Welcome to Silicon Embryo, my personal research project exploring a different path to Artificial General Intelligence (AGI).
Most modern AI (like Large Language Models) relies on massive parameter counts and backpropagation. But biological brains work differently: they use highly efficient structures (like small-world topologies) and continuous online learning rules (like predictive coding).
This project is an attempt to build a Spiking Neural Network (SNN) based on the philosophy: "Structure for Compute, Memory for Intelligence." Instead of just building a simulator, the goal is to "grow" a brain—using simulated developmental genes and letting intelligence emerge from continuous sensory input and learning.
I'll be honest, this is highly experimental. Here's what has been successfully built and verified so far:
- A production-grade LIF Engine (
src/engine.py): A GPU-accelerated Leaky Integrate-and-Fire engine using Compressed Sparse Row (CSR) matrices for synaptic propagation. It can simulate ~670K neurons and 168M synapses at 256 FPS on a consumer laptop (RTX 4060). - Predictive Coding Learning Rule: I recently replaced standard STDP (which failed at spatial pattern storage in sparse networks) with a Predictive Coding rule. The network continuously minimizes prediction errors, learning online at every step.
- Hyperdimensional Computing (HDC) codec (
src/hdc.py): Used as the sensory/motor interface to map text to 10,000-dimensional binary spike patterns. - Developmental Genes (
src/genome.py): A system that "grows" the network topology from scratch.
I am currently pivoting from a traditional "train-then-test" simulator mindset to a true "Living Brain" loop. The network will run continuously, perceiving, thinking, learning, and acting without discrete epochs.
Make sure you have PyTorch installed (CUDA is highly recommended).
python example_run.pyThis is a raw, ongoing experimental project. You'll find a lot of rough edges. Feedback, ideas, and discussions are highly welcome!