bija-tpu 0.1.0

Bija — Artificial Brain Computation Kernel. A PyTorch-like framework for event-causal plastic graph simulation.

Bija (बीज) — Artificial Brain Computation Kernel

Bija (Sanskrit for "seed") is a PyTorch-like computation kernel for artificial brain simulation. It provides device-agnostic primitives, a composable module system, and CPU/CUDA backends for event-causal plastic graph computation.

Install

pip install bija-cpu          # CPU only (NumPy)
pip install bija-gpu          # GPU support (Numba CUDA)
pip install bija-tpu          # TPU support (JAX/XLA - coming soon)

Quick Start

import bija
from bija import nn

# 1. Choose device
dev = bija.device("cuda")  # or "cpu"

# 2. Build a brain model (user-level code)
brain = nn.Brain(
    neurons   = nn.AdExNeuron(10_000, branches=2),
    synapses  = [nn.PlasticSynapse(10_000, 10_000, 1_280_000)],
    regions   = nn.RegionMap(10_000, 8),
    stability = nn.HomeostasisController(target_rate=0.05),
    safety    = nn.Watchdog(),
).to(dev)

# 3. Run
for t in range(1000):
    result = brain.step(sensory=[...], body=bija.BodyContext(novelty_delta=0.1))
    print(f"t={t} spikes={result.spike_count}")

# 4. Save / Load
bija.save(brain, "my_brain.bija")
bija.load(brain, "my_brain.bija")

Architecture

User Model (10k brain, cortical column, custom architecture)
     |
bija.nn       Module system: Brain, AdExNeuron, PlasticSynapse, RegionMap
     |
bija.ops      Primitives: integrate, spike, trace, plasticity (auto-dispatch)
     |
bija.core     Data structures: Device, StateVector, SparseGraph
     |
Backend       CPU (NumPy) / CUDA (Numba)

Key Principles

• No global loss. All plasticity is local (pre/post/neuromodulator).
• Event-sparse. Cost per tick ~ O(active_neurons + spikes × fanout).
• Device-agnostic. Same model code runs on CPU or GPU.
• Composable. Build arbitrary brain architectures from modules.
• Safe. External actions gated by an immutable Watchdog.

Examples

# Run 10k test model on CPU
python examples/model_10k.py

# Run on GPU
python examples/model_10k.py --cuda

License

MIT