Technical Portfolio
Video Summarization Using Reinforcement Learning With Attention
Aug - Dec 2020
Designed a method for video summarization that incorporates a self-attention mechanism to generate importance scores for the frames to generate video summaries.
Used Reinforcement Learning in the form of REINFORCE policy gradient method with a diversity-representativeness reward to train the model.
Employed the rank comparison method "Kendall's tau" as a means to compare generated summaries in addition to the generic F-score metric, enabling more meaningful comparison between various methodologies.
Skeletal Animation in PyOpenGL
December 2020
Implemented the complete graphics pipeline for rendering a Collada(.dae) file, including the model, texture mappings, keyframes, and bone information.
Implemented the interpolation between different keyframes using linear and quaternion models for translations and rotations, respectively.
Through The YouTube Rabbit Hole
November 2020
Productive usage of YouTube requires the user to have continuous self-control to stay on the relevant topics and not be lured away into watching irrelevant content - which generally makes the user feel exhausted, inefficient, and frustrated.
Empirical analysis of how the dissimilarity of content decreases as users go down the trail of recommended videos on the popular video streaming site, affirming the rabbit-hole phenomena' existence.
Music Harmonization Using Reinforcement Learning
November 2020
Modelled melody-based harmony generation in the form of states, actions, values, and rewards.
Trained an artificial neural network to represent the value function using a modified SARSA algorithm that considers the immediate past action in addition to the current state.
Sourced the training data from MIDI files representing a compilation of popular music. Best results achieved rewards that were 11.09 times the expected rewards for a random choice.
GradCAM Implementation
September 2020
Utilized the pre-trained InceptionV3 to carry out transfer learning on the P29-Cross-Pure-Dogs dataset. Managed to achieve a test-set accuracy of 85.9%.
Implemented the Grad-CAM methodology by Selvaraju et al. to visualize regions of the test image that are seminal to the classifier's output.
RISC-V Emulator
Jan - June 2020
Developed a RISC-V Instruction Set Architecture emulator from scratch in C++.
Implemented 1st order prediction pipeline, memory access, and cycle updates.
CMOS Sound Synth
November 2019
Designed and built a sound synthesizer circuit to compose electronic music using generic analog ICs(OpAmps, Inverters, Mux, Demux).
Incorporated a pattern arpeggiator circuit with controllable frequency parameters.
Exploring the Scope of Diversification of Wasteland Usage
November 2019
Analysed the impact of barren and vacant land in Indian villages.
Investigated possible reasons for this phenomenon and suggested plausible solutions to help tackle this problem in Dekhwala, our model village.