Nisarg Dave.

1. Explainable AI for Understanding Decisions and Data-Driven Optimization of the Choquet Integral using Superior Super-Computer of Michigan Tech (Scientific Computing using HPC)

Timeline: Oct 2018  

 

Building different Deep neural networks based on different architecture on ImageNet dataset. Investigating the new XAI approach using Fuzzy Choquet integral introduced by Dr. Timothy Havens & team.

Working on Michigan Tech's High-performance computing infrastructure for Scientific Computing based AI applications. Incorporating modular, powerful & parallelize code on HPC/ Supercomputing infrastructure. Developing model on Superior Supercomputer on Original ImageNet (full) dataset. 

Developed the different DNN’s using AlexNet, ResNet, CaffeNet, GoogleNet architecture. The code was compatible to handle the parallelized tensor flow enabled GPU version. It’s a fully automated process oriented SIL based workflow, which does several tasks in a single command. The code was designed by considering the need for Sustainable computing so it can be retrained using transfer learning in the future. 

Mainly used, 
Python, BASH Scripting, Tensorflow for code development.

 

2. Explanation in AI System ( using Transfer learning on Inception Classifier) - DARPA Project

Timeline: May 2018  | Team: Shane Mueller

 

As a Graduate research assistant working on a project of DARPA(US Army research agency) under Dr. Shane Mueller. The research majorly focused on the transfer learning ability of Machine learning.

We did train our own Google Inception like classifier for different examples. The examples contain different shape, size, and orientation of different types of images. We did test the retrained model to the new test cases for finding its robustness. The main investigation here is to check whether the machine learning model can perform as accurately as it was before without tempering test cases. The objective is to make the deep learning models robust enough for any use test case (ex. the test image that easily identifiable by humans but not by machine)

I did use the Amazon web services, python, tensor flow, and shell scripting. Moreover, for creating new test cases of image, I did use Photoshop and ImageMagick for creating image transformations.

The project included the work in the areas including,
• Deep Neural network 
• Convolution neural network
• Machine learning using Amazon Web service ( Elastic computer engine)
• Classification analysis using R

Overall I used Ubuntu Linux environment with Python 3 and Tensor flow. (followed traditional Data Science cycle)

• Data munging/wrangling
• Exploratory analysis
• Deep learning model creation
• Model training
• Model Validation & testing (Google cloud vision)
• transfer learning
• Parallel, GPU computing integration

 

3. Experiential User guide for Facebook's Business page

Timeline: May 2018  | Team: Erin Richie, Sivaramakrishnan Sriram, Shruti Ganjoo

 

Developed the set of experiential user guide lessons for Facebook's business pages. The EUG Development was held in four different parts such as,

1. Design & data collection
2. EUG Implementation
3. Methods
4. User guide lessons

We did find the most important points and flaws. After finding the key pain of customers. We did focus on that misconceptions and then provided the possible explanatory solution for the same. We used different methods for explaining the process or point to the users. We even discussed the detailed implementation of the methods in the next section. For a more clear understanding, we used the background information with the possible backing points.

 

4. Optimization using Artificial Bee Colony

Timeline: May 2018

 

Using another swarm intelligence tool for getting optimized values of fx and x within the given range of benchmark functions. I used the Artificial bee colony algorithm for optimizing the values for given benchmarks in a given range. The optimization was carried out by a process inspired by nature, using the nature of honey bees during their food hunting process.  

There are three types of bees,
- Employed bees
- Scouts
- OnLookers

Employed bees will gather food from food sources. onlookers will look for newly available sources, they will see the dance of Employed bees ( dance frequently if nectar amount is more ). Scouts will search for the new food sources. The optimization here is to find the best available food source and the associated cost for the same. The process continues until all number of onlookers gets the food source and all the best food sources occupied by bees. so the final optimized output is the best location for the food source and associated cost for the same. 

For Artificial bee colony, I did perform the 100 runs on each benchmark functions(Ackley, Branin, Dejong, Rosenbrock, Rastrigin) Moreover, I did report the fxbest and xbest found for the each of the benchmarks. In this way, I used the algorithm inspired from nature to solve the optimization problem.

 

5. Optimization using Particle Swarm Optimization

Timeline: May 2018

 

Performed one of the Swarm optimization procedure, Particle swarm optimization. The goal is to optimize the benchmark functions such as Ackley, Branin, Dejong, Rastrigin, and Rosenbrock. Developed the PSO function which optimizes the objective function in the given interval range. After getting it optimized the next goal is to find the optimal parameters for all benchmarks. I did some trial and runs to find the best parameters overall. After this, I did find the best parameter for each function individually. Now I tried getting the early stage convergence using the given parameter set. In short, I was trying to get less number of steps/iterations for each benchmark. In this manner, I did analyze the optimization of benchmarks using swarm optimization techniques in detail.

 

6. Adversarial attack on Google Inception v3 Classifier

Timeline: April 2018 | Team: Erin Richie, Sivaramakrishnan Sriram

 

Performed the Adversarial attack on Image classifier. The Deep neural network based model Inception v3 by Google has good capabilities and deep insights of image classification but still, there are ways to fool the DNN based image classifiers. We created adversarial samples using the Fast gradient step algorithm. We used Caffe Deep learning framework (python environment) and implemented a Targeted and non-targeted attack for the classifier. 

In this experiment, we studied the nontrivial exploits, different contrastive examples and some of the outlier cases. We established and demonstrated the systematicity of the attack. we answered the questions such as, How the attack works and how to make users aware of the usage of the system. Moreover, we created explanations for "How and when the system generates an error?". We proposed different ways of explanations for the users to make them understand the working and cautions of the system. Mostly all the Deep learning or Machine learning based systems without any defending mechanism can be affected by an attack so making systems robust against adversarial attacks is necessary. The project was part of the course explainable AI (XAI) or explanation in AI systems.

 

7. Genetic Algorithm Project

Timeline: April 2018

 

Developed Generic genetic algorithm code in MATLAB which finds the optimal maxima of function and minimal optimal of the parameter. Played with Genetic algorithm process steps and tried different parameter values to identify the best scheme for the general GA process. Developed the general procedural function and tested the function on the set of different parameter values. I did set the optimal parameters which overall generates the best results for our benchmark test function.

 

8. Iterated Prisoners' Dilemma using Genetic algorithm

Timeline: April 2018

 

Studied various algorithmic theories of Game Theory, including the iterative prisoner's dilemma. Developed the Genetic algorithmic approach based simulation which identifies the best possible strategy for the prisoner's dilemma problem. In the domain of game theory, the no regret situation and Nash equilibrium should be achieved. I did implement the Nash equilibrium and performed the test on the Axelrod's tournament using the different strategies. I did analyze the strategy results and trade-offs for each of them. The analysis was carried out on the different set of strategies which actually won the competition. Furthermore, I analyzed the prisoner's behavior on the 0-memory, 1 memory ... n memory types different forms of Iterative problem. After considering the previous moves, I did speculate the next possible moves using the best strategy found for the problem.

 

9. A/B WOZ study of Explanation and Justification of Decisions made by Deep neural network based Visual Search application for visually impaired

Timeline: Feb 2018 - Mar 2018 | Team: Erin Richie, Sivaramakrishnan Sriram

 

Conducted the WOZ Study for the Deep neural networks based AI System. An actual woz study for the visual search application. The analytical and experimental study for finding feasibility, soundness, an explanation for the given system. we considered the convolutional neural networks and Recurrent neural network based Deep neural network system which captures the environment and describes it in natural language to the users. We did a fake setup and perform the experiment on different users. After experiencing the entire system the user will write the feedback. Based on the feedback by users, we did analyze the level of explainability, trust, and robustness of the system.

 

10. Fuzzy Clustering Project

Timeline: Mar 2018

 

Implemented Fuzzy C means clustering algorithm in MATLAB. I did also code other two variants such as GK-FCM and PCM (Possibilistic C Means) algorithms. Experimented on different initialization types on each algorithm. Moreover, I did try different parameter values and chose the optimal parameters for each of the algorithm. After this, I performed experiments on iris data for all of the three algorithms. I did try out different initialization scheme such as random selection, selecting from Gaussian distribution & selecting from data points in vector form.

Performed 100 iterations for each algorithm using the best initialization & reported label accuracy.
Estimated the parameter optimums using the same process and identified the ideally best case parameter selection for each of them on the iris dataset.

 

11. Business case for Deep neural network based Visual Search application for visually impaired

Timeline: Jan 2018 - Feb 2018 | Team: Erin Richie, Sivaramakrishnan Sriram

 

The Project was part of the Explainable Artificial Intelligence system. We have prepared an in-depth business case for the Deep neural network based AI system. The system was for visually impaired people. The system is the application of visual search. We discussed the device which captures the live environment and process it using high-level deep learning networks to describe it in a natural language so that visually impaired people can understand what's going on around there. We need to consider the aspect of Explainable AI into this. How to make AI systems explainable and understandable in front of others. We discussed different use cases such as CEO, Engineer, Users, regulator, etc. For each of them, we came out with system specification details and possible explanations. Explaining a complex system in a simple and complete manner was our task. The models we discussed were based upon Convolutional neural network and LSTM - Recurrent neural network. This project was fun because XAI is the new research topic and I love doing research. Still, the question is simple, How to make AI systems self explainable? but the answer is very complex. We tried giving the best explanations for each of the use cases. The case was specific but the perspective was wide because we need to see through each different lenses so I found this project very interesting and challenging.

 

12. Fuzzy Controller for controlling the Robot movement

Timeline: Feb 2018

 

Designed the Fuzzy controller for the Robot moving in the specified arena. The Fuzzy logic was designed for the Robot to achieve the specified goals. The task is to achieve all the goals in the given time frame. The time or number of steps should be as minimum as possible. The optimization for a total number of steps was performed using different tweaks to the existing controller. The overall fuzzy logic procedure including designing the fuzzifier, Fuzzy rules, mapping, and defuzzifier was performed. At last the fuzzy inference system was designed to integrate everything. The simulation was done in the Fuzzy toolbox of MatLab.

 

13. Company Analysis of OcuGlass

Timeline: Sep 2017 - Dec 2017 | Team: Aoi Buto, Chirag Dave

 

This Project was part of Managing Innovation and Technology course. The Project was about doing research on the selected firm. We opted OcuGlass LLC. an only American acid etched glass manufacturer. We studied the firm, their business process, innovation strategy, technological details, marketing strategies, financial state, competitors, business strategies and business opportunities. We visited the company, factory and we attended meetings with the CEO. We understood the firm from the client, student and organization's point of view. At last, we came up with SWOT analysis and Recommendations. The recommendations were based upon the various theories we studied in class. The theories related to innovation and technology management. From the process to the product, each and every business process step was studied and analyzed by us. We presented a thorough analytical report of the facts and recommendations to the CEO and class. The report was in depth business case report. We suggested possible solutions to their main problems. For each and every solution we presented the alternatives with the possible way of implementation. The implementation guide was in detail and very precise.

From this project, I learned a lot of things regarding the business aspect and the technological aspect of innovation in the organization. This was really very memorable project experience.

 

14. Consumer Analytics for Rozsa Center for the Performing Arts

Timeline: Sep 2017 - Dec 2017 | Team: Mary Jennings, Ankit Kapoor, Chirag Dave, Priyansh Agarwal

 

The Project was part of Information system management and data analysis course. The objective of the project was to pick an organization, find the biggest problem of it and solve it. We decided to work with the Rozsa Center for the performing arts. It's the arts center of MichiganTech. We studied the Rozsa as an organization. The study included core business processes, business model, business strategies, IT strategies and their integration in business, Information system management and Problems affecting the organization. The Rozsa center organizes the events and manages them. They even rent their space to renters. Their main revenue stream is consumers who buy the tickets! They didn't have any kind of data collection mechanism so they were not using data science for improving their business. I suggested them to collect the data so We headed in the direction to establish the data collection mechanism for Rozsa, we established the mechanism. In parallel, I developed the machine learning model which uses the previous ticketing data and predicts consumer behavior. The application is customer discovery and identification, understanding and knowing your consumer's behavior. We developed a prototype of a predictive system for the same and established surveys and questions and collection mechanism. At last, we presented the whole business case in front of the class and solved the problem. It was one of my great team experience.

 

15. Fraud Analytics & Modelling

Timeline: Jan 2017 - May 2017

 

The Project was part of my final semester at Nirma University. The project was on Fraud analytics and modeling. I was a data scientist intern, the primary task was to analyze the various fraud detection methods and their performance. I have implemented and studied the performance of logistic regression, Support vector machines, Random forests and convolutional neural network. Demonstrated the pros and cons of all methods. I've generated a detailed draft of comparative analysis of all methods.

In the second part, the research on differential privacy mechanism was carried out by me. I was thinking of privacy-preserving machine learning. I did the research on privacy-preserving machine learning modeling using differential privacy and presented as a whole prototype project in front of a jury of organization and university. 

Tools used: Python 3, Anaconda Jupyter Notebook, Scikit learn, Other python libraries

Domains explored: Machine learning, Deep learning & Statistical computing, Computer Security.

 

16. Research on Differential Privacy in Machine learning

Timeline: Sep 2016 - Dec 2016 | Team: Ketul Majmudar, Himen Sidhpura

 

This Project was part of the course Information retrieval systems. After studying machine learning and intelligent systems, I was interested in making them secure so I started my research on the same. Under the guidance of Prof. Sapan Mankad, I studied several mechanisms for privacy preserved machine learning. Reviewed the different mechanisms and compared it with various systems. I wrote the Review paper for an in-depth review of that system and presented the mathematical model for differentially private machine learning-based system. The proposed and discussed system had complex architecture integrated with a data driver and noise generator. I've explored the various areas of computer science such as statistical modeling, mathematical modeling and differential privacy in machine learning. My research consisted of architectural details, working details, sequential modeling and even architecture of whole new privacy-preserving mechanism. The final outcome was a review paper of the mathematical model of differential privacy.

 

17. Optical Character Recognition for Handwritten digits

Timeline: Sep 2016 - Dec 2016 | Team: Shefali Jain

 

This was part of my minor project, I was eager to apply my machine learning skills to the real-world problem so I picked out a handwritten digit recognition system. We started exploring the various algorithms for character recognition system. we've used the MNIST dataset for the recognition.

First I started comparing various algorithms for OCR. I've used KNN classifier, Naive Bayesian, Convolutional neural network, recurrent neural network, and Boltzmann machines.

We came up with an idea to implement CNN for OCR. We used Matlab for the same and implemented the working model of CNN for recognition. It was basic CNN which takes images as a input and classifies the number into various classes as a output. 

The final product was a working model of CNN for character classification. The task performed by us contains data munging, feature engineering, feature extraction and classification based on principle components.

 

18. Research & comparative analysis of Deep learning networks

 

This was one of my personal project for research and exploration of various deep learning models such as Feed forward neural networks, back-propagation neural networks, Convolutional Neural networks, Recurrent Neural networks, long-short term neural networks and Generative additive networks, Boltzmann machines, MCMC based networks, etc.

 

19. Predictive Analytics in Health Sciences (Diabetes Case study based analytics)

Timeline: July 2016

 

Developed Diabetes analytics based on historical data of behavioral science of disease. The analytics were based on determined fixed factors of disease. The tools used were Python - Jupyter notebook and related python libraries such as sklearn. 

 

20. Search engine based Information retrieval system

Timeline: June 2016 - July 2016

 

Designed a Python-based Information retrieval system. The system was kind of document retrieval system based on TF-IDF score. Incorporated the information retrieval concepts and learned the different ways to search queries and optimizations.

 

21. 3D FPS Game using Unreal Engine 4

Timeline: Jan 2016 - April 2016 | Team: Ketul Majmudar

 

Created 3D FPS Game Using Unreal Engine 4. I was curious to learn game development concepts so I decided to join my friend Ketul for his game development project. I like to learn new things, moreover, I do believe in diversifying my skills. we decided to create the game prototype by the end of the project. We have used the unreal engine for creating the game. The 3D graphics were created by Ketul and I was responsible for the blueprints based coding workflow. The unreal has C++ based blueprint workflow for coding the game. The coding environment is unique and different than another scripting so it was a really very new learning experience.

 

22. Recommender System in R

Timeline: April 2016

 

Designed Multivariate Regression Model For Predicting Movie Success using R. During the class of machine learning we designed the system which predicts the success of movies by considering various factors. It was a multivariate linear regression model which predicts the success rate of a movie based on historical data provided by IMDB.

 

23. Research Case study of Cloud Computing in Analytics and Machine Learning

Timeline:  March 2016 | Team: Barkha Patel

 

Research and Literature Survey on Cloud Computing. During the class of cloud computing, we started investigating the importance and integration of cloud computing in business. The application of cloud computing in machine learning and big data age. AI Systems and their dependence on cloud and big data were examined by us. We did create a business and technological case study based on this topic. Research on the current state of the art and current technologies was really very Informative experience. We have studied many AI systems and commented on their relation/integration with cloud computing. (AI Systems such as IBM Watson, Google DeepMind AI, Apple Healthkit, Intelligent assistants such as Siri)

 

24. Event management System Website for Nirma University

Timeline:  Sep 2015 - Dec 2015 | Team: Ketul Majmudar, Himen Sidhpura

 

In Nirma University, We have different Students Associations and Clubs like Rotaract Club, Computer Society of India, IEEE Student Chapter IT-NU, ISTE, etc. These clubs organize Technical/NonTechnical events, Activities, and Workshops. We have designed a Website which will handle all events and activities online. Our website can take care of event registration, payments, creating and managing events and all clubs related tasks. At the end of the project, we came up with the working prototype of the website. We need to do some refinements before publishing it or making it live! But at the end of the project, we had the prototype model of event management system for our university.

 

25. Online Quiz System ( Client Server Application )

Timeline:  Nov 2015

 

Developed Quiz Application using Client Server Programming in C as a part of Data Communication Networks Innovative Project.

 

26. Compression algorithm comparative analysis

Timeline:  Jan 2015 - April 2015

 

Conducted Research on Compression Techniques and their Performance Analysis on Different multimedia applications.
Wrote one Review Paper based on Analysis and Comparison of Different Compression Techniques.

 

27. NISC- No Instruction set Computer Logic Design

Timeline: April 2015 |Team: Himen Sidhpura, Ashutosh Joshi

 

Designed Simulation of NISC based Computer Processing Unit Design in Logisim. Studied NISC and RISC in depth. This was the innovative project for computer organization subject.

 

28. Shopping mall management system

Timeline: April 2014

 

Developed Shopping Mall System in C++ as a part of the Innovative assignment of Art of Programming Course.