Exploring the intersection of Artificial General Intelligence, Spiking Neural Networks, and cognitive architectures.
Currently researching adaptive world models...
Spiking Neural Networks
Developed a bio-inspired SNN architecture for real-time pattern recognition with 90% accuracy and 10x energy efficiency compared to traditional ANNs. Uses temporal spike encoding for feature extraction.
Key Finding:
Temporal coding significantly improves energy efficiency while maintaining accuracy
World Models & Embodied Learning
Implemented predictive world models enabling sample-efficient navigation in complex environments. Agent learns environment dynamics through embodied interaction and generalizes to unseen scenarios.
Key Finding:
Embodied learning reduces sample complexity by 5x compared to model-free approaches
Reasoning & Meta-Learning
Developed a meta-learning framework that enables rapid adaptation to new reasoning tasks with minimal examples. Combines gradient-based meta-learning with compositional representations.
Key Finding:
Compositional meta-learning achieves human-level performance on novel reasoning tasks
Consciousness & AI Ethics
Developed computational metrics for measuring consciousness-like properties in AI systems based on Integrated Information Theory and Global Workspace Theory.
Key Finding:
Integrated information correlates with emergent self-awareness behaviors
World Models & Embodied Learning
Novel approach using SOMs to reduce high-dimensional sensory input while preserving semantic structure. Enables efficient online learning in resource-constrained environments.
Key Finding:
SOM-based compression maintains 95% semantic fidelity at 50x compression ratio
Consciousness & AI Ethics
Framework for embedding ethical reasoning capabilities into autonomous systems using deontic logic and value alignment techniques.
Key Finding:
Deontic logic enables consistent ethical decision-making under uncertainty
Temporal spike encoding achieves 10x energy efficiency while maintaining 90% accuracy on pattern recognition tasks.
Embodied learning reduces sample complexity by 5x compared to model-free approaches in navigation tasks.
Compositional meta-learning achieves 92% accuracy on novel reasoning tasks.
Integrated Information Theory correlates with emergent self-awareness behaviors.
90% accuracy, 10x energy efficiency breakthrough
92% few-shot accuracy on novel reasoning tasks
5x sample efficiency improvement in embodied learning
Multi-modal memory integration in SNN-based agents
Experimenting with storing episodic traces in spiking architectures.
Compression of sensory input with self-organizing maps
Investigating low-dim encodings for faster online learning.
If you're working on adjacent problems or have data to share, let's connect.