Evolving Unipolar Memristor Spiking Neural Networks

Gerard David Howard, Larry Bull, Ben de Lacy CostelloAcalci 2015 · DOI

TL;DR

This paper explores the potential of unipolar memristors as a simpler, more robust alternative to bipolar memristors for synapses in neuromorphic spiking neural networks. By using a genetic algorithm to evolve networks with these unipolar synapses, the researchers demonstrate that they can achieve comparable or better performance on robotics tasks, particularly in terms of faster adaptation and evolution, while being easier to manufacture.

What Problem Does This Paper Address?

Neuromorphic computing aims to mimic the brain's architecture and function in hardware, often using spiking neural networks (SNNs) with specialized synaptic elements. Memristors, with their stateful and context-sensitive resistance properties, are promising candidates for these synapses. While much research has focused on bipolar memristors, which offer gradual, analog weight changes akin to biological plasticity, this paper investigates unipolar memristors. Unipolar memristors, which switch between two distinct states (conductive and resistive) and are less sensitive to voltage polarity, are simpler to fabricate and potentially more durable. The challenge addressed is whether these simpler unipolar memristor synapses can effectively support the learning and adaptive capabilities required for neuromorphic systems, and if they offer advantages over their bipolar counterparts and traditional non-plastic synapses.

Methods

The researchers employed a Genetic Algorithm (GA) to evolve the topology and parameters of spiking neural networks. The GA was used to automatically discover high-performing network configurations for specific tasks. The core novelty lies in comparing three types of synapses within these evolved networks:

  • Unipolar Memristor Synapses: These synapses switch between two resistance states (low and high) based on a threshold of accumulated coincidence events. They are ambivalent to spike order.
  • Bipolar Memristor Synapses: These synapses exhibit gradual, analog weight changes based on the timing of pre- and post-synaptic spikes (Spike-Time Dependent Plasticity - STDP).
  • Constant Synapses: These are non-plastic, fixed-weight connections, serving as a baseline.

The networks were evaluated on two robotics tasks: a reactive phototaxis task and a dynamic T-maze adaptation task. The GA was configured to evolve networks for a task-dependent number of generations, with fitness assigned based on performance in each task. Self-adaptive mutation rates were also incorporated into the GA to allow for more efficient exploration of the network space.

Datasets

Not discussed in the paper. The paper uses simulated environments for two robotics tasks (phototaxis and T-maze) as benchmarks.

Key Insights and Ideas

  • Simplicity vs. Performance: Unipolar memristors, despite their binary switching, can achieve performance comparable to or better than bipolar memristors, suggesting that biological realism in plasticity isn't always necessary for effective neuromorphic computation.
  • Faster Evolution and Adaptation: Unipolar memristor networks evolved significantly faster and adapted more quickly to dynamic environments in the T-maze task. This is attributed to their simpler attractor space, which allows the GA to more efficiently find useful network configurations.
  • Manufacturability Advantage: The inherent simplicity of unipolar memristors (binary switching) makes them easier to fabricate reliably and more durable, a crucial factor for large-scale neuromorphic hardware.
  • "Sweet Spot" for Evolvability: Unipolar memristors strike a balance between the complexity of bipolar memristors (richer behaviors but harder to evolve) and the simplicity of constant connections (easy to evolve but lack adaptability).
  • Weight Oscillators: Unipolar networks leverage "weight oscillators" – synchronized switching of synapses – to create dynamic connectivity maps and generate appropriate output actions, demonstrating a novel mechanism for behavior generation.

Key Results

  • Phototaxis Task: Unipolar memristor networks achieved better final best fitness and evolved solutions statistically faster than bipolar and constant networks. They also required significantly fewer generations to "solve" the task (0.76 generations on average for unipolar vs. 14.7 for bipolar and 77.6 for constant).
  • T-Maze Task: Unipolar memristor networks again demonstrated significantly faster adaptation, solving the task in an average of 12.47 generations, compared to 47.93 for bipolar and 27.73 for constant networks.
  • Fitness: While unipolar networks adapted faster, the final average fitness across both tasks was comparable between unipolar and bipolar networks, and often superior to constant networks.
  • Durability: The paper suggests that the fewer total switches required by unipolar memristors could lead to more long-lasting physical networks.

Limitations and Future Work

  • Scalability: The paper acknowledges that a lingering question is the scalability of unipolar synapses to more complex tasks. However, they hypothesize that the reduced attractor space of unipolar networks might actually make them more scalable than bipolar networks.
  • Biological Realism vs. Efficiency: The authors raise an interesting point about the trade-off between biological realism and computational efficiency in neuromorphic engineering, suggesting that non-biologically realistic architectures like those using unipolar memristors might offer a more direct path to efficient neuromorphic computing.
  • Further Exploration: The paper implicitly suggests future work in exploring the full potential of unipolar memristor networks in more complex scenarios and on hardware implementations.

Summary updated

BibTeX
@inproceedings{howard2015evolving,
  author = {Gerard David Howard and Larry Bull and Ben de Lacy Costello},
  title = {{Evolving Unipolar Memristor Spiking Neural Networks}},
  booktitle = {Acalci},
  year = {2015},
  doi = {10.1007/978-3-319-14803-8\_20},
}
Figure