Roshni Borse

Echo state network on FPGA -Developed a 100-neuron free-running Echo State Network (ESN) model for time-series forecasting, where the network generates future predictions by feeding its own outputs back as inputs -Used Optuna to tune hyperparameters such as spectral radius, leakage rate and input scaling for the Mackey–Glass and Santa Fe laser benchmark datasets -Analyzing the effects of reduced-precision quantization (float-to-integer) on reservoir dynamics and stability on free-running prediction accuracy

Satellite image denoising on hardware for ISRO Objective : Developed a high performance OBNLM IP core for real time speckle noise removal - Python to C++ code conversion of the algorithm was performed for IP generation using Vivado HLS - Ported the IP in Vivado tool flow and build peripherals according to Zynq MPSoC - Significant improvement in timings was achieved

1) Implementation of Neural Network on FPGA for laser locking -Classification of Jet tagging model was performed using neural networks and a Keras accuracy of 75.62% was achieved -Used HLS4ML python package to generate IP of the frozen model and obtained an accuracy of upto 75.60% after conversion -Built peripherals around Zynq UltraScale+ MPSoC using Vivado 2019.1 -Processing such as pruning, quantization, and compression was performed before deployment -Testing was performed on baremetal as well as

-Worked on an industrial project sponsored by L&T, Mumbai along with a team of 5 during M.Tech course. -The project involves Mathematical modelling for Autonomous Underwater Vehicle (AUV).