Driven Research Assistant experienced in the field of deep learning and machine learning. In process of completing PhD degree (Expected Graduation: Summer 2022).
The primary topic of this research is Deep Learning in Computer Vision.
Contributed by developing some novel techniques of data subset selection for deep neural network training.
Developed a novel mini batch selection algorithm for deep learning.
Instructed courses, Checked assignments, proctored tests, provided grades and took recitations according to university standards.
The primary topic of research was communication networks. Contributed in the development of a fall detection system which is capable of detecting an event, the fall of a human, using devices like a motion sensor, a heat sensor, and a vibration sensor.
Configured and deployed Master Data Management(MDM) Integration in the prevalent Siebel based system using IBM Datastage for the client The Teachers Insurance and Annuity Association of America-College Retirement Equities Fund (TIAA).
GPA: 3.6/4.0
GPA: 3.3/4.0 (WES)
Deep Learning
undefined[1] S. Banerjee, S. Chakraborty, “Deterministic Mini-batch Sequencing for Training Deep Neural Networks”, AAAI Conference on Artificial Intelligence 2021
[2] S. Banerjee, S. Chakraborty, “Budgeted Subset Selection for Fine-tuning Deep Learning Architectures in Resource-Constrained Applications”, IEEE International Joint Conference on Neural Networks (IJCNN) 2020
[3] S. Banerjee, S. Chakraborty, “DeepSub: A Novel Subset Selection Framework for Training Deep Learning Architectures”, IEEE International Conference on Image Processing (ICIP) 2019
[4] Z. Zhang, A. Mukherjee, S. Banerjee , "Fall Detection Devices, Systems, and Methods", (U.S. Patent No. 10,722,148). Current Assignee Florida State University Research Foundation Inc; 2020
In this project, we presented a novel algorithm to generate a deterministic sequence of mini-batches to train a deep neural network (rather than a random sequence). The proposed mini-batch sequencing strategy is deterministic and independent of the underlying network architecture and prediction task. [1]
In this project, we proposed frameworks to select a subset of the training set to induce a model with maximal generalization capability. This is helpful for training in a resource-constrained environment where the full training set can not be used to train a model. [2] [3]
The goal of this project was to use the video footage from a camera installed on the back of a vehicle to detect the presence of objects (especially kids) and to trigger an alarm in case of a dangerous situation. Such technology can be immensely useful to reduce child casualties. As a part of this project, we had also worked on the problem of pedestrian path prediction. Given the location of a pedestrian in an image, the objective was to predict his position at short time intervals (1 or 2 seconds) into the future. This can be tremendously useful to activate and trigger emergency braking in case of a potentially dangerous situation.(Summer 2018)
• Easybox, A sandboxing application for Linux: Easybox is a light-weight software that monitors all programs that try to access the directory. It was implemented using a File system in User Space (FUSE). Access of a file is given to a program only when it meets certain conditions. It intercepts the system calls of a filesystem to update the tables. It can also block illegal access. It uses database tables to store the record of every program and file pairs. The database which store the pairs also is considerably small. On average, the database takes up about 0.001 MB of memory per file in the sandbox. (COP5611 - Advanced Operating Systems)(Spring 2018)
• ECG Analysis, Diagnosis of Cardiovascular Abnormalities From Compressed ECG: In this project, we have implemented a mechanism to detect Cardiovascular Abnormalities from a compressed ECG signal with EM clustering and other data mining techniques. (CIS 5930 - Data Mining)(Fall 2017)
• Connecting similar products on the internet: "Matching Titles with Cross Title Web-Search Enrichment and Community Detection" using google search API. We used the methodology used in the abovementioned paper. In this project, we used Abt-Buy Dataset that contains 1097 matching product titles for electronic products. We matched the titles of Abt and Buy that refer to the same product. We have shown that this technique beats others like Jaccard Coefficient and TF-IDF cosine -term frequency X Inverse document frequency. (COP 5725 Database Systems)(Spring 2017)
• Keystroke security: An android application that provides cellphone security by analyzing the user’s keystroke patterns. We implemented a mobile user verification system which utilizes a mobile device’s touch screen for detecting keystroke dynamics. (COP 5659 Mobile Programming)(Spring 2017)
• Chatroom application: A Java chatroom application that enables a user to open and join multiple chat rooms. It also had provisions for private-chat, File Sending, Video recording with playback & sending and emoji support. (COP 5570: Concurrent, Parallel, and Distributed Programming) (Fall 2016)
• Myshell: A simplified Unix shell implemented using C. It had built-in commands like "exit", "cd", "pwd", "set", pipes, background execution, I/O redirection. (COP 5570: Concurrent, Parallel, and Distributed Programming) (Fall 2016)
• Messenger app: A simple messenger application using C++ where one can send friend requests and chat with friends. A peer to peer system using socket-based networking. BSD Sockets are used for communications between the client and the server. (COP 5570: Concurrent, Parallel, and Distributed Programming) (Fall 2016)
• Retail Management System: It is a software for handling a retail store database using Visual Basic and MySQL. It offered functionalities such as inventory management, reporting, purchasing and generating bills. (B. Tech Grand Project)(Fall 2014)