ecram-short-term-plasticity-neuromorphic

name	ecram-short-term-plasticity-neuromorphic
description	Cross-layer device-circuit-system co-design framework for leveraging non-equilibrium ECRAM dynamics as computational resources for short-term plasticity in neuromorphic circuits. ECRAM devices naturally exhibit volatile ionic dynamics that produce transient conductance modulation, which can be exploited for STP rather than treated as unwanted variability.
category	neuromorphic-computing
tags	["neuromorphic","short-term-plasticity","ECRAM","spiking-neural-networks","device-circuit-co-design","temporal-processing"]
created	2026-05-14
source	[{"title":"Leveraging Non-Equilibrium ECRAM Dynamics for Short-Term Plasticity in Neuromorphic Circuits","url":"https://arxiv.org/abs/2605.11243","arxiv_id":"2605.11243","date":"2026-05-13"}]

ECRAM Short-Term Plasticity for Neuromorphic Circuits

Overview

This skill describes a cross-layer device-circuit-system co-design methodology that transforms volatile ECRAM (Electrochemical RAM) device dynamics from a tolerated artifact into a computational resource for implementing short-term plasticity (STP) in neuromorphic circuits.

Key Innovation

Rather than treating non-equilibrium ionic dynamics in ECRAM as undesirable variability, this framework exploits activity-dependent conductance modulation as a native hardware substrate for temporal computation and STP.

Core Methodology

1. Device-Level: ECRAM Characterization

• Key Property: ECRAM devices exhibit transient conductance modulation (~1.5 KOhms per spike)
• Mechanism: Non-equilibrium ionic dynamics produce temporary conductance changes
• Behavioral Model: Compact model suitable for circuit-level simulation based on experimentally characterized devices
• Energy: 2 pJ per spike operation

2. Circuit-Level: Delay-Feedback LIF Architecture

• Architecture: Delay-feedback leaky integrate-and-fire (LIF) neuron co-designed with ECRAM synapses
• Key Features:
  • Tunable delay-feedback spike-generation path
  • Transient device dynamics directly modulate neuron excitability
  • Activity-dependent conductance modulation with negligible circuit overhead
  • Extends across multiple neuron topologies

3. System-Level: STP Behaviors

The architecture demonstrates two key STP behaviors:

• Synaptic Facilitation: Transient increase in synaptic efficacy following activity
• Intrinsic Excitability Modulation: Activity-dependent changes in neuron firing threshold

4. Network-Level: Temporal Filtering

• Individual synapses act as tunable temporal filters within SNNs
• Frequency-selective spike processing emerges from the device dynamics
• Enables temporal computation without additional circuit complexity

Implementation Patterns

Pattern 1: ECRAM Behavioral Model

class ECRAMBehavioralModel:
    """Compact behavioral model of ECRAM device dynamics."""
    
    def __init__(self, conductance_change_per_spike=1.5e3):  # Ohms
        self.conductance_change = conductance_change_per_spike
        self.state = 0  # Initial conductance state
        
    def update_conductance(self, spike_count, time_delta):
        """Update conductance based on spike activity."""
        # Non-equilibrium ionic dynamics
        delta_g = self.conductance_change * spike_count
        self.state += delta_g
        # Natural decay over time
        self.state *= decay_factor(time_delta)
        return self.state

Pattern 2: Delay-Feedback LIF Neuron

class DelayFeedbackLIF:
    """LIF neuron with delay-feedback for STP."""
    
    def __init__(self, ecram_synapse, delay_tau):
        self.synapse = ecram_synapse
        self.delay_tau = delay_tau  # Tunable delay parameter
        self.membrane_potential = 0
        self.spike_history = []
        
    def process_input(self, input_spike):
        """Process input spike through ECRAM synapse."""
        # Update synapse conductance
        conductance = self.synapse.update(input_spike)
        # Modulate neuron excitability
        self.membrane_potential += conductance * input_spike
        # Check threshold (modulated by recent activity)
        threshold = self.base_threshold * self.get_excitability_modulation()
        if self.membrane_potential >= threshold:
            self.spike_history.append(current_time)
            return 1  # Spike
        return 0

Pattern 3: Frequency-Selective Processing

class FrequencySelectiveSNN:
    """SNN with frequency-selective temporal filtering."""
    
    def __init__(self, synapses):
        self.synapses = synapses  # ECRAM-based synapses
        
    def process_spike_train(self, input_spikes):
        """Process spike train with frequency selectivity."""
        output = []
        for spike in input_spikes:
            # Each synapse acts as tunable temporal filter
            filtered = self.synapses.filter(spike, frequency_band)
            if filtered > threshold:
                output.append(spike)
        return output

Applications

1. Temporal Pattern Recognition: Frequency-selective processing for sequence tasks
2. Working Memory: Short-term plasticity as a mechanism for transient memory
3. Adaptive Filtering: Tunable temporal filters for signal processing
4. Neuromorphic Sensing: Event-driven temporal feature extraction
5. Reservoir Computing: Rich dynamics for temporal computation

Key Advantages

Aspect	Traditional SNN	ECRAM-STP SNN
STP Implementation	Explicit circuits	Native device dynamics
Energy per spike	~10-100 pJ	~2 pJ
Circuit Overhead	Additional components	Negligible
Temporal Resolution	Limited by clock	Continuous device dynamics
Adaptability	Fixed parameters	Activity-dependent

Hardware Implementation

Device Requirements

• ECRAM devices with characterized transient conductance modulation
• Conductance change: ~1.5 KOhms per spike
• Retention time: Tunable via device materials
• Energy: 2 pJ per spike operation

Circuit Design

• Delay-feedback LIF neuron topology
• ECRAM synapse crossbar array
• Minimal additional control circuitry
• Scalable to large arrays

Research Extensions

1. Multi-timescale STP: Combining short-term and long-term plasticity
2. Heterogeneous Arrays: Mixed device types for diverse temporal filters
3. Learning Rules: STDP combined with STP for unsupervised learning
4. System Integration: Full neuromorphic chip implementation
5. Application-Specific: Optimized architectures for temporal tasks

References

• Primary: "Leveraging Non-Equilibrium ECRAM Dynamics for Short-Term Plasticity in Neuromorphic Circuits" (arXiv:2605.11243)
• Related: ECRAM device characterization, STP in biological systems, neuromorphic circuit design

Activation

• ecram short-term plasticity
• neuromorphic temporal processing
• device-circuit co-design
• spiking neural network hardware
• non-equilibrium dynamics
• memristive synapses