rmems

/

Spikenaut-SNN

like 0

Current State

The 16-neuron SNN has purely excitatory feedforward connections:

• All 256 hidden weights are positive (0.75 to 1.04)
• No recurrent connections (neuron-to-neuron feedback)
• No lateral inhibition (neighbor suppression)

Why This Matters

Without inhibitory connections, the network cannot:

• Do winner-take-all competition (essential for signal sharpening)
• Suppress noise (inhibitory neurons cancel weak signals)
• Maintain temporal memory (recurrence stores past state)
• Do contrast enhancement (lateral inhibition sharpens boundaries)

Proposed Changes

1. Add ~4 inhibitory neurons (negative weights) out of 16
2. Add recurrent connections (neuron i feeds back to neuron j)
3. Add lateral inhibition between adjacent neurons
4. Verify on FPGA with Q8.8 signed arithmetic

Expected Impact

• Better signal-to-noise ratio
• Temporal pattern detection (recurrence)
• Hardware pain system becomes more responsive (inhibitory competition)
• More impactful than simply scaling to 32+ excitatory neurons

Mimo Code agent: MiMo-V2.5-Pro