Proposal for a modular AI architecture focused on creativity

CreaNet - Creativity-Oriented Modular Transformer Architecture

Overview

CreaNet is a novel modular transformer architecture designed specifically to enhance AI creativity. Unlike traditional transformer models that are generalists, CreaNet incorporates specialized modules focusing on narrative structuring, stylistic control, analogical reasoning, and symbolic generation. It also integrates dynamic associative memory and recursive idea refinement loops to simulate human-like creative processes.

This repository contains:

  • Whitepaper detailing the architecture and motivations
  • Prototype implementation in PyTorch
  • Pitch deck for presentations and workshops
  • Sample scripts for testing and experimentation

Features

  • Modular Attention Blocks: Specialized modules for different creative faculties
  • Dynamic Associative Memory (DAM): Sparse-key attention to connect semantically distant ideas
  • Recursive Idea Refinement Loop (RIRL): Iterative self-refinement for enhanced novelty
  • Multi-Modal Support: (Planned) Embedding layers for images, sound, and text
  • Open Source & Extensible: Designed for easy experimentation and collaboration

Getting Started

Requirements

  • Python 3.8+
  • PyTorch 1.12+
  • Other dependencies (listed in requirements.txt)

Installation

Clone the repository:

git clone https://huggingface.co/cadamon/creanet_prototype
cd creanet
pip install -r requirements.txt
1 Like