Multilayer NLAs
use multi-layer inputs for your natural language autoencoders.
cat ./projects.md
Projects
Stuff I've tinkered with.
Machine learning
Fantastic Biases and Where to Find Them
mechanistic study of racial and gender bias in gpt-2 and gpt-neo using linear probes, activation patching, attribution patching, and circuit tracing. steered model outputs to show reduced bias without MMLU degredation.
3D Scene Reconstruction
an end-to-end tour through SfM, MVS, GANs, neural radiance fields, and gaussian splatting for reconstructing scenes from images.
Face Recognition on LFW
a comparative analysis of traditional machine-learning classifier systems for face recognition on the LFW dataset.
Multimodal Stress Counselor
Moshi-based duplex conversational agent that utilizes audio (micrsoft S5) and visual (QuantumConv) contexts to provide real-time stress counseling.
FastSim with CERN
flow-based, diffeomorphic generative models for deep-learning based particle-shower simulation in cylindirical coordinates for high-energy physics.
Core systems
Distributed Program Execution
byzantine fault tolerance in peer-to-peer networks, simulated with OMNeT++.
Huffman Encoder
a mini file-compression system built around Huffman coding.