Project Dates: Apr - May 2026
- A highly-optimized Multilayer Perceptron (MLP) neural network built to classify grayscale handwritten digits
- Optimizations achieved a 156X training + testing speedup over the serial baseline while maintaining a classification accuracy of 96.57%
- The MLP was written in C, contained 4 layers + 80,000 parameters, and was optimized for the Intel i9-14900
- Optimizations included loop unrolling, AVX vectorization, OpenMP multithreading, and matrix operation fusion
- Full report in repo root
- Resources: 3Blue1Brown + Michael Nielsen's Neural Networks and Deep Learning
- Coding & Compiling: VSCode + GCC
- Optimization libraries: Intel AVX Vector extension + OpenMP Multithreading
- Claude Code
- Synthesized video and textbook resources to construct serial baseline backpropagation and feedforward algorithms within 3 days from scratch
- Targeted optimization hotspots according to Amdahl’s Law by prioritizing the largest activation/weight/bias matrices and their functions
- Directed brainstorming sessions with partner, accelerating integration of serial/parallel optimizations
- Achieved 78X speedup using techniques from class, then utilized Claude Code to identify 2 additional optimization hotspots, improving to 156X speedup
- Recorded optimizations in 22-page final report and streamlined 9-min presentation to the full class
This codebase consists of three separate code architectures:
- optimized contains our most efficient code. Navigate to the directory and compile/run on Linux systems with "make run".
- serial_baseline contains our code derived from first principles. Navigate to the directory and compile/run on Linux systems with "make run".
- python_reference_code contains the neural network from Michael Nielsen's "Deep Learning and Neural Networks". Navigate to the directory/src and compile/run on Linux systems with "python3 script.py".