EMI-RNN
Multiple Instance Learning for Sequential Data Classification on Resource Constrained Devices
Microsoft Research, India
Advisors: Dr. Prateek Jain & Dr. Harsha Simhadri
Developed a multiple-instance-learning based algorithm (EMI-RNN) that recovers the distinguishing signature of minimum length for each class in time series classification. Smaller signatures results in smaller computational costs and effective use of classification model's capacity thereby improving performance while reducing compute by up to 72x. For nice data, showed linear convergence to global optimum in the number of non-noise samples in a more general non-homogeneous setting.
Accepted to NIPS '18.


EMI-RNN
GesturePod: Machine Learning Based Gesture Recognition on Resource Constrained Devices
Microsoft Research, India
Advisors: Dr. Prateek Jain, Dr. Harsha Simhadri & Dr. Manik Varma
Developed an efficient machine learning pipeline to enable GesturePod, a low resource microcontroller based device, to perform robust, low-latency gesture recognition. The ProtoNN algorithm powered prediction pipeline along with communication and storage stack works with just 32kB RAM on a 48MHz processor. The ML based approach as opposed to a rule based approach gives as extensibility in the sense that users can add new gestures to the system with very low effort - programmability.
In submission, CHI '19.
Microsoft's demonstration at NIPS '18.


EMI-RNN
WakeWord: Keyword Spotting on a Pi0
Microsoft Research, India
Advisors: Dr. Prateek Jain & Dr. Harsha Simhadri
Developed a small, fast and accurate classifier based on LSTM and ProtoNN to enable real-time keyword spotting on Raspberry Pi3. Developed EMI-RNN to make it possible on even smaller devices (Raspberry Pi0, MXChip
Accepted to NIPS '18 as part of MLPCD2 workshop.


EMI-RNN
Talk-Bot: Federated Human Detection for Collaborative Multi-angle Videography
IIT Patna, India
Advisors: Prof. Arijit Mondal & Prof. Jimson Mathew
An attempt at automating in-class-room videography at low costs. Multiple Raspberry Pi 3 based recording devices, each powered by a cascaded person of interest tracking stack (MOG + Particle filter + Viola Jones and DNN face recognizers), tracks the speaker and audience from various camera angles and records and broadcasts the talk in real time. All decisions on selecting camera angles, activating appropriate audio devices, identifying the current active speaker and so forth is arrived at through a distributed consensus.
Runner up at Grand Challenge, ISED 2016.


Nagging Naagin
Nagging Naagin: The Q-Learning Snake
IIT Patna, India
Advisor: Prof. Arijit Mondal
Taught an agent to play the classic Snake game through reinforcement learning. Created a custom version of the game to allow for a multi-bandit formulation (snake and adversary who places food). Implemented and analyzed various search and RL algorithms - reflex agents, min-max tress, expectimax trees, Q-learning and approximate Q-learning.


EMI-RNN
Universal IoT Gateway with Disaster Resilient Communication Pathways
CSI, SUTD, (formerly at NUS), Singapore
Advisors: Dr. Vishram Mishra & Prof. Lim Hock Beng
Developed a Universal IoT Gateway - a gateway that can command and communicate with all and any IoT devices, regardless of the manufacturer or communication protocol the specific device supports, be it BLE, Bluetooth, WiFi or ZigBee. This cross-protocol communication is made possible by an ontology based kernel that understands device specific properties and communication atoms. The ability to communicate with multiple protocols enables the device to also double as a disaster resilient communication pathway - a mesh network at the MAC layer.


EMI-RNN
Multi-node BFS for the Map-reduce Paradigm on Hadoop
IIIT Delhi, India
Advisor: Prof. Debajyoti Bera
Single source BFS, have well established algorithms for the map-reduce paradigm. These algorithms though have poor load balancing properties when working with real world graphs (social graphs, connectivity graphs) because of high sparsity in such graphs. We explore one potential solution to this problem - increasing the number of initial source nodes from which BFS is performed, that is, creating a multi-source-node BFS within map-reduce.


EMI-RNN
Single Cycle RISC-V Micro Architecture Processor and its FPGA Prototype
IIT Patna, India
Advisor: Prof. Arijit Mondal
Developed a RISC-V based single cycle micro architecture processor optimized for low-cost embedded devices, its bare bones simulator and an FPGA prototype. Additionally wrote a custom assembler-linker-loader tool chain to run native programs on the prototype.
Published at ISED 2018.


EMI-RNN
GSoC 2015: WITCH on a Board
Advisors: David Anders & Tom King
The WITCH was used as part of the Atomic Energy Research Establishment, Harwell, Oxfordshire in the late 1940s and hence the details of its workings were classified till early 2000s. WITCH works on very peculiar and outdated electronics (dekatron tubes instead of vacuum tubes, base 10 arithmetic as opposed to base 2 arithmetic, etc). In this project, I created a fully functional simulator for the WITCH with the MinnowBoard Max as a base. Because of the arcane technology used, a lot of the details of how the WITCH works exactly is still not understood and the project partly is responsible for asking new questions and getting the details correct.
Helped raise £50,000 for WITCH-E and The National Museum of Computing, England.