Understanding the Perception of Courteous and Humorous Behavior using Prosodic and Lexical Features
Computationally understanding the perception of a speaker’s characteristics is an important task for speech processing, behavioral outcomes and dialogue systems. Our work focuses on the classification and understanding the perception of two such characteristics: courtesy and humor. We analyze ratings surveyed from the SpeedDate corpus to build models using prosodic and lexical features of the speakers. We find that lexical features like Hedge-words and prosodic features like Intensity Min SD are important for detecting courteous while Exclamations and listener’s Intensity are important for funny. Using an Adaboost Classifier, we detect the funny with an accuracy of 0.74 and courteous at 0.76.
Multimodal Sentiment Analysis in Online Videos
In this paper, we address the problem of detecting sentiments in video clips with multimodal information. Using a corpus of opinionated videos from Youtube, we extract textual, visual and audio features for sentiment classification. The videos are divided up into utterances and annotated as expressing positive, neutral or negative sentiments. We combine the features extracted from each modality into a shared representation and classify the sentiments based on the trimodal features. Our experimental results demonstrate that a combination of features from multiple modalities improves the overall performance of sentiment classification.
Artificial Intelligence for 2048
A fun expectimax artificial intelligence algorithm implementation to play 2048. It uses gradients as heuristics. It manages to win the game 100% of the time and achieves an average score of 60,631. I also compare it to minimax and to other heuristics like empty tiles and monotonicity. The implementation is based entirely in client side javascript.
Outfluenza: Live Influenza Tracking Analytical Engine
In the US alone, about 3600 deaths occur because of Influenza. The disease is seasonal and has led to thousands of deaths worldwide every year without fail.
Current online influenza tracking softwares use past data and are not constantly utilizing all the new cases that occur everyday. The exception to this rule is Google Influenza tracking which tracks influenza by saving information about people who search for influenza on Google Search. However, this is not very reliable as it is not a good estimate of the spread of influenza. One person might search for the disease multiple times. Also, there might be false positives as people without the disease might look it up for academic purposes as well.
Also, there is a severe lack of prediction algorithms that are readily available and help arm health organizations with facilities and inventories to tackle an outbreak of influenza.
We aim to tackle this problem by collecting e-prescription data. E-prescriptions are a much stronger estimate of the spread of the disease as it tracks patients and prescribers that receive or issue medication for influenza. Additionally, as technology is becoming more widespread (explained below), more and more pharmacies are moving towards using the convenience of e-prescriptions.
Ray Tracer in C++
Over the course of a week in October, I decided to venture into a branch of computer science I had not yet explored. Having taken no classes on graphics at Cornell, the entire field was a black box. I started off by learning about rasterization and other graphics concepts through online tutorials until I finally decided that I knew just about enough to program my own ray tracer. This ray tracer does not perform any optimizations on object detection and can be greatly improved. However, it was an excellent starter project for graphics.
|
HIV Progression Prediction
This project analyses the progression of the human immunodeficiency virus by finding markers in the HIV sequence that predict a change in the severity of the infection. This severity is measured using two parameters: the viral load in a patient and the CD4 count. Using the nucleotide sequences of patients’ Reverse Transcriptase (RT) their Protease (PR) at the beginning of therapy, we test multiple machine learning algorithms using n-fold cross validation tests on the training data. The machine learning algorithms used are Multinomial Bayes Classifier, Less Naive Generative Bayes Classifier, Decision Trees, Linear Classifiers with Support Vector Machines and k-Nearest Neighbours with Needleman Wunsch dynamic programming algorithm. We discovered that there was a 62% direct correlation between the severity of the virus load and the probability of the patient improving. Through our experiments, we have ruled out certain features including codon transicions, nucleotide transitions and codon placements in patients’ DNA sequences as predictive influences. However, we have also concluded that the viral load, the CD4 count and certain positions in the reverse transcriptase sequences are directly related to the progression of the HIV virus.
Click here to read more
Click here to read more
Sign Language Translation
Abstract: This paper describes the design and implementation of a portable glove based sign lan- guage translator with LCD Display, Speech Synthesis, and Sign Language Education Software. It consists of a power glove with flex sensors, accelerometers and contact sensors. This glove is used to detect positions of fingers inorder to analyse sign language. The glove is connected to a ATMEGA64 microcontroller that analyses the signals from the sensors and compresses it and transmits it across the transceiver next to it. The signal is translated to text using two different algorithms to minimize errors. Using alpha compression, we store a dictionary of words in the base station that then takes the text being signed and converts it to speech. Finally, the signals are passed to an educational software that teached the user how to sign faster by racing against time.
|
|
The project was featured/published in numerous websites and articles including: Engaget, Slashgear,DeafTechNews, The Economic Times, Language Translation Inc., Element 14, Gadgets, The Tech Journal, Nix Times, and many others.
An Interactive Sink Utilizing "Fun Theory"
Research over the past few years has shown no change in people’s behavior towards washing their hands before or after handling money, using the bathroom, touching pets, or eating food. The percentage of people who are affected by communicable diseases has not changed even though numerous ventures have been spurred by organizations including the American Microbiology Society to spread awareness. Fun Lavoratories aims to decrease communicable diseases such as Influenza, Streptococcus, and Staphylococcus by propagating proper hygiene pratices. Our objective is to increase the number of people who wash their hands after using the bathroom by providing them with an incentive. The incentive will be in the form of an entertaining light and sound show. As people use the sink, a proximity sensor will trigger a microcontroller to start the light show using an array of Light Emitting Diodes (LEDs). It will also begin an integrated sound system that will play either a soothing tune or will project excerpts from current news articles. We have built a prototype that can be installed in public restrooms above the sink and will be protected by a waterproof casing. The cost of building a unit for each sink is approximately $112, making the adoption of such a unit extremely cost effective. We project that it will decrease the spread of diseases and provide an overall fun environment for its users.
Click here to read more.
Click here to read more.
Ocaml based Real Time Strategy Game
This is a RTS game coded in ocaml, a funcitonal programming language. THe graphics are done using java jFrame. The game starts off with two town centers and 6 villagers on two teams. The point of the game is to destroy the opponents town center. Two Artificial Intelligence bots play against each other by controlling one team each. The AI's are also coded in ocaml. Disclaimer: the graphics are from Age of Empires 2. This is an academic project and not a commercial one. |
|
Compiler for Xi
For a course at Cornell, I implemented a Compiler for an object oriented programming language called Xi. There are some known problems that will be fixed. The source code can be found in the repository.
Operating System
This project represents a Simulated Operating System that is build on top of Windows. It contains a preemptive Thread package for building concurrent applications. It supports network communication between different devices through a reactive ad-hock Network Level Protocol. At the transport level, packets can be sent either through an UDP or an algorithm similar to TCP. In addition, it supports a shell that reacts with a unix based file system.
The source code can be found here
The source code can be found here
Tachometer and Motor Control
The aim of this project was to build a digital tachometer that measures the speed of a motor using an infrared emitter through the fan blades of the motor and measured by a phototransistor and comparator circuit. A proportional{integral{derivative algorithm is
used to set the speed of the fan appropriately. Lastly, the system can be modied by a user by entering parameters for the PID control and set the desired speed of the fan through the UART. The aim of the project is to determine the accuracy of the PID control algorithm.
Click here to read more.
used to set the speed of the fan appropriately. Lastly, the system can be modied by a user by entering parameters for the PID control and set the desired speed of the fan through the UART. The aim of the project is to determine the accuracy of the PID control algorithm.
Click here to read more.
Plethysmograph
This project is a finger-based plethysmograph. A light is shined at receiver and, when a finger is placed between the two, a buzzer and light respond on every pulse. In addition, the signal can be saved on a computer and put through a simulated digital circuit to give a pseudo-real time pulse count. At a high level, the project involves the use of photo diodes and photo detectors, passive and active filters, anolog circuit design, anolog to digital conversion and digital system design. As a person inserted their finger into the clip, it would detect the subjects heartbeat and output a signal into the amplifier. The amplifier would then amplify the signal and send it to the filter to remove all the low and high frequency disturbances and noise. The filter would then send the clean signal across to the analof to digital convertor to create a digital signature corresponding to the heartbeat. This digital signal was then responsible for running the buzzer, LED and the digital readout.
Here is the project report.
Here is the project portfolio.
Here is the project report.
Here is the project portfolio.
Music Synthesis using Frequency Modulation
Will be uploaded soon.
Tracking System using a MSP430
Will be uploaded soon.
Quadcore Processor with a Cache and Ring Network
Will be uploaded soon.
31: A fun, simple card Game in Java
This is just a fun game I came up with because it's so interesting to think about the strategies and math behind it.