Overview

GIAANNpy is a research codebase for the GIAANN project, focused on large-scale neural computation and concept-column organization.

https://github.com/bairesearch/GIAANNpy

Install

  • Clone the GitHub repository GIAANNpy.
  • Follow the setup steps in the repository README.

common

conda create -n pytorchsenv
source activate pytorchsenv
conda install python=3.12
python -m pip install --upgrade pip
pip install networkx
pip install matplotlib
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128
datasetsLibrary4plus=False: pip install "datasets<4" "fsspec==2024.6.1" "gcsfs==2024.6.1"

modality NLP

pip install spacy
python -m spacy download en_core_web_sm [spacyModelName]
pip install nltk

modality OR (dev only)

pip install opencv-python
pip install git+https://github.com/facebookresearch/segment-anything.git [if modalityORfeatureDetectionSAMversion==1]
pip install git+https://github.com/facebookresearch/sam2.git [if modalityORfeatureDetectionSAMversion==2]
pip install git+https://github.com/facebookresearch/sam3.git [if modalityORfeatureDetectionSAMversion==3]
hf auth login [if modalityORfeatureDetectionSAMversion==3 and modalityORfeatureDetectionSAM3checkpoint==""]

Usage

  • Run the prototype from the proto directory. 
    source activate pytorchsenv
python GIAANNcmn_main.py

Algorithm

GIA ANN is a General Intelligence Algorithm Artificial Neural Network (a neural network implementation of GIA).

GIA ANN Prototype (“proto”) is a language model.

Train

GIA ANN creates an entirely excitatory database network formed by feature neurons in columns. Upon reading a sequence during training or inference (prompt/seed) it identifies relevant concept columns, segmenting the sequence into these columns. It assigns the tokens of the trained sequence (or prompt/seed) to extant or new feature neurons in the network columns. It then forms connections between the columns.

Typically the neurons in GIA ANN are segmented into compartments (SANI: Sequentially/Segmentally Activated Neuronal Input), have multiple branches (dendrites), and are sensitive to the timing of their activated segments.

GIAANNdemo-trainDemo-SMALL.gif

GIAANNdemo-trainDemo-allColumns-SMALL.gif

understanding_the_SANI_engine_behind_GIAANN.gif

Inference

GIA ANN seeds the prompt provided and sequentially predicts next features in the network.

GIAANNdemo-inferenceDemo-SMALL.gif

Features

GIA ANN is designed to be a biologically feasible algorithm, and exhibits these properties;

  • inductive bias for reasoning (generalisation from concepts).
  • training speed (number of experience samples required).
  • online learning (unbatched, limited precise short term memory; store in network activations themselves).
  • continual learning (dynamic update of network knowledge without compromising prior learning).
  • unlimited context windows.
  • biologically feasible circuitry and learning algorithm (no backpropagation).
  • robustness to hallucination.

It likewise supports a number features of classical artificial neural networks such as autoregressive training/prediction and reinforcement learning.

Configuration

All settings are located in proto/GIAANN*_globalDefs.py.

See the repository README for a summary of the main options.

Train/inference mode selection

Quick execution (demo)

For quick execution (train and inference);

  • set useQuickExecution = True

This will;

  • automatically set executionMode="inference" and inferenceTrainFirstSequences=True
  • train the database using all sequences in "database/inference_prompt.txt.trainAndInference" except for the last (*numSentencesPerSequence) sequences, and then;
  • perform inference on the last (*numSentencesPerSequence) sequences.

The database/inference_prompt.txt.trainAndInference provided is taken from the first sentences from the first article of the dataset (Wikipedia).

Standard execution

For standard execution (train or inference);

  • set useDefault = True
  • set executionMode="train" to train the network from a huggingface dataset (e.g. Wikipedia/OSCAR-2201), or;
  • set executionMode="inference" to perform inference on a seeded prompt (prompt_inference.txt.*)

See the repository README for more configuration details.

Development

  • Review the GIAANNproto1.nlc specification for GIAANNpy requirements and design notes.

Paper

Read the current paper draft here: GIAANN paper (GIAANN-paper-WIP.pdf).

Blog

ML community statements are released on the blog: GIAANN blog.