multi-timescale-conductance-spiking-networks

name: multi-timescale-conductance-spiking-networks description: "Multi-Timescale Conductance (MTC) Spiking Networks methodology — sparse, gradient-trainable framework with rich firing dynamics. Derives differentiable conductance-based neurons with fast/slow/ultra-slow timescales, enabling direct BPTT without surrogate gradients. Benchmark: Mackey-Glass chaotic time series forecasting. Activation: MTC-SNN, conductance spiking, BPTT spiking, multi-timescale neuron, Mackey-Glass forecasting, I-V curve shaping, 多时间尺度电导脉冲网络."

Multi-Timescale Conductance Spiking Networks (MTC-SNN)

Gradient-trainable SNN framework where neural dynamics emerge from shaping the current-voltage (I-V) curve via tunable fast, slow, and ultra-slow conductance elements. Enables direct backpropagation through time (BPTT) without surrogate-gradient approximations.

Paper Reference

• Title: Multi-Timescale Conductance Spiking Networks: A Sparse, Gradient-Trainable Framework with Rich Firing Dynamics for Enhanced Temporal Processing
• Authors: Alex Fulleda-Garcia, Saray Soldado-Magraner, Josep Maria Margarit-Taulé
• arXiv: 2605.11835v1 (cs.NE, cs.AI, cs.LG)
• Date: 2026-05-12
• PDF: https://arxiv.org/pdf/2605.11835.pdf

Core Problem

SNNs face a fundamental tradeoff:

1. LIF neurons are simple but strip away rich temporal dynamics — limited control over excitability, restricted firing repertoire
2. Surrogate gradients introduce forward-backward mismatch — limits faithful learning of complex temporal dynamics for regression
3. Sparsity managed indirectly via threshold/loss regularization rather than emerging from interpretable neuron mechanisms

Key Innovation: Conductance-Shaped I-V Curves

Circuit-Theoretic Foundation

Builds on Ribar & Sepulchre's reduced conductance-based framework where neuron behavior is controlled by shaping its I-V curve via parallel interconnection of positive and negative conductance elements at different timescales.

Three-Timescale Architecture

Governing Equations

Voltage-gated conductance dynamics:

τx · dUx/dt = -Ux + Vm
Ix± = αx± · tanh((Vm - Ux) / δx±)

Membrane potential:

τm · dVm/dt = -Vm + Σ Ix± + Iinput

Where:

• τx: time constant for state variable Ux relative to Vm
• αx±: maximal conductance (gain) of the channel
• δx±: voltage range where element is active

Firing Regimes

Single model smoothly transitions between:

• Tonic spiking: constant firing rate under sustained input
• Phasic spiking: burst at stimulus onset then silence
• Bursting: clusters of spikes separated by quiescent periods
• Spike frequency adaptation: decreasing rate under constant input

Discrete-Time Differentiable Formulation

Key Breakthrough: No Surrogate Gradients

Unlike LIF/AdLIF models that require surrogate gradient approximations (ArcTan, SLAYER), the MTC model is fully differentiable because:

• Continuous voltage trajectory Um(t) is inherently smooth
• Spike generation emerges from continuous nonlinear dynamics
• State variables (Ux) evolve differentiably
• Standard BPTT applies directly

Discretization

Explicit Euler-Forward discretization:

Um[t+1] = Um[t] + dt/τm · (-Vm[t] + Σ Ix±[t] + Iinput[t])
Ux[t+1] = Ux[t] + dt/τx · (-Ux[t] + Vm[t])

Synaptic Transduction Model

Raw action potentials normalized to [0,1] via semi-digital communication function:

• Suppresses sub-threshold activity (s(t) = 0 for Vm < Vth)
• Maintains differentiability at spike onset for gradient computation
• Approximates nonlinear relationship between pre-synaptic voltage and neurotransmitter release

Experimental Results: Mackey-Glass Forecasting

Setup

• Task: Chaotic Mackey-Glass time series regression (γ=0.1, β=0.2, n=10, τ=17)
• Prediction horizon: d = 675 timesteps (~5× Lyapunov time)
• Architecture: Feedforward spiking network (no recurrence); temporal memory from intrinsic neuron dynamics
• Training: Adam optimizer, 10,000 epochs, batch size 128, Cosine Annealing LR
• Readout: Linear decode + 4th-order low-pass filter for continuous signal reconstruction

Baselines

Results

• MTC outperforms LIF and AdLIF on Mackey-Glass forecasting
• Achieves higher sparsity naturally from conductance dynamics
• Rich firing patterns captured without recurrent network overhead
• Feedforward-only architecture with intrinsic temporal memory

Why This Matters

1. Eliminates Surrogate Gradient Mismatch

The forward dynamics and backward gradients are consistent — no ad-hoc surrogate functions needed. This is critical for regression tasks where approximation error, noise, and spike discretization can severely degrade continuous-valued outputs.

2. Conductance-Shaped Excitability as Computation

The neuron's I-V curve is itself a computational mechanism — biological circuits exploit conductance modulation to attune to input statistics. MTC captures this explicitly.

3. Analog Circuit Compatibility

Localized conductance elements implementable with compact transconductance blocks (subthreshold MOS) — I-V characteristics need not be exact tanh functions, any approximately monotone nonlinearity suffices.

Implementation Guide

Step 1: Define Conductance Elements

class MTCNeuron:
    def __init__(self):
        # Fast timescale (negative conductance)
        self.tau_f = 1.0  # → instantaneous
        self.alpha_f_minus = 1.0
        self.delta_f_minus = 0.1
        
        # Slow timescale (positive conductance)
        self.tau_s = 10.0
        self.alpha_s_plus = 0.5
        self.delta_s_plus = 0.2
        
        # Ultra-slow timescale
        self.tau_us = 100.0
        self.alpha_s_minus = 0.3
        self.alpha_us_plus = 0.2

Step 2: Forward Pass (Differentiable)

def forward(self, Vm, U_f, U_s, U_us, I_input, dt):
    # Update filtered voltages
    U_f = U_f + dt/self.tau_f * (-U_f + Vm)
    U_s = U_s + dt/self.tau_s * (-U_s + Vm)
    U_us = U_us + dt/self.tau_us * (-U_us + Vm)
    
    # Compute conductance currents
    I_f = self.alpha_f_minus * torch.tanh((Vm - U_f) / self.delta_f_minus)
    I_s = self.alpha_s_plus * torch.tanh((Vm - U_s) / self.delta_s_plus)
    I_s_minus = -self.alpha_s_minus * torch.tanh((Vm - U_s) / self.delta_s_minus)
    I_us = self.alpha_us_plus * torch.tanh((Vm - U_us) / self.delta_us_plus)
    
    # Update membrane potential
    Vm = Vm + dt/self.tau_m * (-Vm + I_f + I_s + I_s_minus + I_us + I_input)
    
    return Vm, U_f, U_s, U_us

Step 3: Training with BPTT

# Standard PyTorch training loop — no surrogate gradients needed
optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)
for epoch in range(10000):
    output = model(input_sequence)  # BPTT through conductance dynamics
    loss = mse_loss(filtered_output, target)
    loss.backward()  # Exact gradients through differentiable dynamics
    optimizer.step()

Hyperparameter Tuning Strategy

1. Phase-space analysis: Analyze I-V curves and Vm traces of individual neurons under varying inputs to ensure rich temporal behavior
2. Conductance time constants: Grid search over τs, τus ratios relative to τm
3. Gain parameters: Balance αx± to avoid saturation or silence
4. Voltage sensitivity: Adjust δx± to control activation sharpness

Applications

1. Chaotic time series forecasting: Mackey-Glass, Lorenz, financial time series
2. Event-based sensor processing: DVS cameras, neuromorphic audio
3. Temporal pattern recognition: Speech, gesture, EEG decoding
4. Neuromorphic hardware deployment: Analog circuit implementation
5. Low-power edge inference: Sparse event-driven computation

Activation Keywords

• MTC-SNN
• multi-timescale conductance spiking
• conductance-based spiking neuron
• BPTT spiking neural network
• I-V curve shaping neuron
• surrogate-gradient-free SNN
• Mackey-Glass spiking forecasting
• 多时间尺度电导脉冲网络

Related Skills

• snn-learning-survey — Comprehensive SNN learning rule survey
• spiking-neural-network-analysis — SNN paper analysis patterns
• spikingjelly-framework — SpikingJelly framework usage
• ei-network-chaos-synchrony-theory — E/I network dynamics theory

Limitations & Open Questions

• Higher per-neuron computational cost than LIF
• More hyperparameters to tune (conductance gains, timescales)
• Validation needed on larger-scale tasks beyond Mackey-Glass
• Recurrent MTC networks not yet explored
• Hardware implementation details pending

Future Directions

1. Recurrent MTC-SNN architectures
2. Self-supervised pre-training for conductance neurons
3. Cross-modal temporal processing
4. Real neuromorphic chip deployment (Loihi, SpiNNaker)
5. Integration with attention/spiking transformer architectures