PI-WNN: Physics-Informed Weightless Neural Network Engine

A simple memory network teaching physics to tiny hardware, bypassing power-hungry models.

Comment

Inspiration

Traditional deep learning models are power-hungry, opaque black boxes that frequently violate basic physical laws because continuous equations have trust issues with floating-point approximations. I wanted to see if we could completely side-step heavy multi-billion parameter networks. The goal was to build an ultra-fast, hardware-bounded architecture that enforces physical constraints directly inside raw, bit-packed memory arrays.

What it does

WiSARDINES is a physics-informed weightless neural network (PI-WNN). Instead of shifting floating-point weights via backpropagation, it maps tracking telemetry into deterministic binary tuple addresses across RAM discriminators.

It handles complex data structures using both hard and soft physical constraints:

  • Hard Constraints: Telemetry that violates known conservation laws (like mass-tier routing or energy ceilings) is masked entirely at the hardware boundary and returns a strict (-1).
  • Soft Constraints: Unseen binary addresses receive logical priors using sparse Laplace smoothing rather than multiplying the dense RAM footprint.

How I built it

The architecture is split into a robust testing pipeline and a bare-metal execution core:

  • The Physics & AI Sandbox: Written in Python, featuring custom telemetry encoders, real-world data pipelines, and full continuous baselines to challenge GBTs and PyTorch architectures.
  • The Simulation Engines: Built-in tools including an IsingTransitionSimulator for Metropolis frequencies and a chaotic EnergyConstrainedDoublePendulum that maps step updates back to an initial energy budget.
  • The Embedded Core: A zero-dependency, strict C99 module (embedded_core.c) optimized for zero-overhead, highly predictable execution loops on low-power microcontrollers.
  • Quantization: Employs Gray-coded thermometer quantization to maintain a strict binary Hamming distance of exactly (1) between adjacent interior data ranges.

Challenges I ran into

Hardware has a brutal sense of humor, and exponential growth is the punchline. When scaling a tracking tuple size from 8 to 16 bits, the lookup table doesn't double in size—it multiplies the dense RAM footprint by exactly 128 times ((2^{16}) vs (2^8)).

To keep WiSARDINES viable for tiny embedded chips without crashing the memory ceiling, I bypassed dense matrix allocations completely. I designed a highly optimized address-partitioning pipeline utilizing cuckoo_table.py for exact address storage, switching the engine into a sparse, hashed storage footprint that only scales with activated states.

What I learned

When you write physics-informed code, optimization becomes the ultimate tool for engineering equity. By seeding algebraic data priors (like identity, bit-reversal, and complement orbits) straight into the memory structure, you gain material accuracy lifts on heavy physics benchmarks—like the 1-million-row UCI HIGGS dataset—without needing a massive corporate infrastructure budget to back you up.

Built With

  • c
  • c99
  • data-structures
  • embedded-systems
  • gray-code
  • hashing-algorithms
  • machine-learning
  • physics-informed-ai
  • python
  • telemetry
  • unit-testing
Share this project:

Updates

posted an update —

Project Launch: Unleashing WiSARDINES (PI-WNN)

I'm incredibly excited to lock in the initial release of WiSARDINES (Physics-Informed Weightless Neural Network Engine) for the Global Innovation Build Challenge!

Instead of throwing massive, power-hungry deep learning models at physics problems, I built a framework that teaches physical boundaries directly to weightless, bit-packed lookup tables running on tiny edge hardware.

What's packed into this release:

  • embedded_core.c: A dependency-free, strict C99 core optimized for zero-overhead execution loops on microcontrollers.
  • Cuckoo Table Address Storage: Solved the exponential RAM footprint growth of weightless logic using optimized sparse dictionary mapping.
  • Symmetry & Invariant Constraints: Hard-masking for physical boundaries (like mass tiers) and soft-score logical priors using Laplace smoothing.
  • Empirical Validation: Proven track records streaming the 1-million-row UCI HIGGS physics archive.

The complete codebase, architectural profiles, and simulation scripts are live on GitHub. Take a look under the hood: https://github.com/t4rxnn-devel/WiSARDINES-Physics-Informed-Weightless-Neural-Network-PI-WNN-

Log in or sign up for Devpost to join the conversation.

Submission history