EMBER: Autonomous SNN-LLM Hybrid Cognition

🔑 Enhanced Key Takeaways

• •The architecture utilizes dedicated 'person concept cells' within the SNN, drawing inspiration from medial temporal lobe research to facilitate specific person-topic associative learning.
• •The system is implemented in PyTorch and designed for consumer-grade hardware, specifically utilizing an NVIDIA RTX 5070 Ti for the SNN substrate and an RTX 4060 Ti for the embedding model.
• •The researchers are actively investigating safety mitigations including weight auditing, association quarantine, and topology-aware gating to manage potential conflicts between learned associations.

🛠️ Technical Deep Dive

• •SNN Substrate: 220,000-neuron network implemented in PyTorch.
• •Hierarchical Organization: Layer 1 (5,000 sensory neurons) for embedding-to-spike transformation; Layer 2 (150,000 concept neurons) for STDP-based associative fabric; Layer 3 (25,000 category neurons) for abstract grouping; Layer 4 (10,000 meta-pattern neurons) for higher-order regularities.
• •Hardware Allocation: SNN runs on NVIDIA RTX 5070 Ti (16GB); embedding model runs on NVIDIA RTX 4060 Ti (8GB).
• •Encoding: Z-score standardised top-k population code, designed to be dimension-independent.
• •Learning Mechanism: Spike-timing-dependent plasticity (STDP) combined with reward-modulated learning and inhibitory E/I balance.

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline