Siddharth Satyam
University of California San Diego
I am a recent MS graduate in Machine Learning at UC San Diego. Last summer, I worked as a Machine Learning Software Engineer Intern at Valeo, developing computer vision algorithms for autonomous driving perception using LiDAR point clouds. I have experience in Semantic Segmentation, Object Detection, Denoising etc. Recently, I worked on a multi-modal material estimation model using audio-visual cues utilizing CLIP, Whisper, and LLama.
Things I can do:
Publications
- H. Nikam, S. Satyam and S. Sahay, "Long Short-Term Memory Implementation Exploiting Passive RRAM Crossbar Array," in IEEE Transactions on Electron Devices, doi: 10.1109/TED.2021.3133197.
- S. Satyam, H. Nikam and S. Sahay, 2021. "Energy-Efficient Implementation of Generative Adversarial Networks on Passive RRAM Crossbar Arrays," arXiv preprint arXiv:2111.14484.
Projects
Multimodal Material Estimation using Large Vision Language Models
Developed a multi-modal model for object material estimation from audio-visual cues, leveraging transformer-based approaches and pre-trained language models.
Spectrum Based Fault Localization Using Graph Neural Networks
Implemented the spectrum based fault localization problem as a graph network of test cases and components. Created a graph neural network implementing message aggregation from test nodes to component nodes and vice versa. Generated component bug suspicion probabilities using component node embeddings through feed forward networks.
Long Short-Term Memory Implementation Exploiting Passive RRAM Crossbar Array
Performed a time series analysis using LSTM networks for the prediction of airline passenger dataset. Simulated a fixed amplitude in-situ training and evaluated effects of update symmetry and RRAM device variations. Performed a comparative study to predict the enhanced accuracy and energy efficiencies of LSTM implementation in passive over active 1T-1R crossbar arrays.
Energy-Efficient Implementation of Generative Adversarial Networks on Passive RRAM Crossbar Arrays
Implemented Generative Adversarial Networks to synthesize realistic looking images of the MNIST dataset. Performed a simulation of a fixed amplitude training requiring in-situ computations on passive RRAM crossbars. Performed evaluation of energy dissipation and accuracy considering the effects of RRAM device-to-device variations. Performed a detailed study on the effects of noise inputs to the generator from random sampling by pseudo random noise generators and noise from true random noise generators on the accuracy of the output images.
Neural Network training leveraging Weight Binarization
Created a digit classification neural network using the MNIST dataset and updated weights through gradient descent. Mapped the network weights to binary resistance values in a passive RRAM crossbar array. Performed an evaluation on the trade-off between accuracy and decreased memory and energy consumption.
Deep Learning techniques on GIS Data to predict Social factors - Course Project SOC479
Studied relevant literature implementing deep learning techniques utilizing Geographical Information Systems. Worked on the idea of using neighborhood vector polygons to create prediction systems for domestic violence rates. Discussed the relevance of using prediction systems for areas with under-reported domestic crime rates.
Numerical solution of ODE systems using Runge-Kutta methods - Course Project ME685
Performed higher order explicit schemes to solve systems of differential equations through discretization techniques. Studied the discretization errors and compared the time complexity, consistency and stability of different order schemes. Performed Regula Falsi technique to carry out boundary value problems involved in fourth order Runge-Kutta methods. Numerically simulated velocity distribution of turbulent flow between parallel plates using Navier Stokes equations.
Work Experience
Software and ML Engineer Intern
Larsen and Toubro Infotech
Worked on building a regression model to predict carbon footprint in supply chain management companies. Performed Document Information Extraction using machine learning techniques to use the data for prediction analysis. Worked on DBMS to integrate data from in-memory relational databases to web applications using OData services. Created a workflow to perform Robotic Process Automation in Document Information Extraction service.
Curriculum Vitae
Talks
Neural Correlates of emotional responses to music - CGS609
Not Academics
The wall paintings that I participated in and supervised at the Open Air Theatre, IIT Kanpur.
Some of my recent paintings and sketches.
