Projects
Open Source Projects
Generative AI
- Diffusion Model
- Diffusion models are a class of generative models that learn to reverse a gradual noising process to produce high-quality data, such as images, from random noise. Inspired by probabilistic principles, they are increasingly popular in AI image generation. I have coded a basic diffusion model following the Denoising Diffusion Probabilistic Model (DDPM) paper, which is now publicly accessible.
https://github.com/JDan-16/Diffusion_model1
- Stable Diffusion
- Stable Diffusion is a generative model designed to create highly detailed images from textual input. It operates by progressively refining random noise into coherent visuals, leveraging diffusion techniques for efficient and realistic image synthesis. I have implemented the architecture of Stable Diffusion v1.5 in PyTorch.
https://github.com/JDan-16/Stable_Diffusion_v1_5
More open source projects coming soon...
Corporate Endeavors
Hexagon Tensor Processor
Company: Qualcomm
I contribute to the Hexagon Tensor Processor (HTP) team, focusing on deep learning operation kernels and model graph optimization. HTP is the neural processing unit (NPU) designed from the ground up for generative AI, integrated into all Snapdragon chipsets, including the Snapdragon X Series. This role requires me to stay current with state-of-the-art deep learning and generative AI models. To know more about HTP, see link.
Life Sciences Project
Company: Ceremorphic, Inc.
I led a software team dedicated to constructing an AI-powered software framework tailored for in-house custom hardware within the life sciences domain. I developed proficiency in integrating diverse deep learning models, including Vision models, Language models (transformer, BERT), other Generative models (GAN, VAE), Graph based models (GCN), and more.
Furthermore, I have been responsible for architecting and designing a compiler from inception, tailored for analog compute circuits in the realm of life sciences applications.
Deep Learning Accelerator
Company: Ceremorphic, Inc.
Inventing a power-efficient method, I optimized convolution and max-pool operations in a combined way, achieving 3-4x power efficiency without modifications during backpropagation—resulting in a filed patent. My role extended to designing algorithms for quantization-aware training of neural networks, minimizing machine cycles. Additionally, I contributed to deploying posit mathematics for neural network operations in hardware.
AI powered Security and other Projects
Company: Ceremorphic, Inc.
Under the mentorship of Prof. Boris Murmann, I explored microarchitectural side-channel attacks (SCA) like Flush+Reload, Flush+Flush, Prime+Probe, etc., developing AI-inspired algorithms to mitigate such threats. My comprehensive efforts encompassed building and filing numerous intellectual properties (IPs) and patents. Other projects include exploration of fundamental quantum machine learning concepts and circuits.
Academic Projects
IIT Gandhinagar
M.Tech Thesis: An Approach towards Building Energy-Efficient Architectures for Neural Networks
Under the guidance of Dr. Joycee Mekie, I derived a formula for the minimum exponent size in floating-point representation for neural network weights, ensuring accuracy, and gave mathematical explanation for the relationship existing between mantissa and depth of a model. Additionally, proposed methods for bit error resilience in SRAM-based floating-point representation.
Other minor projects include implementation of an elementary pipelined processor and a multilayer perceptron neural network on an FPGA platform.
Assam Engineering College
B.E. Final Year Project: Prediction of Water Usage Based on Weather Data Pattern Using Neural Network
Collaborating with Dr. Rashi Borgohain and Mr. Tanmoy Goswami, I designed a Raspberry Pi-based device with interfaced sensors to collect soil type and daily water usage data. Based on the collected data and weather data from other sources, I designed and trained a neural network. Leveraging this neural network, I wrote RESTful APIs, enabling accurate water usage predictions via a user-friendly web interface.