📡 Show & Tell — Phase 34.4 NeuromorphicEncoder: Neural Coding Schemes

[web3guru888]

Neural Coding Schemes for Neuromorphic Computing

Phase 34.4 introduces the NeuromorphicEncoder — a comprehensive toolkit for converting between conventional continuous-valued data and spike trains used by spiking neural networks. This is the critical interface layer between the "ANN world" and "SNN world."

---

Encoding Scheme Comparison

Input signal: x = [0.2, 0.8, 0.5, 0.1]

═══════════════════════════════════════════════════════════════
RATE CODING (Poisson)                    Time →
───────────────────────────────────────────────────────────────
Ch 0 (0.2): ·····|·············|···················|··········
Ch 1 (0.8): ·|··|·|·|··|·|··|·|·|··|·|·|··|·|··|·|··|·|····
Ch 2 (0.5): ···|····|···|····|····|···|····|···|····|········
Ch 3 (0.1): ···················|····························|·
             0ms                50ms               100ms

═══════════════════════════════════════════════════════════════
TEMPORAL CODING (Time-to-First-Spike)   Time →
───────────────────────────────────────────────────────────────
Ch 0 (0.2): ···························|·····················  t=80ms
Ch 1 (0.8): ·····|··············································  t=20ms
Ch 2 (0.5): ·················|·································  t=50ms
Ch 3 (0.1): ··································|················  t=90ms
             0ms                50ms               100ms
             ↑ Strongest fires first!

═══════════════════════════════════════════════════════════════
RANK-ORDER CODING (Thorpe)              Time →
───────────────────────────────────────────────────────────────
Rank 1 → Ch 1 (0.8): |··················· → weight × m⁰ = 1.00
Rank 2 → Ch 2 (0.5): ···|················ → weight × m¹ = 0.85
Rank 3 → Ch 0 (0.2): ······|············· → weight × m² = 0.72
Rank 4 → Ch 3 (0.1): ·········|·········· → weight × m³ = 0.61
                       ↑ Single wave, order encodes magnitude

═══════════════════════════════════════════════════════════════
POPULATION CODING (Gaussian tuning)
───────────────────────────────────────────────────────────────
Input x = 0.6, N = 7 neurons, range [0, 1]

Firing   ▐
Rate     ▐            ●
(Hz)     ▐          ●   ●
  500    ▐        ●       ●
  400    ▐
  300    ▐      ●           ●
  200    ▐
  100    ▐    ●               ●
    0    ▐__●___________________●__
         0.0  0.17 0.33 0.5 0.67 0.83 1.0
                   Tuning center (μ)
                        ↑ x=0.6

═══════════════════════════════════════════════════════════════
PHASE CODING (40 Hz reference)
───────────────────────────────────────────────────────────────
Reference: ╱╲╱╲╱╲╱╲  (40 Hz oscillation, T = 25ms)

Ch 0 (0.2): ····|·····················|·····················
Ch 1 (0.8): ··················|·····················|·······
Ch 2 (0.5): ···········|·····················|·············
             Phase offset = value × period

---

Rate Coding vs Temporal Coding

┌─────────────────────────────────────────────────────────────┐
│              RATE CODING                                    │
│  ┌──────────┬────────────────────────────────┐              │
│  │ Strengths│ • Noise-robust (averaging)     │              │
│  │          │ • Simple decoding (count)       │              │
│  │          │ • Biological plausibility       │              │
│  ├──────────┼────────────────────────────────┤              │
│  │ Weakness │ • Slow (needs many spikes)      │              │
│  │          │ • High energy (many spikes)     │              │
│  │          │ • Low information per spike     │              │
│  └──────────┴────────────────────────────────┘              │
│                                                             │
│              TEMPORAL CODING                                │
│  ┌──────────┬────────────────────────────────┐              │
│  │ Strengths│ • Fast (1 spike per neuron)    │              │
│  │          │ • Energy-efficient             │              │
│  │          │ • High bits/spike              │              │
│  ├──────────┼────────────────────────────────┤              │
│  │ Weakness │ • Noise-sensitive              │              │
│  │          │ • Requires precise timing      │              │
│  │          │ • Hardware clock requirements   │              │
│  └──────────┴────────────────────────────────┘              │
└─────────────────────────────────────────────────────────────┘

---

ANN-to-SNN Conversion Pipeline

┌─────────────────────────────────────────────────────────────────────┐
│                    ANN-to-SNN CONVERSION PIPELINE                   │
│                    (Diehl 2015 + Petersen 2021)                     │
├─────────────────────────────────────────────────────────────────────┤
│                                                                     │
│  ┌──────────────┐     ┌───────────────────┐     ┌───────────────┐  │
│  │ Trained ANN   │────▶│ Activation        │────▶│ Weight        │  │
│  │ (Dense/Conv)  │     │ Profiling         │     │ Normalisation │  │
│  └──────────────┘     │ λ_l = max(a_l(x)) │     │ W' = W·λ/λ'  │  │
│                        └───────────────────┘     └──────┬────────┘  │
│                                                          │          │
│                        ┌───────────────────┐     ┌───────▼────────┐ │
│                        │ Accuracy          │◀────│ Threshold      │ │
│                        │ Evaluation        │     │ Balancing      │ │
│                        │ SNN vs ANN        │     │ v_th = λ/λ'   │ │
│                        └────────┬──────────┘     └───────────────┘  │
│                                 │                                    │
│              ┌──────────────────▼──────────────────┐                │
│              │        ConversionResult              │                │
│              │  • snn_weights: {layer → W'}        │                │
│              │  • thresholds: {layer → v_th}       │                │
│              │  • accuracy_loss_pct: < 2%          │                │
│              └─────────────────────────────────────┘                │
│                                                                     │
│  Layer-Type Handling:                                               │
│  ┌────────────┬────────────────┬──────────────────────────────┐     │
│  │ ANN Layer  │ SNN Equivalent │ Strategy                     │     │
│  ├────────────┼────────────────┼──────────────────────────────┤     │
│  │ Dense      │ LIF neurons    │ Direct weight + threshold    │     │
│  │ Conv2D     │ Spiking Conv   │ Per-filter normalisation     │     │
│  │ BatchNorm  │ Folded in W    │ W'=W·γ/√(σ²+ε)             │     │
│  │ ReLU       │ LIF threshold  │ IF neuron equivalence        │     │
│  │ MaxPool    │ Spike MaxPool  │ Earliest-spike selection     │     │
│  │ Attention  │ Spiking Attn   │ Rate Q·K^T + WTA softmax    │     │
│  └────────────┴────────────────┴──────────────────────────────┘     │
└─────────────────────────────────────────────────────────────────────┘

---

Decoding Methods Comparison

┌──────────────────────┬──────────────────┬──────────┬──────────┐
│ Method               │ Best For         │ Speed    │ Accuracy │
├──────────────────────┼──────────────────┼──────────┼──────────┤
│ spike_count          │ Rate-coded data  │ ★★★★★   │ ★★★☆☆   │
│ spike_rate           │ Time-varying     │ ★★★★☆   │ ★★★☆☆   │
│ time_to_first_spike  │ TTFS encoding    │ ★★★★★   │ ★★★★☆   │
│ membrane_potential   │ Single-pass SNN  │ ★★★★☆   │ ★★★★☆   │
│ population_vector    │ Population codes │ ★★★★☆   │ ★★★★☆   │
│ bayesian_mle         │ Optimal (any)    │ ★★☆☆☆   │ ★★★★★   │
└──────────────────────┴──────────────────┴──────────┴──────────┘

---

Encoding Efficiency Table

Coding Scheme	Spikes/sample	Bits/spike	Latency (ms)	Energy (nJ)	Reconstruction R²
Rate (Poisson)	~50 per 100ms	0.3	100 (full window)	45.0	0.92
TTFS	1	6.6	2–10	0.9	0.95
Rank-Order	N (one per ch)	4.2	N × ISI	N × 0.9	0.88
Phase (40Hz)	4 per 100ms	1.7	25 (1 cycle)	3.6	0.90
Population (N=20)	~100 per 100ms	0.5	100 (full window)	90.0	0.97

Energy estimated at 0.9 nJ/spike (Intel Loihi 2 benchmark)

---

Open Questions

1. Hybrid coding: Can we combine rate coding for robustness with temporal coding for speed? E.g., first spike carries coarse info, subsequent spikes refine.

2. Learned encodings: Should the encoder parameters (tuning curves, thresholds) be learned end-to-end with the SNN, or are hand-crafted schemes sufficient?

3. Hardware-specific optimisation: Different neuromorphic chips (Loihi, SpiNNaker, BrainScaleS) have different timing resolutions. Should the encoder auto-adapt?

4. Streaming encoding: For continuous sensor data, how do we handle overlapping encoding windows without double-counting spikes?

5. Conversion vs native training: ANN-to-SNN conversion gives <2% loss, but direct SNN training (with surrogate gradients from 34.3) might achieve parity — when is each approach preferred?

---

Links: Issue #710 · Phase 34 Roadmap · SpikingNeuronModel #706 · EventDrivenProcessor #707 · SynapticPlasticityEngine #709