This repository contains the research paper "Bridge Neural Networks: Direct Neural Pathways for External Knowledge Integration" which introduces the "Abhay Architecture" for external knowledge integration in language models.
figures/: SVG diagrams for the paperlatex/: LaTeX implementation of the paper using the Tau templatemarkdown/: Markdown version of the paperwebsite/: Academic website using Jekylloutput/: Generated PDFs and other outputs
This project includes a build system that can generate the paper in different formats:
- For Markdown PDF: Pandoc and XeLaTeX
- For LaTeX PDF: A complete LaTeX distribution (TeXLive, MikTeX, etc.)
- For Website: Jekyll and Ruby
- For SVG Conversion: Inkscape or rsvg-convert
# Generate PDF from Markdown
./build.sh markdown
# Set up academic website
./build.sh website
# Generate academic PDF using LaTeX
./build.sh latex
# Generate all formats
./build.sh all
# Show help
./build.sh helpThe paper includes the following figures:
figure1_bridge_architecture.svg: Bridge Neural Network Architecturefigure2_information_flow_comparison.svg: Comparison between RAG and Bridge Neural Networksfigure3_pruning_bridge_allocation.svg: Pruning-Guided Bridge Allocationfigure4_mathematical_framework.svg: Mathematical Frameworkfigure5_training_curriculum.svg: Training Curriculum
The "Abhay Architecture" refers to the proposed bridge neural network design that creates direct neural pathways for external knowledge access. Key components include:
- Bridge detector neurons
- Neural query encoder
- External knowledge service
- Response integrator
This architecture allows language models to access external knowledge without consuming context window tokens, maintaining full reasoning capacity while adding factual accuracy.
The Markdown PDF version is a simpler format with a clean design, suitable for initial reviews and sharing.
The LaTeX PDF uses the Tau template which provides a professional academic layout with proper formatting for equations, figures, and references.
The website version is built using Jekyll and provides an online accessible version of the research with interactive elements.
Pankaj Doharey
ZenDiffusion.art
pankaj@zendiffusion.art
MIT License