ViciousSquid/Dosidicus

Raise a neural network as a pet - a transparent cognitive sandbox where a digital squid learns, grows, and rewires itself via Hebbian learning & Neurogenesis

What is Dosidicus?

Dosidicus electronicus is an open‑source, Python‑based simulation that lets you raise a digital pet squid whose brain is a tiny, fully‑transparent neural network. The project is positioned as an educational neuro‑tool and a simulation game: you feed, clean and interact with the squid while watching every neuron fire, every synapse change, and every new neuron appear as it learns.

Core ideas

Eight‑neuron starter brain – seven “drives” (hunger, curiosity, etc.) plus a single sensory input. The network grows new neurons only when the existing structure cannot solve a task.
Biologically‑inspired learning – Hebbian updates and Spike‑Timing‑Dependent Plasticity (STDP) are implemented in pure NumPy (see src/plasticity.py). Each weight change records what caused it, why it happened and which experience triggered it.
Full provenance – the system logs a human‑readable explanation for every synaptic modification and for every newly created neuron, so you can ask questions like “Why did this weight go from 0.31 to 0.47?”.
Dual‑memory system – short‑term and long‑term stores let the squid retain experiences over time.
Extensible engine (STRINg) – a custom simulation core built from scratch in NumPy, with no TensorFlow/PyTorch dependencies. The same engine powers the GUI, a head‑less trainer, and the test suite.

How you use it

Clone & set up – git clone … && python -m venv venv && pip install -r requirements.txt.
Run – python main.py launches a PyQt5 window where you can interact with the squid.
Docker – pre‑built images are provided for a GUI version (requires X11/WSLg) and a head‑less CLI trainer that can output a trained brain JSON file.
Customization – the Brain Designer wiki page explains how to add or delete neurons and behaviours; pre‑made brain configs live in custom_brains/.
Sharing – saved brain files can be exported and uploaded to the SQUID‑EXCHANGE branch so others can explore your creature’s neural history.

Who might find it useful?

Students & hobbyists who want a concrete, visual illustration of how simple spiking‑type learning rules can produce adaptive behaviour.
Researchers looking for a lightweight, fully‑observable sandbox to prototype ideas about structural plasticity or credit assignment without the overhead of large deep‑learning frameworks.
Game designers interested in emergent AI agents that can be inspected and tweaked.

What the repo contains

src/ – the STRINg engine, plasticity algorithms, brain‑designer utilities, UI widgets, and plugin framework.
main.py – entry point for the PyQt5 GUI.
headless/ – command‑line trainer that runs the same engine without graphics.
custom_brains/ – example network configurations.
Dockerfiles with two build targets (gui and headless).
Extensive wiki (57 pages) covering the manifesto, engine overview, brain design, STDP theory, and changelog.

TL;DR

Dosidicus is a transparent neural‑network sandbox wrapped in a cute virtual‑pet game. It lets you watch a tiny spiking network learn in real time, inspect every weight change, and even grow new neurons, all with a simple Python/NumPy codebase and an optional PyQt5 GUI.

All information above is taken directly from the repository’s README and linked wiki pages.