Skip to content
D-CSIL

Research Domain

Novel Neural Architectures

Hybrid recurrent models (HELIX H-E-R blocks), selective state-space models, Hebbian plasticity with neuromodulation and sleep-style consolidation (SynapNet / HibbieFormer), and physics-inspired field layers (FieldTransformer).

MEASUREDCURRENT EVIDENCE LEVEL

Research Question

Do recurrent hybrids and biologically inspired plasticity offer practical advantages over standard transformers at scales one machine can train?

HELIX models stack HarmonicStateLayer → ElasticAttentionGate → ResonantExpertRouting blocks — a recurrent/sparse-attention/MoE hybrid trained entirely on Apple Silicon. HELIX-200M (209M params) is mid-training at validation PPL 39.4.

The SynapNet/HibbieFormer line trains with explicit plasticity phases — hebbian, synaptic scaling, and pruning on a 50/30/20 schedule with adaptive neuromodulation. The completed TinyStories run logged 21,476 steps across 215 plasticity cycles.

Experiments in this domain