Skip to content

CLAUDE.md

Latest commit

 

History

History
53 lines (45 loc) · 1.93 KB

CLAUDE.md

File metadata and controls

53 lines (45 loc) · 1.93 KB
Raw

PDNA — Pulse-Driven Neural Architecture

Project Overview

Codename "Heartbeat". A neural network with an intrinsic oscillatory "pulse" that maintains and evolves state independently of input. Built on Liquid Time-Constant (LTC) networks augmented with learnable oscillatory dynamics.

Tech Stack

  • Python 3.10+ managed with uv
  • PyTorch 2.x
  • ncps (Neural Circuit Policies) for LTC/CfC base
  • torchdiffeq for ODE solving
  • hatchling build backend

Project Structure

src/pdna/          # main package
tests/             # pytest tests
configs/           # model/experiment configs
scripts/           # training & eval scripts
notebooks/         # analysis notebooks
data/              # datasets (gitignored)
docs/              # project plans and documentation

Commands

  • uv sync — install dependencies
  • uv run pytest — run tests
  • uv run pytest tests/test_foo.py -v — run specific test
  • uv run python scripts/<script>.py — run a script

Conventions

  • Source code lives in src/pdna/
  • All model variants (A–F) share identical hyperparameters except architectural differences
  • Use layered git commits — one logical change per commit
  • Config-driven experiments: hyperparameters in configs/, not hardcoded
  • Tests mirror source structure: src/pdna/models/foo.py → tests/test_foo.py

Key Architecture

  • Variant A: Baseline LTC (control)
  • Variant B: LTC + Noise (random perturbation control)
  • Variant C: LTC + Pulse (structured oscillation only)
  • Variant D: LTC + SelfAttend (self-attention on own state only)
  • Variant E: Full PDNA (pulse + self-attend)
  • Variant F: Full + Idle (pulse + self-attend + idle ticks)

Core Equation

τ(x) · dh/dt = -h + f(h, x; θ) + α · pulse(t, h) + β · self_attend(h)
pulse(t, h) = A · sin(ω · t + φ(h))
self_attend(h) = W_self · σ(h)

Task Plan

See docs/task_plan.md for the full 13-task execution plan with dependencies.