Daniel Sam Pete
Thiyagu

Click this Link for more info.

Click this Link for a ppt.

Using gradient descent based training algorithms incorporated with an attention mechanism, we were able to build a deep learning model that learns to generate image descriptions given an unseen image. Inspired by past work in treating the caption generation task as a neural translation task, our deep learning network tries to represent natural language, visual features and visual focal points in the same embedding while using a recurrent network to generate sentences. We implemented and evaluated our attention model with the traditional neural approach to image captioning as well as non-deep learning methods. Our evaluations were done using the bleu, meteor and cider metrics. The results show that attention-based deep learning models can learn to produce good image captions. We also show visualizations to infer how a model is able to focus on certain parts of the image while predicting the next word.

• Technologies used: Python, Keras
• Date: November 2016 - December 2016

Click this Link for more info.

In a micro-blog such as twitter, there are a wide array of tweets on each topic of interest according to the user.

In order to enable users to quickly view a summary about a topic searched, a system is proposed in which it involves two major steps.

1. Speech-act based classification of the tweets
2. Generating the summary with the classified tweets

In this approach, given an input query topic, the best relevant tweets are retrieved by the system using tweepy API. Then speech act recognition is done, which is a multiclass classification problem solved using the word based linguistic features and features based on twitter. The Speech Act Recognition recognizes speech acts such as statement, question, comment/suggestion and the rest are tagged as miscellaneous. These four type of speech acts are the most common types found in twitter tweet stream. The classification is done using Logistic Regression after testing that it worked better than SVM. The unwanted/miscellaneous tweets are ignored and then an abstract summary is formed using the remaining tweets. Phrase Extraction and Entity Extraction is done to each tweet and the necessary information is used. Information about entities used in the summaries are stored in a database to avoid redundant information. Summarization is done based on k means algorithm with which an abstract summary is formed which is based on the uniqueness of the tweet and the similarity and various other features which would be used in forming a valid summary. The Summarization is carried out by gathering information from the most informative, time-based and novel tweets from each cluster formed through k means algorithm. Novelty is identified using IDF values and other such measures.

• Technologies used: Python, Octave
• Date: December 2014 - April 2014

Click this link for more info

Source Code: Click this link

We won the best use of JIBO in HackUMass 2016

First, ask Jibo to take a guess your drawing. It will take a picture and send it off to be classified by our custom-trained tensor flow neural networks. The first step uses openCV to detect the playing area within the board and crop off the rest of the image, and make it black and white. The image then gets passed off the our tensorflow model to be classified. We build trained neural network using hundreds of images pulled automatically from Bing Image Search. We used the "line drawing" category to limit it to pictures that look more like something drawn a whiteboard. We then supplemented the training data with a few examples of our own hand-drawn pictures. This network gives us a likelyhood for each possible label, which is sent back to Jibo. Then Jibo will go through these items one by one, listening for a "yes" or "no" from the user. There is an adorable celebration and frustration animation that plays when it gets the answer right or wrong. This continues until it exhausts its entire vocabulary, at which point it insults your artwork in protest!

• Technologies Used: jibo,javascript,atom,python,opencv,tensorflow,flask,docker,npm,whiteboards,node.js,neural-networks
• Date: October 2016

A natural language processing project which does an autocorrect of a subset of the grammatical rules present in the Tamil. It involved generating the various parts of speech from the stem words and using a syntactic parser to aid in verifying the correctness of the grammar and building an effective algorithm was built to rectify the errors if found. Repo link : https://github.com/danielsamfdo/tamil_grammar_auto_correct

• Technologies Used: Python
• Date: February 2013 - April 2013

Created a two player and a one player game for Chess. The one player game utilized Alpha Beta Pruning Algorithm.

• Technologies Used: Java
• Date: April 2012

An online website for a church written using laravel framework which provides the admins reports on the attendance and details of the users. It also utilizes AdminLTE responsive framework.The backend database is a mySql DB. You can check it out in https://church-management.herokuapp.com/

The repo for this is in thislink

• Technologies Used: Laravel
• Date: April 2016

An online website application for a religious organization written using html and firebase which helps the admins to reach out to other users on their updates to their youtube channel and events.The backend database is a noSql DB. You can check it out in https://danielsampete.000webhostapp.com/www/

The repo for this is in thishttps://github.com/danielsamfdo/hlmAndroidApp

• Technologies Used: Firebase
• Date: August 2017

Click this Link for source code: https://github.com/danielsamfdo/garbage_collector

I worked on the design of a memory allocator as well as a garbage collector that triggers at various instances when it requires memory as part of course requirement for COMPSCI 630.

• Technologies Used: C++
• Date: Feb 2017 - March 2017

Harness unlabeled multilingual data and learn a common representation under which structure and meaning is shared across various languages. Mentors: Oracle personnel.

The repo for this is in thishttps://github.com/Shamya/multiLingRep

The wiki for this is in thishttps://github.com/Shamya/multiLingRep/wiki

This PPT provides additional Information ( https://docs.google.com/presentation/d/1NM_EbD5U62_F4p8pruH4nyEK3w9tMfqOIGdOKRPZAtQ/edit#slide=id.g1f26bf60b6_0_45 )

• Technologies Used: Python, Tensorflow, Scikit, Nltk
• Date: Jan 2017 - May 2017

Creating a gamut of visual analytics for delivering insights in the health care domain using Salesforce Analytics. I worked particularly on the Total Joint Replacement Condition analytics with salesforce development and publishing app packages to salesforce

• Technologies Used: Salesforce, Wave Analytics, Python, pgsql
• Date: May 2017 - Aug 2017

Developed an emergency notification application that sends real-time notifications using Laravel.

Designed an automatic website generator and wireframes for the product design of the emergency notification portal

• Technologies Used: Laravel, Web Development, WireFraming Aws
• Date: Nov 2016 - May 2017

Improved performance of various modules such as sending bulk invitations by 10x and rendering of web pages by 2x

Implemented and enhanced core modules for preventing past mentor matching, signup workflow of the portal, an internal tool for automation of migration in the mentor portal and career review generation in the career development portal

Resolved critical level issues such as the inability of the customer to access the portal due to SSO issues.

Worked on multiple feature enhancements like Admin Favorite views, Bulk Loads, a few enhancements in Mentor Matching as well

Full-stack developer in Customer Success Engineering Team, Chronus iOS application Team and NCH Team

• Technologies Used: Ruby on Rails framework, Thinking Sphinx, MongoDB, MySQL
• Date: Jun 2014 - Jul 2016

In the bin is a location-based web application where users can submit locations of publicly-accessible trash and recycle bins near by. Hacking for Good 2017 hackathon on June 17, 2017 @ Harvard Innovation Labs. Submission by team HackU: Pankaj Bhambhani, Nitin Kishore Sai Samala, Daniel Sam Pete Thiyagu, Daigo Fujiwara

Presentation slides
https://docs.google.com/presentation/d/100pHPRdu0_BviOYKNsm83SPsRO9klxhmizqk1lga0lc/edit?usp=sharing

• Technologies Used: Firebase Web, Google Maps JavaScript API, Material Design Lite
• Date: Jun 2016 - Jun 2016

Click this Link for more info: https://drive.google.com/file/d/0B09veTD03OS_TU45VWtZT0RFd3c/view
Click this Link for a poster: https://drive.google.com/file/d/0B09veTD03OS_VnRaUm9TZDF4WFU/view
In this project, we present a comparative study to solve the problem of rating the similarity of two sentences. Our methods range from some well known corpus based methods to little known strategies for computing semantic similarity of the two given sentences. This includes computation of semantic similarity using semantic knowledge bases like WordNet, supervised learning methods like Linear Regression, Support vector machines and Multilayer Perceptron, as well as methods utilizing Word2Vec. We also use various semantic similarity metrics like Jaccard Coefficient and Containment Coefficient to aid the computation of semantic similarity. Some of our methods performed better than the standard baseline algorithm

• Technologies Used: python,scikit-learn,numpy
• Date: Nov 2016 - Dec 2016

Click this Link for source code: https://github.com/danielsamfdo/memcached

In this course project for COMPSCI 630, we were working as a team. We designed a Memcached system which can allocate a certain size of memory and added algorithms for replacement like LRU, Random Replacement when the size of the cache becomes full.

• Technologies Used: C++, Socket Programming
• Date: Apr 2017 - May 2017

Click this Link for source code: https://github.com/danielsamfdo/carSafetyStats

In this individual course project for COMPSCI 687, I designed a system to evaluate the effectiveness of alarm signals based on Reinforcement learning algorithms like Qlearning and SARSA with models based on cautious, moderately cautious and irresponsible drivers.

• Technologies Used: Python
• Date: Nov 2017 - Dec 2017

Click this Link for source code: https://github.com/danielsamfdo/bnn

In this group project for COMPSCI 682, We designed a binary neural network for image classification tasks like CIFAR10 and MNIST using standard architectures like VGG, We also experimented with different levels of quantization and with binary activations and weights.

• Technologies Used: Python, Pytorch
• Date: Nov 2017 - Dec 2017

Click this Link to know more: https://devpost.com/software/my-home-memories

This is an Alexa based App created to helps us share personal memories and knowledge like never before.

• Technologies Used: Python Flask, AWS, Alexa
• Date: Nov 2017