Skip to main content

Implemented Papers

SpikeSEG synthesises methods from four publications. This page summarises each paper's contribution and how it maps to the codebase.

[1] Kheradpisheh et al. 2018 -- STDP-Based Spiking Deep CNNs​

Title: STDP-based spiking deep convolutional neural networks for object recognition Venue: Neural Networks, vol. 99, pp. 56--67

Contributions used:

  • Simplified STDP rule with soft-bounded multiplicative updates (Δw=a±⋅w(1−w)\Delta w = a^{\pm} \cdot w(1-w)).
  • Three-layer spiking convolutional architecture with max pooling.
  • Winner-Take-All lateral inhibition (global + local).
  • Convergence criterion Cl<0.01C_l < 0.01.
  • Weight initialisation near 1 (μ=0.8\mu = 0.8).

Codebase mapping: spikeseg.learning.stdp, spikeseg.learning.wta, spikeseg.core.neurons.

[2] Kirkland et al. 2020 -- SpikeSEG​

Title: SpikeSEG: Spiking segmentation via STDP saliency mapping Venue: 2020 IEEE IJCNN, pp. 1--8

Contributions used:

  • Encoder-decoder architecture with tied weights.
  • Transposed convolutions and max unpooling to trace classification spikes back to pixel space (saliency map).

Codebase mapping: spikeseg.models.decoder, spikeseg.models.spikeseg.

[3] Kirkland et al. 2022 -- HULK-SMASH​

Title: Unsupervised spiking instance segmentation on event data using STDP features Venue: IEEE Transactions on Neural Networks and Learning Systems

Contributions used:

  • HULK (Hierarchical Unravelling of Linked Kernels): decode each classification spike individually to get per-instance pixel masks.
  • ASH (Active Spike Hashing): compress 4D spike activity into a 2D binary feature-time matrix.
  • SMASH score: Jaccard(ASH) x IoU(BBox) for grouping instances into objects.

Codebase mapping: spikeseg.algorithms.hulk, spikeseg.algorithms.smash.

[4] Kirkland et al. 2023 -- IGARSS Space Domain Awareness​

Title: Neuromorphic sensing and processing for space domain awareness Venue: IGARSS 2023

Contributions used:

  • Layer-wise subtractive leak (λ\lambda = 90% and 10% of θ\theta in layers 1 and 2).
  • 10x higher STDP learning rates (a+=0.04a^{+} = 0.04, a−=0.03a^{-} = 0.03) for faster convergence.
  • Volume-based evaluation with informedness as the primary metric (target: 89.1%).
  • Application to the EBSSA satellite dataset.

Codebase mapping: EncoderConfig.from_paper("igarss2023"), scripts/evaluate.py --volume-based.

Additional References​

#Citation
[5]W. Maass, "Networks of spiking neurons: The third generation of neural network models," Neural Networks, vol. 10, no. 9, pp. 1659--1671, 1997.
[6]G.-Q. Bi and M.-M. Poo, "Synaptic modifications in cultured hippocampal neurons," Journal of Neuroscience, vol. 18, no. 24, pp. 10464--10472, 1998.
[7]S. Afshar et al., "Event-based object detection and tracking for space situational awareness," IEEE Sensors Journal, vol. 20, no. 24, pp. 15117--15132, 2020.
[8]G. Orchard et al., "Converting static image datasets to spiking neuromorphic datasets using saccades," Frontiers in Neuroscience, vol. 9, p. 437, 2015.

See Citation for BibTeX entries.