❓ Q&A — Phase 34.4 NeuromorphicEncoder

[web3guru888]

Common Questions about Phase 34.4 — NeuromorphicEncoder

A collection of Q&A pairs covering neural coding schemes, ANN-to-SNN conversion, and spike train encoding/decoding.

---

Q1: What are the trade-offs between rate coding and temporal coding?

Rate coding (Poisson) represents information via average firing frequency over a time window. It's noise-robust because errors in individual spike times average out, and decoding is trivial (count spikes). However, it requires many spikes — and hence more time and energy — to achieve precision. A 100ms window at 200 Hz gives ~20 spikes per neuron.

Temporal coding (TTFS, rank-order, phase) encodes information in precise spike timing. A single spike can carry ~6-7 bits if the timing resolution is fine enough. This yields ultra-low latency (one spike per neuron) and extreme energy efficiency, but requires precise clocks and is sensitive to jitter. Our TimeToFirstSpikeEncoder achieves 0.95 R² reconstruction with just 1 spike per channel.

Rule of thumb: Use rate coding when noise tolerance matters (sensor data, safety-critical); use temporal coding when speed/energy is paramount (edge inference, robotics).

---

Q2: How does ANN-to-SNN conversion preserve accuracy to within 2%?

The Diehl et al. (2015) + Petersen et al. (2021) pipeline works in three phases:

1. Activation profiling: Run calibration data through the ANN, recording the 99.9th percentile activation per layer (λ_l). This avoids outliers skewing the normalisation.

2. Weight normalisation: Scale weights so that the maximum expected input to each layer is 1.0: W'_l = W_l × (λ_{l-1} / λ_l). This ensures SNN neurons operate in their linear regime.

3. Threshold balancing: Set v_thresh = 1.0 for all layers (post-normalisation). The key insight is that a ReLU neuron with output a = max(0, Wx + b) is equivalent to an integrate-and-fire neuron that fires when V_mem ≥ v_thresh, with the firing rate proportional to the ReLU activation.

The <2% accuracy loss comes from temporal quantisation — the SNN needs enough timesteps to accumulate spike counts that faithfully represent the continuous activations. With 256 timesteps, the rate resolution is ~0.4%, which bounds the accuracy gap.

---

Q3: How does population coding handle high-dimensional inputs?

Population coding uses N neurons per input dimension, each with a Gaussian tuning curve centered at a different point in the input range. For a D-dimensional input, we need N×D neurons total.

Dimensionality management:

• For images (e.g., 28×28 MNIST = 784 dims), using N=10 neurons/dim gives 7,840 population neurons. This is feasible on Loihi 2 (1M neurons) but expensive.
• In practice, population coding is best for low-dimensional continuous spaces (robot joint angles, sensor values, control signals) where D < 50.
• For high-dimensional inputs, rate coding or TTFS is preferred — they use 1 neuron per dimension.

Tuning curve width (σ): Controlled by the beta parameter. sigma = range / (N-1) / beta. Larger beta → narrower curves → more precise but sparser representation. The default beta=1.5 gives good overlap between adjacent neurons.

---

Q4: What are the statistical properties of Poisson spike trains?

Poisson spike trains have several important properties:

• Inter-spike intervals (ISI) follow an exponential distribution: P(ISI = t) = λ·exp(-λt)
• Coefficient of variation CV = 1.0 (ISI std / ISI mean) — maximum variability for a renewal process
• Spike count in window T follows Poisson distribution: P(n) = (λT)^n · exp(-λT) / n!
• Fano factor = 1.0 (variance / mean of spike count)

Our implementation adds a refractory period (default 2ms), which modifies these properties:

• CV < 1.0 (more regular than pure Poisson)
• Effective maximum rate = 1000/refractory_ms = 500 Hz
• ISI distribution becomes a shifted exponential

The EncodingMetrics.mean_firing_rate_hz and sparsity fields capture these statistics for benchmarking.

---

Q5: How does conversion handle different ANN layer types?

Layer	Conversion Strategy	Notes
Dense/Linear	Direct weight transfer + threshold balance	Core conversion; 1:1 neuron mapping
Conv2D	Same weights, spiking conv operation	Weight-norm applied per output filter
BatchNorm	Folded into preceding layer: W' = W·γ/√(σ²+ε)	Must be folded before activation profiling
ReLU	Becomes the LIF neuron's threshold mechanism	Exact equivalence with IF neuron (no leak)
Dropout	Simply removed	Training-only regularization
MaxPool	Earliest-spike-wins per pooling window	Most active neuron fires first
Residual connections	Membrane potential addition at skip junction	Requires careful threshold calibration
Attention	Rate-coded QK^T with winner-take-all softmax approximation	Active research area; 3-5% extra accuracy loss typical
LayerNorm/GroupNorm	Similar to BatchNorm folding but per-instance	Requires running statistics from calibration set

The ANNtoSNNConverter currently handles Dense and Conv2D natively. BatchNorm folding is automatic. Attention conversion is experimental.

---

Q6: What latency considerations affect encoding choice?

Classification latency depends on encoding scheme:

Scheme	Min latency	Why
TTFS	1-2 ms	Strongest feature fires immediately
Rank-Order	N × ISI	All channels must fire once
Phase	1 cycle period	Need at least one oscillation cycle
Rate	Full window (50-200ms)	Need enough spikes for reliable count
Population	Same as rate	Rate-based sub-encoding

For real-time applications (robotics, control), TTFS encoding + single-pass SNN inference achieves <5ms total latency. The trade-off is that later, weaker features are encoded with less precision.

For batch classification (image recognition, NLP), rate coding with 100-256 timesteps gives best accuracy. The extra latency (10-25ms at 0.1ms dt) is acceptable.

Our EncodingMetrics.latency_to_first_spike_ms captures this for benchmarking.

---

Q7: How should encoding be adapted for specific neuromorphic hardware?

Different chips have different capabilities:

Intel Loihi 2:

• 0.1ms timestep resolution → all schemes work well
• 128 compartments/core → population coding feasible
• Native TTFS support via programmable dendrites
• Energy: ~0.9 nJ/spike

SpiNNaker 2:

• 1ms timestep → phase coding with fine offsets loses precision
• ARM cores → can run encoder in software alongside SNN
• Energy: ~10 nJ/spike (higher but more flexible)

BrainScaleS-2 (analog):

• Accelerated 1000× → rate coding window shrinks to µs
• Analog neurons → threshold balancing more critical (device mismatch)
• Energy: ~0.05 nJ/spike (lowest)

TrueNorth (IBM):

• 1ms timestep, binary weights → rate coding with stochastic rounding
• 256 neurons/core → population size limited
• Energy: ~26 pJ/spike

The EncoderConfig parameters (dt_ms, max_firing_rate_hz, refractory_period_ms) should be set to match hardware specs. Future work: automatic hardware profile detection.

---

Q8: How can we perform information-theoretic analysis of encoding quality?

The benchmark_encoding() method computes several information-theoretic metrics:

Mutual Information I(X; S):

• Measures how many bits of the input X are preserved in the spike train S
• Computed via binning: discretise both input and spike count, build joint histogram
• information_rate = I(X; S) / T_window gives bits/second

Coding Efficiency (bits per spike):

• efficiency = I(X; S) / total_spike_count
• TTFS achieves highest efficiency (~6-7 bits/spike) because each spike carries maximum temporal information
• Rate coding achieves ~0.3 bits/spike due to Poisson variability

Reconstruction Accuracy (R²):

• Encode → Decode round-trip; compute R² between input and output
• Rate coding: ~0.92 R² at 100ms window
• Population coding: ~0.97 R² (best, but most expensive)
• TTFS: ~0.95 R² (excellent efficiency-accuracy trade-off)

Lower bounds on information rate can also be computed analytically for Poisson processes: I ≥ ∫ [r(x)·log(r(x)/r̄) - (r(x) - r̄)] dx where r(x) is the tuning curve and r̄ is mean rate.

---

Q9: Can encoding schemes be combined for hybrid coding?

Yes — this is an active research direction. Some promising combinations:

1. Burst coding: First spike time encodes the primary value (TTFS), subsequent burst of spikes refines precision (rate within burst). Combines low latency with high accuracy.

2. Multiplexed coding: Different frequency bands carry different information. Low-frequency oscillation (theta, 4-8 Hz) carries position; high-frequency (gamma, 30-80 Hz) carries fine features. Phase coding relative to each band.

3. Population + temporal: Population coding for the value, with TTFS within each population neuron to encode certainty. More confident predictions → earlier spikes.

4. Multi-scale rate coding: Short initial window (5ms) for coarse classification, extended window (100ms) for fine-grained. Anytime computation — can read output at any point.

The NeuromorphicOrchestrator (34.5) will support pipeline configurations where different input dimensions use different encoding schemes.

---

Links: Issue #710 · Show & Tell #718 · SpikingNeuronModel #706 · SynapticPlasticityEngine #709