDICTION
SYSTEM (SLEPS)
DELHI TECHNOLOGICAL UNIVERSITY
COMPUTATIONAL SOCIAL SCIENCE

WHAT IS COMPUTATIONAL
SOCIAL SCIENCE
 Computational Social Science (CSS) is the science the focuses on investigating
areas affecting Election Results Prediction :
 Social Media and Social Network Analysis
 Behavioural Analysis : Psychological Analysis, Geographic Comment Analysis,
Sentiment Analysis.
 Physiological Trait Analysis and Competence Judgement based on an Event
 Response Time Analysis based on Scenario ( Image, Movie, Audio)
 Deliberation in judgement based on Gender, Place of Birth, Native Tongue and
Genealogy.
 Study of Impact of any Event whether Positive or Negative that directly or
indirectly influence the Comments of the People and that makes up HEADLINE!!.

WHAT IS CSS (CONTD.)
 Other Potential Application Areas of Computational Social Science are:
 ENVIRONMENT RECOGNITION : Adaptation to New Location, New Language
and Vocal Transformation.
 Organizational Behaviour : Code of Conduct, Managerial Hierarchy.
 Detecting Friendly, Flirtatious, Awkward and Assertive Speech in Speech
Dates.
 Social Network Research in Higher Education

NEED FOR CSS
 There is tremendous need for transformation form static social science to
more dynamic and quantitative computational social science due to
following reason
 High information content on Social web and Trading Web , example : Twitter
(#) and Wall Street Journal
 Sentiments in the comments : Through Likes, replies, image { emoticons}
 Responses to Queries, News and Judgement, Recognition Judgement
 Hits on the site, for Example comparing site access and reviews using two
case studies of TRIVAGO and IBIBO.com
 Revolutionize Organizational Behaviour based on Customers response For
example a detailed study at MIT showed that Expedia’s customer service
was improved based on the number of complaints by the customers.

ELECTION PREDICTION :
HOT APPLICATION OF CSS
 CSS incorporate foundation topic of vibrant field like Sociology, psychology,
Mathematics, History and Geography with real role played by TIME
TRAVEL!!
 During the months of elections, there is large amount of data (near about
100 TB) on the table, which if properly analysed can augur the election
result.
 We try to map data ( both numerical and Boolean ) of any presidential or
gubernatorial election using social traits to statistical learning using neural
networks, support vector machines.
 This create a new field, that combines Social Behaviours with high end
statistics, more profoundly called Machine Learning aided Computational
Social Science (ML-CSS)

ELECTION PREDICTION: HOT
APPLICATION OF CSS
 Days of Election has seen rise in Twitter Tweets, Facebook status update.
 Moreover, people pass comments ( in Favour or Against) based on
 Facial Appearances
 Standard Trait Based Analysis ( Character Sketch)
 Geographical Background
 Educational Background
 Gender Based Vote Casting
 Ethnicity based Vote Casting

BLENDING ML AND CSS 
ML-CSS
 Machine Learning : It is a field of intelligence and learning that perfectly
incorporates statistical techniques to educate a system based on clustering
and classification.
 FOOD for Machine Learning is a data set comprising of independent and
dependent attributes.
 What CSS Contains?
 Web-metrics Data { Social Networking site, Forums, Blogs, Yahoo Portal}
 News data { NDTV, AAJ TAK, BBC, CNN-IBN}
 Historical Data (Data Warehouse) { Mostly used in election prediction}
 Video Clips, Audio Clips and Reviews { by Experts, by third party analyst
etc}

POPULAR ML TECHNIQUES
IN CSS
 Neural Networks :
 Support Vector Machine:
 Fuzzy Logic:
 Association Rules:

WHAT IS SLEPS
 Self Learning Electoral Prediction System (SLEPS) is a system being
proposed to integrate Machine learning and Computational Social Science
to increase the credibility and authenticity of electoral polls.
 SLEPS is a proposed model that uses large heterogeneous dataset along
with dual neural network for efficient electoral result prediction.
 SLEPS uses heterogeneous data set stored in various data bases and feed
it into a neural network after assigning a proper set of hidden neurons.
 The first layer Neural Network is trained and output of which is submitted
to 2nd
layer neural network.
 After the neural network is trained, it is tested on real time data

SLEPS INTERNAL
 The proposed system is called self learning due to presence of flexible
table structures which expands column wise.
 SLEPS model incorporates all features that affect elections prediction
process that is
 Facial appearance with varied response time (100ms, 250ms, 1s, 2s,
Unlimited)
 Unreflective and Deliberate judgement by the voters
 Character Traits like Trustworthiness, Attractiveness, Integrity etc.
 Influence on vote share, competence judgement and electoral outcomes
(dependent variables)
 Background data of the Contestant

SLEPS DATABASE SCHEMAS
 Layer 1 Schemas
 Character_Traits {Contestant, Trustworthiness, Attractiveness,
Likeability, Authenticity, Integrity, Competence_judgement}
 All the attributes are scaled from 0-9 in which competence judgment is layer one
output ( class label)
 Participant_Response_Traits { Contestant, Binary Competent, Continuous
Judgement, Recognition Judgement, Competence Judgement}
 Binary Competent is scaled from 0-1 { yes or no}
 Continuous Judgement is scaled from 0-9. The faces of the contestant are shown
in cycle continuously.
 Recognition Judgement is scaled from 0-1. This field reflects people’s behaviour
when they look for similarity in faces of contestants.

 The Participant Response Trait is created for different response time wiz {100ms, 250ms, 1s,
2s, unlimited}
 SLEPS plot the graph between the competency point and time exposure for different
contestant.
 SLEPS calculated influence parameter taking all the subsets of the field (2^n).
 For example: 100ms exposure for Gubernatorial election, 250ms exposure to improve the
prediction by calculating the change in competence judgement
 Judgement_Types {Contestant, Unreflective Judgement, Deliberate Judgement, vote
share, Competence Judgement }
 Deliberate judgement and unreflective judgement is calculated for 3 different categories
 Careful thinking and good judgement ( unconstrained time)
 Response Deadline within 2s
 250ms replication condition

 Contestant_Traits { Contestant, Work_done, Family_BG, Category,
Ethnicity, Gender, Competence Judgement}
 Work_Done : Work done by the contestant in past mapped on scale 0-9
 Family_BG : Family Background of the contestant mapped on scale 0-9
 Category : To Which category the contestant belong on scale 0-9
 Ethnicity : To which religious class the contestant belongs (based on ethnic
classes in the country)
 Gender Based : Gender of the contestant on scale 0-1

SLEPS DATABASE SCHEMA
 Layer 2 neural network schemas
 Layer2Schema { Contestant, Vote_share, Election_outcomes}
 Contestant : it is the primary key in all the database schemas
 Vote_Share: it is count of number of votes given by the participant to the
contestant
 Election_outcomes : it is a binary field which is categorized into {winner,
runner-up}
 All the Schemas used in SLEPS are flexible but provides prediction based on
pervasive influential parameters.
 Election_Outcomes is the final output dependent parameter of SLEPS

ARCHITECTURE OF SLEPS
back propagation Can be modified based on
following ways
1. Added Association Rules
2. Can be integrated with
Fuzzy Logic
3. Change in the hidden
layer activation function
4. Following figure is for
one schemas, similar
structure has to be
followed for all the
SLEPS schemas
5. SLEPS functionality is
integrated with MYSQL
and Web Help

SLEPS IN ACTION
Dual Neural Network training in SLEPS with convergence in 300 epochs and 400 epochs

SLEPS IN ACTION
Error Histogram with 20 bins. Error = target-output
Performance based comparison between three
machine learning techniques (MLT)
Analysis of dataset with regression
and root mean square values

SLEPS IN ACTION
0
2
4
6
8
10
0
2
4
6
8
10
12
SLEPS OUTPUT
Expert 1 Expert 2 Expert 3
Expert 4 Election Outcome
Expert 1 Expert 2 Expert 3 Expert 4
Election
Outcom
e Vote Share
5 5 5 5 10.5 51597
5 6 4 6 10 910910
5 3 4 5 10.4 140870
5 4 4 5 10.3 318293
5 4 4 5 10.3 353315
5 5 3 5 9.6 172387
5 5 2 5 8.7 93864
5 5 2 5 8.7 122930
3 3 8 6 2.2 1.10E+07
5 5 4 5 10.2 463278
9 8 8 9 1.9 1.98E+07
5 5 6 5 7.336 551633
5 2 2 4 9.364 51323
5 6 4 6 10 292711
5 6 4 6 10 310854
5 2 2 4 9.36 52443
5 8 4 7 9.992 174870
5 6 6 6 10 276656
5 6 6 6 10 305110
5 4 5 5 10.1 578136
5 7 7 7 6.8 1254290
5 4 1 5 7.5 41998
4 7 5 6 6.5 1773256

CLASS DIAGRAM FOR SLEPS
Class diagram showing
One to one
Correspondence
Among the modules

SEQUENCE DIAGRAM OF
SLEPS
The sequence diagram of SLEPS
Showing the timely passes of
Message in SLEPS in each generation

COMPONENT DIAGRAM OF
SLEPS
Component Diagram of SLEPS
Represents all the executable
Modules in SLEPS

CONCLUSION AND FUTURE
WORK
 SLEPS is an efficient proposed election prediction framework based on
facial data of the contestant, background data and opinion data.
 SLEPS works towards predicting competence judgment.
 SLEPS work on dual Neural Network Model which provides election
outcomes.
 SLEPS can be build further to predict election of any country, presidential
and gubernatorial election.
 SLEPS is a self learning election prediction system which is continuous
evolving with the help of web module incorporated in it.

REFERENCES
 “A computational approach to politeness with application to social factors” Cristian Danescu-
Niculescu-Mizil, Moritz Sudhof, Dan Jurafsky, Jure Leskovec, Christopher Potts. Proceedings of ACL,
2013. Nominated for Best Paper Award.
 “Clash of the Contagions: Cooperation and Competition in Information Diffusion” by S.
Myers, J. Leskovec. IEEE International Conference On Data Mining (ICDM), 2012.
 “Classroom Ordering and the Situational Imperatives of Routine and Ritual” by David Diehl,
and Daniel A. McFarland. Sociology of Education 85, 4: 326-349. 2012.
 “Defining and evaluating network communities based on ground-truth” by Jaewon Yang and
Jure Leskovec. IEEE International Conference On Data Mining (ICDM). Brussels, 2012.
 “Detecting Friendly, Flirtatious, Awkward, and Assertive Speech in Speed-Dates” by Rajesh
Ranganath, Dan Jurafsky, and Daniel A. McFarland. Computer Speech and Language 27, 1: 89-
115. 2012.
 “Differentiating Language Usage Through Topic Models” by Daniel A. McFarland, Christopher D.
Manning, Daniel Ramage, Jason Chuang, Jeffrey Heer, and Dan Jurafsky. In press. Poetics. 118 (6),
1596-1649. 2013.

Deep learning review

More Related Content

What's hot (9)

Viewers also liked (20)

Similar to Deep learning review (20)

Recently uploaded (20)

Deep learning review