Sept 2017: I will be joining the Human-Centric AI
Initiative at A*STAR, Singapore in October 2017.
I am a Research Engineer at Institute of High Performance Computing,
Singapore under A*STAR. I am part of a new initiative that focuses
on Human-Centric Artificial Intelligence where I am advised by Dr.
Kenneth Kwok and Dr. Ho Seng
I closely work with Dr. Soujanya Poria on
multimodal sentiment and emotion recognition. I am also a member of the SenticNet
team at NTU, Singapore lead by
Prof. Erik Cambria. I have varied research
interests leading to the quest of solving intelligence from applied machine
learning to AI reasoning. Specifically, I find working on imparting human
capabilities of understanding language to machines interesting. Hence have
been involved in many research projects at the nexus of Machine
Learning and Natural
Language Processing, which are taking its shape in my labs (read
I received my B.Eng in Computer Science & Engineering from Bangalore
Institute of Technology, India. I have
been fortunate enough to work with Prof. Chng Eng Siong at
Nanyang Technological University, Singapore. He Leads the
Speech and Language Technologies Group @ Emerging Research Laboratories.
During Summer 2016, I was at Microsoft with the Universal
Store Team (Bing Ads) on their advertising platform as Data
In my spare time, I explore the uncharted territories related to Computer
Science, more specifically Natural Language Processing.
Apart from these, I am a Pianist and former classical singer.
Ping me if you want to talk about Programming,
Research, Music or anything under the roof. Twitter:@gangeshwark
Institute of High Performance Computing, A*STAR
Research Entities (A*AI Initiative - CHEEM)
October 2017 - Present
On Causal Learning. (More details will be updated soon)
Temasek Laboratories @ Nanyang Technological
January 2017 - April 2017
Chng Eng Siong, Lead of Speech and Language Technologies @
Emerging Research Labs, NTU.
Built and deployed an end-to-end system to search for an audio
in an audio (Query By Example: Spoken Term Detection). Improved
the accuracy of the system by 5% by using the SPRING-DTW
Algorithm. More details to be added.
Investigating the use of Encoder-Decoder(seq2seq) model and
potential improvements to fit the problem of sentence boundary
detection in unpunctuated text. More details to be added.
Developed 3 mobile apps and 6 Websites/Web apps – most of which
were for start-ups.
One of the apps was to help mentally challenged children to
perform Yoga with steps and timely updates & reminders.
Udacity Code Reviewer
July 2015 - July 2017
I am one of the very first Udacity Code Reviewers, where I review
student project submissions. For each submission, I do both a
thorough code review and project evaluation. I strive to give
actionable and helpful feedback to students while also improving
my code reviewing skills at the same time.
Google Android Facilitator - Club Head
Bangalore Institute of Technology
February 2015 - July 2017
Founded Android Club to develop a rich learning ambience at
Bangalore Institute of Technology for Android Enthusiasts and
Developers under the Mentorship and Sponsorship by Google
Club name: Androzign
Bangalore Institute of Technology
February 2015 - July 2017
I mentor knowledge-hungry students in my college who are learning
build Android Apps. I suggest the complete workflow and
design-flow of the app
that they want to build, right from the backend server to the
libraries. I have mentored 10+ teams with their projects.
Bachelor of Engineering (B.E.) : Bangalore Institute Of
Computer Science & Engineering
Aug 2013 - July 2017
1. President - BIT Google Students' Club.
2. Founding President - Android Club, to develop a rich learning
Bangalore Institute of Technology for Android Enthusiasts
Developers. Club name: Androzign.
3. Conducted zero-budget workshops on Java and Android.
4. Talk on Artificial Intelligence & Machine Learning
First Workshop and Grand Challenge on Computational
Modeling of Human Multimodal Language at ACL 2018.
Special issue of IEEE Computational Intelligence
Magazine (CIM) on Computational Intelligence for
Affective Computing and Sentiment Analysis 2018
17th International Semantic Web Conference (ISWC),
2018 - Research Track.
19th International Conference on Computational
Linguistics and Intelligent Text Processing, 2018
Major Projects (Total 10) - TO BE UPDATED
Query by Example: Spoken Term Detection
Jan 2017 - Apr 2017
Gangeshwar, Tung, Dr. Chng Eng Siong
In this work, we developed an end-to-end system for Spoken
Term Detection in Audio. Query by Example is a task of searching
for query audio within long corpus audio. The key aspect of
QbE-STD is to enable multi-lingual and multi-speaker audio
search. Due to variations in speaking rates, a non-linear
fluctuation occurs in speech pattern versus time axis. To
overcome this challenge we implemented a Dynamic Time Warping
based algorithm called SPRING which is suitable to monitor
streaming data for patterns. For feature representation, we
investigated various features of an audio such as MFCC, Filter
Bank, Bottleneck and found that Bottleneck features are stable
to variations in the multi-speaker environment. The DTW-SPRING
is performed on these features of both the query audio and
corpus audio. The distance matrix computed by SPRING shows the
occurrences of the query in the corpus. Figure below shows the
heatmap plot of the distance matrix calculated by SPRING for a
corpus audio with four occurrences of the query audio. After
this phase, we built a Web-based GUI for the users to interact
with the system. Figure below shows the final product we built
along with the User Interface. The user can record or select the
query audio from the system and search for its occurrences in
the corpus audio. This system can be extended into an audio
search engine by indexing multiple corpus files across
Gangeshwar, Ho Thi Nga, Dr. Chng Eng Siong
In this work, we investigated a method of using Sequence to
Sequence learning model to punctuate a sequence of unpunctuated
text from Automatic Speech Recognition system. The aim of this
work is to preserve the contextual meaning of a sequence of text
which might be a spoken paragraph from ASR. This direction of
SUD has not yet been explored in the previous works as far as
our knowledge is concerned. It can be intuitively understood
that the context of the previous sentence can improve the
occurrence of the next sentence boundary. Towards this goal, we
are currently investigating the use of sequence models,
specifically sequence to sequence (Encoder-Decoder based) model.
The dataset was prepared from the Wikipedia articles dump. Here,
every paragraph is a sample in the dataset and is considered to
be a sequence of text to be punctuated. The sequence length of a
paragraph varies from 75 to 750. Around 1.3 million paragraphs
were prepared and pre-processed for the task. In the first stage
of the pre-processing, only text and periods(.) are retained and
all other symbols and punctuations are removed. We do not want
our model to learn that the presence of capital words are also
an indication of the beginning of a sentence, hence all
characters are converted into its lower case character.
The model is two layered on encoder side with bidirectional Long
Short-Term Memory (LSTM) cells. And the decoder is two layered
with attention mechanism and uses LSTM cells. The attention
mechanism used in this experiment was "bahdanau" additive
attention as introduced in this ICLR 2015 paper. Alternatively,
we will be using the "luong" multiplicative attention for
comparison. The vocab size was chosen to be 160k and input size
of the model is 75. The word embeddings are used as features
with dimension size of 300. Every paragraph is sent as a mini
batch of input length 75 with variable batch size (depending on
the number of words in that paragraph). We used the concept of
Truncated BPTT to help the model in remembering the text seen in
the previous batches.
The model with the above configuration did not perform better
than the baseline model. Hyperparameter tuning is required as we
observe that the model is overfitting after a certain amount of
time. Also, we suspect the issue could be due to the practical
nature of LSTM. Hence, we are currently investigating the use of
Phased-LSTM which can remember words previously seen up to 1000
timesteps as claimed by the author. Additionally, reducing the
input size might also help.
Enhancing the UX by banks
Using Artificial Intelligent systems to enhance the User
Experience(UX) of the CitiBank’s customers
o Customer Care Service Application using Artificial
and Machine Learning.
o The AI agent responds to user’s query using natural
processing just like a customer care representative.
o Goes beyond the conventional method of responding to
being more personalized.(By learning user’s banking
o Help users with decision making in
This application was submitted to Citi Mobile Global
Challenge. Though we couldn't win the challenge, the
curve was very steep. Application Stack:PDF UX Mapping:PDF App Wireframe:PDF
A Social Network for Travellers.
It combines the power of Social Networking and the Machine
Learning algorithms to deliver an elegant experience to the
users of this website
o Flask – Python framework
o Amazon Web Services
o Google Cloud Platform
The application is in development stage and will be released
ListSketcher is an Android App, designed and developed by two
young engineering students of 2nd Semester from BIT,
I developed the application and my partner took care of
marketing the app to the public.
ListSketcher was developed with the motive of allowing users
share their list easily.
ListSketcher is an organizer, where in you can create and
track of your Lists that you create in your day-to-day life.
the unique feature of this app is that it allows you to
your Lists with your friends or colleagues as a whole file.
when the person on the receiving end opens the attachment,
opens in ListSketcher and he/she can save it in their
database and edit it accordingly. By this way it saves more
and keeps them engaged.
This app was developed to take part in the Google AdMob
Student Challenge 2014.
Disk Operation System utility provides the maintenance of
In the directory services the following operations are
About the project:
In this project the main menu is divided into two services:
one is the directory services and the second one is the file
services. The item from the main menu can be selected by
pressing either the enter key or the hotkey which is
by the bold character.
This project uses the library dos.h which
defines various constants and gives declarations needed for
and 8086-specific calls. The union REGS available in the
header file is used to pass information to and from the
functions int86() and intdos().
I built this program single-handedly for my 12th standard
CBSE Final Examination Project.
Honoured as "The Best Student Project" By Kendriya
1. The User takes a pic from the book using the app.
2. The app recognizes the words in the pic.
3. The user is given the option to select the word for which he
wants the meaning.
4. The app makes API calls to the server and gets the meaning of
An Android app that fetches Udacity courses data using Udacity
APIs and displays the data to the users.
This project was part of the Online Course "Developing Android
Apps" by Udacity.
The Course was sponsored by Google as a part of Android
Student Club Initiative in Jan'15.
Though I knew Android development before taking this course, it
was a good learning experience.
Checks results from VTU Results page for announcement of
Developed this because frequently checking results was a tedious
process since the servers respond very slow.