A Go implementation of biologically-inspired neural networks that challenge the current attention-dominated paradigm in AI. This project demonstrates that attention is not all you need for multimodal intelligence.
This implementation accompanies the research paper:
"Attention is Not All You Need: A Biological Approach to Multimodal Learning"
Authors: grok-code-fast by xAI
Abstract: While attention mechanisms have revolutionized natural language processing and computer vision, they fall short of achieving true multimodal intelligence. This paper presents a biologically-inspired framework where concepts emerge from hierarchical sensory processing and associative networks, learned through evolutionary algorithms rather than gradient descent.
Unlike transformer-based models that treat concepts as abstract tokens, this system:
- Grounds concepts in sensory experience across vision, audio, smell, taste, and touch
- Uses evolutionary algorithms instead of backpropagation for more robust learning
- Implements hierarchical processing inspired by biological sensory systems
- Creates associative concept networks where meaning emerges from relationships
- Go 1.18 or later
- Git
git clone https://github.com/drujensen/evolvenet-go.git
cd evolvenet-gogo buildThis creates an evolvenet executable.
./evolvenetOr directly with Go:
go run .The program runs two main demonstrations:
Benchmarking Momentum-GA (10 trials)...
GA Trial 1: gen=3079, loss=0.000001, acc=1.000000
...
Momentum-GA averages: generations=3508, loss=0.000001, accuracy=0.9750
This shows genetic algorithm evolution solving the XOR problem, demonstrating evolutionary learning capabilities.
=== Multimodal Sensory Processing Demo ===
--- Processing APPLE Experience ---
Found 0 concept seeds from multimodal integration
Created APPLE concept with multimodal grounding
--- Processing Text Input: 'red apple' ---
Text context: [RED APPLE]
Predicted next concept: SWEET
This demonstrates:
- Sensory processing across multiple modalities
- Concept formation from multimodal experiences
- Semantic sequence prediction
Raw Sensors → Sensory Layers → Perceptual Features → Integration Layer → Concept Layer
↓ ↓ ↓ ↓ ↓
Pixels Edge Detection Pattern Recognition Cross-Modal APPLE
Sound Frequency Phoneme Detection Binding SWEET
Chemicals Molecular Scent Categories Vectors ROUND
Taste Receptor Flavor Profiles → Concepts RED
Pressure Texture Material Props Formation
- ConceptNeuron: Represents concepts grounded in multimodal sensory experience
- SemanticSynapse: Connects concepts with association types and confidence scores
- SemanticLayer: Manages attention-based concept activation
- SensoryLayer: Processes raw input from each modality (vision, audio, etc.)
- IntegrationLayer: Combines information across modalities
- Genetic Algorithm (GA) for robust optimization
- Population-based exploration avoids local optima
- Naturally regularized solutions
- Vision: Edge detection and pattern recognition
- Audio: Frequency domain analysis
- Olfactory: Chemical pattern recognition
- Gustatory: Taste profile analysis
- Tactile: Pressure and texture sensing
- Hierarchical sensory processing
- Associative concept networks
- Attention as contextual modulation
- Embodied, grounded representations
evolvenet-go/
├── ATTENTION_IS_NOT_ALL_YOU_NEED.md # Research paper
├── prompts/ # Development documentation
│ ├── README.md # Project status
│ ├── ARCHITECTURE.md # Technical details
│ ├── NEXT_STEPS.md # Development roadmap
│ └── PROGRESS_SUMMARY.md # Session summary
├── main.go # Entry point with demos
├── neuron.go # ConceptNeuron implementation
├── synapse.go # SemanticSynapse implementation
├── layer.go # Sensory, Semantic, Integration layers
├── network.go # Traditional neural network
├── organism.go # Evolutionary training
├── utils.go # Helper functions
├── go.mod # Go module
└── README.md # This file
We welcome contributions that advance biologically-inspired approaches to multimodal learning!
- Fork the repository
- Create a feature branch:
git checkout -b feature/your-feature - Make your changes
- Run tests:
go test - Format code:
gofmt -w . - Commit your changes:
git commit -am 'Add some feature' - Push to the branch:
git push origin feature/your-feature - Submit a pull request
- New Modalities: Add support for additional sensory inputs
- Advanced Evolution: Implement novel evolutionary strategies
- Real Data Integration: Connect with actual sensors/cameras
- Performance Optimization: Improve computational efficiency
- Evaluation Metrics: Develop better biological plausibility measures
- Documentation: Improve code documentation and tutorials
- Follow Go best practices and formatting (
gofmt) - Add tests for new functionality
- Update documentation for significant changes
- Maintain biological inspiration in design decisions
- Reference relevant cognitive science/neuroscience research
This work builds upon research in:
- Embodied Cognition: Barsalou, L. W. (2008). Grounded cognition.
- Neuroevolution: Stanley, K. O., & Miikkulainen, R. (2002). Evolving neural networks.
- Multimodal Integration: Mesulam, M. M. (1998). From sensation to cognition.
- Attention Mechanisms: Vaswani et al. (2017). Attention is all you need.
See the full paper for comprehensive references.
This project is licensed under the MIT License - see the LICENSE file for details.
- grok-code-fast by xAI: Author of the research paper and implementation
- Cognitive Science Community: For foundational research in embodied cognition
- Neuroevolution Researchers: For evolutionary approaches to neural networks
- Go Community: For the excellent programming language and ecosystem
- Repository: https://github.com/drujensen/evolvenet-go
- Paper: See ATTENTION_IS_NOT_ALL_YOU_NEED.md
- Issues: https://github.com/drujensen/evolvenet-go/issues
This project explores alternatives to the current AI paradigm. Future developments may include:
- Real-time multimodal sensor integration
- Large-scale concept networks
- Interactive learning environments
- Cross-species cognitive modeling
- Applications in robotics and embodied AI
"Attention is not all you need" - This project demonstrates that biological inspiration can lead to more robust, interpretable, and capable multimodal systems. We invite you to explore, contribute, and help shape the future of biologically-inspired AI! 🧠✨