Projects
Continual Learning for Keyword Spotting (Spring ‘21)
Advisor: Prof. Preethi Jyothi, CS, IITB
Modeled the problem of new keywords arriving at various timestamps as a continual domain adaptation problem and reviewed several state-of-the-art techniques like EWC, CRA and LWF to tackle the forgetting happening in the network while training on the newer keywords. Also designed a joint keyword spotting and speaker verification model as part of an INTERSPEECH challenge. [Report]Efficient Neural Machine Translation (Autumn ‘20)
Advisor: Prof. Pushpak Bhattacharya, CS, IITB
Built a neural machine translation model, with small memory footprint and lower memory computation compared to state-of-the-art transformers, based on RNNSearch with efficiently tuned hyperparameters to achieve a decent BLEU score on the Multi30K dataset. Introduced an adversarial training mechanism to tackle overfitting in the model and improve performance. [Report]Adversarial Attacks and Robust Transformations (Autumn ‘20)
Advisor: Prof. Amit Sethi, EE, IITB
Implemented several white box attacks like FGSM, CWLA, PGD and Deep fool to demonstrate the deterioration in performance of neural networks trained on datasets like MNIST and CIFAR10. Performed a detailed characterization of transformation like JPEG compression, Gaussian smoothing etc. to see the improvement in models with adversarial attacks. [Report]Eigen Faces vs Fisher Faces (Autumn ‘19)
Advisor: Prof. Suyash Awate, CS, IITB
Implemented and compared two popular face recognition algorithms, fisher faces and eigen faces. Analyzed the differences in performances of the algorithms on the Yale Dataset which has sufficient variations in lighting and facial expressions, and provided a mathematical basis for this difference in performance as well. [Report]PCMO Transient Analysis (Summer ‘19)
Advisor: Prof. Udayan Ganguly, EE, IITB
Designed a LTSpice model to verify the lumped model temperature behavior of a PCMO device. Developed a physical and corresponding quantitative MATLAB model to compare the effects of electrical field and temperature on the reset transient of the device. Calibrated the MATLAB model parameters to give accurate results for various devices.[Report1], [Report2]Image Compression (Spring ‘19)
Advisor: Prof. Biplab Banerjee, CSRE, IITB
Implemented k-means clustering algorithm to achieve image compression with significant decrease in size while still retrieving most of the important attributes. Designed a Convolutional Neural Network in Keras to implement a basic classification task, and compare the accuracies achieved using the compressed and original image datasets.[Report]Transient Thermal Feedback Model for PMO RRAM (Spring ‘19)
Advisor: Prof. Udayan Ganguly, EE, IITB
Conducted an extensive literary survey to understand the conduction mechanism in PMO devices. Built a thermal feedback based model on MATLAB to see the role of self heating in PMO DC I-V characteristics and verified it with experimental results. Obtained the thermal transients to explain the distortion in DC I-V hysteresis on varying the ramp rates and the step response of the device.[Report1], [Report2]IITB-Proc, multi-cycle RISC Processor Designing (Spring ‘19)
Advisor: Prof. Virendra Singh, EE, IITB
Developed a 16-bit, 8-register processor which can solve complex problems. Implemented the design of Register File, Datapath, Finite State Machine and other components like ALU, and included operations like add, load, store and jump. Used VHDL for description of the hardware design of various components and integrating them along with the controller FSM using suitable select pins which were toggled depending on the current state.[Report]Hand Gesture Controlled 3-D Hologram (Summer ‘18)
Institute Technical Summer Project, Electronics and Robotics Club, IITB
Designed models and animations in Unity and basic operations like rotation, scaling and changing to other models were incorporated using standard model libraries. Constructed a self-designed IR sensor board to record the input in the form of hand gestures. Used Arduino MEGA to provide the interface between the Unity game and IR sensor board and the final output was presented in the form of a 3D hologram. [Video]