STDP and STDP variations with memristors for spiking neuromorphic learning systems

Teresa Serrano‐Gotarredona, Timothée Masquelier, Themis Prodromakis, Giacomo Indiveri, B. Linares-BarrancoFrontiers in Neuroscience 2013 · DOI

TL;DR

This paper explores how memristors, a novel nanoscale electronic component, can be used to build highly efficient and compact neuromorphic systems capable of Spike-Timing-Dependent Plasticity (STDP) learning. By leveraging memristors' inherent memory and variable resistance, the authors demonstrate a path towards asynchronous, biologically plausible learning in artificial neural networks, moving beyond the limitations of traditional CMOS-based approaches.

What Problem Does This Paper Address?

Building artificial neural systems that truly mimic the brain's learning capabilities has been a long-standing challenge in neuromorphic engineering. A key aspect of biological learning is Spike-Timing-Dependent Plasticity (STDP), where the strength of a synaptic connection changes based on the precise timing of pre- and post-synaptic spikes. Implementing STDP in traditional silicon (CMOS) is notoriously power-hungry and requires significant chip area for non-volatile memory and computation. This paper addresses the need for more compact, energy-efficient, and biologically realistic learning mechanisms in neuromorphic hardware, specifically by investigating the potential of memristors as a solution for implementing STDP.

Methods

The core of this paper lies in exploring how memristors can be integrated with CMOS neurons to implement STDP learning rules. The authors discuss two primary memristor physics models: the "moving wall" model and the "filament creation and annihilation" model. They then detail how these memristor characteristics can be leveraged to implement two types of STDP rules:

  • Pure timing-based STDP: This rule relies solely on the arrival times of pre- and post-synaptic spikes.
  • Hybrid STDP: This rule incorporates pre-synaptic spike timing along with the membrane potential and other state variables of the post-synaptic neuron.

The paper also outlines how these STDP implementations can be realized in cross-bar architectures, which are arrays of memristors suitable for massive parallelism. They further differentiate between additive, multiplicative, and quadratic STDP, linking them to the underlying memristor physics. The authors also discuss single-spike and double-spike STDP variants, highlighting their implementation complexity and potential benefits.

Datasets

Not discussed in the paper. The paper focuses on the theoretical and architectural aspects of implementing STDP with memristors, rather than empirical evaluation on specific datasets.

Key Insights and Ideas

  • Memristors as Synapses: Memristors, with their inherent memory and variable resistance, are a natural fit for emulating biological synapses, offering a compact and energy-efficient alternative to traditional CMOS implementations.
  • Asynchronous Learning: The proposed memristor-based STDP implementations allow for asynchronous operation, meaning learning and computation can occur concurrently and at their own pace, mirroring biological neural systems and avoiding the need for global synchronization.
  • Memristor Physics Matters: The specific physical mechanism of a memristor (e.g., "moving wall" vs. "filament formation") directly influences the type of STDP rule that can be implemented (e.g., quadratic vs. additive).
  • Flexibility in STDP Rules: The paper demonstrates that various STDP rules, including pure timing-based and hybrid approaches, as well as single-spike and double-spike variants, can be realized using memristors.
  • Cross-Bar Architectures: The inherent properties of memristors make them well-suited for dense cross-bar arrays, enabling the construction of large-scale neuromorphic systems.

Key Results

The paper doesn't present quantitative results in terms of performance metrics like accuracy or speedup on specific tasks. Instead, its key results are conceptual and architectural:

  • Demonstration of STDP Implementations: The paper provides detailed theoretical frameworks and circuit-level descriptions for implementing various STDP rules using memristors.
  • Link between Memristor Physics and STDP Type: A clear connection is established between the underlying physics of memristors (moving wall vs. filament) and the resulting STDP learning rule (quadratic vs. additive).
  • Architectural Blueprints: The paper outlines how these memristor-based STDP synapses can be integrated with CMOS neurons in cross-bar architectures for scalable neuromorphic systems.
  • Potential for Artificial Vision: An application in artificial vision is discussed, suggesting that these systems could be used for unsupervised learning and feature extraction from visual data.

Limitations and Future Work

  • Experimental Validation: While the paper presents theoretical frameworks and simulations, it acknowledges that large-scale neural memristive STDP systems have not yet been built. Experimental validation and demonstration of these concepts are crucial next steps.
  • Device Parameter Mismatch: The authors recognize that significant inter-device parameter mismatch in memristors is a major challenge for large-scale systems. They mention that research has begun to analyze this impact and that mechanisms like homeostasis might help compensate.
  • Reliability and Efficiency: Future work will focus on physically building these systems at scale, verifying their reliability, mismatch tolerance, and overall efficiency in real-world applications.
  • Exploring Higher-Order Neurons: The paper suggests future work will involve evaluating memristors in subsequent layers of neural networks, mimicking higher-order neurons to achieve more complex visual feature extraction.

Summary updated

Abstract

In this paper we review several ways of realizing asynchronous Spike-Timing-Dependent-Plasticity (STDP) using memristors as synapses. Our focus is on how to use individual memristors to implement synaptic weight multiplications, in a way such that it is not necessary to (a) introduce global synchronization and (b) to separate memristor learning phases from memristor performing phases. In the approaches described, neurons fire spikes asynchronously when they wish and memristive synapses perform computation and learn at their own pace, as it happens in biological neural systems. We distinguish between two different memristor physics, depending on whether they respond to the original "moving wall" or to the "filament creation and annihilation" models. Independent of the memristor physics, we discuss two different types of STDP rules that can be implemented with memristors: either the pure timing-based rule that takes into account the arrival time of the spikes from the pre- and the post-synaptic neurons, or a hybrid rule that takes into account only the timing of pre-synaptic spikes and the membrane potential and other state variables of the post-synaptic neuron. We show how to implement these rules in cross-bar architectures that comprise massive arrays of memristors, and we discuss applications for artificial vision.
BibTeX
@inproceedings{serranogotarredona2013stdp,
  author = {Teresa Serrano‐Gotarredona and Timothée Masquelier and Themis Prodromakis and Giacomo Indiveri and B. Linares-Barranco},
  title = {{STDP and STDP variations with memristors for spiking neuromorphic learning systems}},
  booktitle = {Frontiers in Neuroscience},
  year = {2013},
  doi = {10.3389/fnins.2013.00002},
}
Figure