Energy efficient hardware for AI!
Where p-bits have primarily been used for optimization & sampling problems, I extended their use to standard feed forward AI models! There are lots of interesting possibilities for building energy efficient AI platforms with pbits. (ps. we're already on it at Ludwig Computing)
Protein Structure Prediction with p-bits!
I've always been fascinated by biological systems, even minored in biology in undergrad. So I tried to see if p-bits could help solve problems in biology. I arrived at an MCMC algorithm (termed Alternating Metropolis-Hastings) for optimizing a protein's 3D geometry given its sequence of amino acids.
FPGA Architectures for Probabilistic Neural networks
Co-built FPGA and microcontroller architectures implementing p-bits to solve np-hard optimization problems, sampling problems, and emulate certain quantum systems. Built both synchronous (clock based) and asynchronous (no clock) architectures. Both designs won 1st place poster awards, 2 years back to back!
Sync Paper
Sync Poster
Async Paper
Async Poster
Microcontroller Paper
Edge AI for Automotive Applications
Combined a set of advanced driver facial recognition models into a single, low-powered asic solution for a tech startup.
I coded a fast C++ application that interfaced with vision models to detect passengers, their age, and presence of mask in a car cabin. Masks were all the rage during covid-19.
ML for Triaging Skin Cancer Cases
Built a system of CNNs that can quickly triage cases of skin cancer from most to least severe, using Keras and multiple fine-tuned VGG16 models.