Ian's Noesis Wiki

Properties1
tagscomputational-neuroscience, neural-networks, neuromorphic, vision, scale-space

3 min read

Covariant Receptive Fields

Covariant spatio-temporal receptive fields are a principled computational model for neuromorphic systems that unifies scale-space theory with spiking neural dynamics. Developed by Pedersen, Conradt, and Lindeberg (KTH Royal Institute of Technology), the theory provides a mathematically grounded way to process spatio-temporal signals in SNNs that is provably covariant to spatial affine and temporal scaling transformations.

Theoretical Foundation

The model combines two well-established frameworks:

  1. Spatial domain: Affine Gaussian kernels over space, drawn from scale-space-theory — the multi-scale representation pioneered by Lindeberg (1994, 2013) that models early visual receptive fields as Gaussian derivatives
  2. Temporal domain: Leaky-integrator and leaky integrate-and-fire (LIF) models over time — the standard spiking neuron formalism from computational neuroscience

The result is a set of spatio-temporal filters that share deep similarities with the receptive field properties found in the mammalian visual system (Hubel & Wiesel, 1962; Lindeberg, 2013, 2021).

Covariance Properties

The key advance is provable covariance: when the input undergoes spatial affine or temporal scaling transformations, the receptive field responses transform in a predictable, mathematically tractable way. This means the network’s representations are stable under natural visual transformations (rotation, shear, dilation, speed changes) — a property that standard ANNs lack without extensive data augmentation.

Application to SNN Training

The paper demonstrates that using these covariant receptive fields as a prior in event-based vision tasks significantly improves SNN training, which is otherwise known to be problematic for event-based inputs. Results show that SNNs equipped with these priors can outperform standard neural networks on event-based vision benchmarks.

This addresses a central challenge identified in the SNN literature: surrogate gradient methods work but don’t exploit the temporal advantages of spiking dynamics. Covariant receptive fields provide exactly this — a theoretically principled prior that guides the network toward representations that are both biologically plausible and computationally efficient.

Relevance to Vision Science

This work is particularly significant for vision researchers because:

  • It directly models the spatio-temporal receptive field properties of V1 simple cells using the Gaussian derivative framework
  • The covariance properties mirror known invariances in biological vision (scale invariance, speed tuning)
  • It provides a bridge between normative theories of visual receptive fields (Lindeberg, 2013, 2021) and neuromorphic hardware implementation
  • Event-based vision sensors (DVS/DAVIS) naturally produce spike-encoded output that SNNs with covariant receptive fields can process natively

References

  • Pedersen, J. E., Conradt, J., & Lindeberg, T. (2025). Covariant spatio-temporal receptive fields for spiking neural networks. Nature Communications, 16, 8231. https://doi.org/10.1038/s41467-025-63493-0 1
  • Lindeberg, T. (2013). A computational theory of visual receptive fields. Biological Cybernetics, 107, 589–635.
  • Lindeberg, T. (2021). Normative theory of visual receptive fields. Heliyon, 7, e05897.

Footnotes

  1. raw/papers/pedersen-2025-covariant-receptive-fields-snn.md ↩

Graph View

Backlinks